New strategies for row-crop management based on cost-effective remote sensors

NEW STRATEGIES FOR ROW-CROP MANAGEMENT BASED ON COST-EFFECTIVE REMOTE SENSORS I n t e r n a t i o n a l P h. D J O R G E M A R T Í N E Z G U A N T E R U N I V E R S I D A D D E S E V I L L A E S C U E L A T É C N I C A S U P E R I O R D E I N G E N I E R Í A A G R O N Ó M I C A 2 0 1 7 T h e s i s D i r e c t o r s Dr. M A N U E L P É R E Z R U I Z Dr. J U A N A G Ü E R A V E G A Universidad de Sevilla Escuela Técnica Superior de Ingeniería Agronómica DOCTORAL THESIS New Strategies for Row-Crop Management Based on Cost-Effective Remote Sensors Authored by JORGE MARTÍNEZ GUANTER Thesis Directors Dr. MANUEL PÉREZ RUIZ Dr. JUAN AGÜERA VEGA Sevilla, June 2017 Informe del Director de Tesis sobre Presentación de Tesis Doctoral El Dr. D. MANUEL PÉREZ RUIZ como Tutor de la Tesis Doctoral titulada “New Strategies for Row-Crop Management Based on Cost-Effective Remote Sensors” realizada por D. JORGE MARTÍNEZ GUANTER en el programa de doctorado DOCTORADO INTERUNIVERSITARIO EN INGENIERÍA AGRARIA, ALIMENTARIA, FORESTAL Y DEL DESARROLLO RURAL SOSTENIBLE, regulado por el R.D. 99/2011 de 28 de enero, INFORMA FAVORABLEMENTE EL DEPÓSITO DE LA MISMA Y LA IDONEIDAD DE SU REALIZACIÓN EN LA MODALIDAD DE COMPENDIO DE ARTÍCULOS, dado que reúne las condiciones necesarias para su defensa. Y para que así conste, firmo la presente. En Sevilla a 20 de Junio de 2017. El Director de la Tesis Manuel Pérez Ruiz Agradecimientos Muchas son las personas que han contribuido directa o indirectamente a la realización de esta tesis doctoral, durante la cual he tenido la suerte de contar con un excepcional grupo de trabajo, sin cuyo apoyo y ayuda esto no hubiera sido posible. En primer lugar, mi más sincero agradecimiento a mis directores de tesis, Manuel Pérez Ruiz y Juan Agüera Vega. Ha sido un camino largo, y a Manuel le agradezco su dedicación continua, su amistad y capacidad de trabajo, todo lo que me ha enseñado y su confianza en mí y mi trabajo desde el primer día. A Juan le debo el haberme iniciado en este mundillo, haber despertado mi interés y curiosidad hace algunos años, y haber compartido su amplio conocimiento para avanzar durante la tesis. Hay personas que hacen y personas que hacen hacer. Vosotros “hacéis hacer haciendo”, un ejemplo muy valioso para quienes queremos seguir vuestros pasos. Gracias. También agradecer a Gregorio Egea su colaboración y ayuda en los trabajos de la tesis, y su apoyo en el día a día. Su iniciativa y valiosas aportaciones han permitido desarrollar muchos de los aspectos de este trabajo, y sus consejos me han ayudado a avanzar en el camino de la investigación. Espero seguir aprendiendo y trabajando contigo. Agradecer también a todos los colegas colaboradores e investigadores con quienes he disfrutado y aprendido a lo largo de la tesis. A los coautores de los papers: a Miguel Garrido (ha sido un placer conocerte y trabajar contigo), a Jose Enrique y Carmen del IRNAS, a Constantino Valero y David Slaughter, todos ellos se han ofrecido a ayudarme en todo momento y sus aportaciones han sido determinantes para este trabajo. Los trabajos que se presentan son colaborativos, y el mérito es de todos. Ojalá que podamos seguir colaborando en futuras investigaciones. Un sincero agradecimiento a mis compañeros de la empresa Soluciones Agrícolas de Precisión. En ella he encontrado además de un grupo de trabajo, un gran grupo humano y unos amigos excepcionales. Sin vosotros esto no hubiera sido posible. A mis amigos. Tengo la suerte de contar con muchos, de muchas partes y siempre estáis presentes allá donde esté. Os debo mucho de mi forma de ser y de la forma en la que valoro las cosas. Gracias por vuestro apoyo. A mi abuela, agradecerle el inmenso amor que me tiene y que comparto, y porque sé la ilusión y alegría que le hace esto. Gràcies Yaya. Sobre todo, a mis padres y hermano. A ellos les debo todo: mi educación, mis valores, mi afán por esforzarme y por hacer bien las cosas. El apoyo que me han dado siempre y su inmenso cariño hacen que esté donde estoy hoy, y harán que siga creciendo como persona. Qué orgullo de padres y hermano tengo. A Marta, por todo lo que compartimos en la vida y por el cariño y apoyo, que es enorme y mutuo. Gracias. 2 Thesis Index Figure List 3 Table List 6 Abbreviation List 7 Abstract 8 Research Contribution 9 I. Introduction……………………………………………………….. 12 1.1 Background 12 1.2 Advances in Automated Weed Control 15 1.3 Crop Water Stress, Aerial Thermography and Smart Water Management 20 II. Overall Summary of the Objectives………………………….. 29 III. Publications. Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet……………………….. 31 Abstract 32 3.1 Introduction 33 3.2 Materials and Methods 35 3.3 Results and Discussion 41 3.4 Conclusions 47 3.5 References 48 III. Publications. Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards.…………. 52 Abstract 53 4.1 Introduction 54 4.2 Materials and Methods 56 4.3 Results 62 4.4 Discussion 68 4.5 Conclusion 72 4.6 References 73 III. Publications. Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: Two Scenarios……………………………………… 77 Abstract 78 5.1 Introduction 79 5.2 Materials and Methods 81 5.3 Results 83 5.4 Conclusions 100 5.5 References 102 IV. General Results………………………………………………… 105 IV. General Discussions of the Results and Future Work….. 108 VI. Conclusions…………………………………………………...... 116 References…………………………………………………………... 118 3 Figure List Introduction Figure 0.1: The canopy–air temperature difference (Tc -Ta) versus air vapour pressure deficit (VPD) for non-water stressed (∆Tpot) and maximally stressed (∆Tdry). Baseline approach of the crop water stress index calculation (Idso et al., 1981). Adapted from Maes and Steppe (2012)……………………………………… 19 Chapter 1 Figure 0.1: Components of the temperature data acquisition system: a) infrared sensor MLX90614 b) evaluation board EVB90614, c) Arduino Leonardo board with Ethernet/SD shield mounted above, and d) hand-held infrared thermometer Fluke 62………………………………………………………………………………….. 38 Figure 0.2: Workflow scheme for the UAV mission planning, survey and processing……………………………………………………………………………….. 39 Figure 0.3: a) Correlation between leaf temperature measured with the reference thermometer and temperature measured with the inexpensive IR sensor in a potted sugar beet plants and; b) field-grown sugar beet plants……… 40 Figure 0.4: Changes in canopy-to-air temperature differences (∆T = Tc-Ta) for FI and DI plants throughout the experimental period. Each point indicates the mean of six potted plants. The error bars denote the standard error……………... 41 Figure 0.5: Relationship between the difference among ∆T of FI and DI plants and the relative evapotranspiration (RET).………………………………………….. 42 Figure 0.6: Boxplot diagrams of mean canopy temperature (Tc) measured at various flight levels with LITAS. Measurements were taken throughout the irrigation season on both a) irrigated and b) deficit-irrigated plots………………… 43 Figure 0.7: Boxplot diagrams of mean canopy temperature (Tc) measured at various flight levels with the thermal camera. Measurements were taken throughout the irrigation season on both a) irrigated and b) deficit-irrigated plots. 44 Chapter 2 Figure 4.1: Thermal mosaic acquired with a FLIR Tau 2 324 thermal camera on board a RPAS model Phantom 2 observing (left) the experimental SHD olive orchard (white rectangle), and (right) details of the hot and cold reference surfaces………………………………………………………………………………….. 58 Figure 4.2: Example of bi-modal histogram of temperatures obtained from a thermal image of an experimental plot………………………………………………… 59 Figure 4.3: (left) Fraction of a thermal mosaic in which FI and 45RDICC trees have been delimited with white and blue dashed rectangles, respectively. The white circles show the hot reference surfaces. Date of flight: DOY 218; (upper right) 4 thermal image plotted in the left panel with a superimposed layer denoting the vegetation pixels selected by the segmentation algorithm based on the full width at half maximum rule (FWHM); (middle right) thermal image with a superimposed layer showing the vegetation pixels selected by the segmentation algorithm based on the full width at one fifth maximum rule (FWFM); (lower right) thermal image with a superimposed layer showing the vegetation pixels selected by the segmentation algorithm based on the full width at one-eighth maximum rule (FWEM). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article) ............................................... 59 Figure 4.4: Example of seasonal effect on the NWSB (∆T = a + b·VPD). The upper panel (a) shows the ∆T vs VPD relationship when data of the period DOY 167– 275 are gathered together. The lower panel (b) shows the same relationship when data are split in three periods: A (June-July, DOY 167–212), B (August, DOY 215– 243), C (September, DOY 244–273). Only clear-sky days were used in the calculations. In both panels, data for GMT = 14 h have been used. The straight lines represent the best fit to the data, whose fitted parameters and coefficients of determination are shown in Tables 2 (panel a) and 3 (panel b)…………………................................................................................................... 62 Figure 4.5: Diurnal variation of the NWSB (∆T = a + b·VPD) for the period June– July (DOY 167–212). The upper panel (a) shows the ∆T vs VPD relationship for GMT 8 to GMT 12. The lower panel (b) shows the same relationship for GMT 13 to GMT 18. Only clear-sky days were used in the calculations. The straight lines represent the best fit to the data, whose fitted parameters and coefficients of determination are shown in Table 4.3……………………………………………….... 63 Figure 4.6: Diurnal evolution of the NWSB-intercept (◦C) for the three different periods of study: A (June–July, DOY 167–212), B (August, DOY 215–243), C (September, DOY 244–273). The lines represent the best-fit to the data………… 63 Figure 4.7: Diurnal evolution of the NWSB-slope (◦C kPa−1) for the three different periods of study: A (June–July, DOY 167–212), B (August, DOY 215–243), C (September, DOY 244–273). The lines represent the best-fit to the data………… 64 Figure 4.8: Relationship between NWSB-intercepts and zenith solar angle for the period (a) 08:00-18:00, (b) 08:00-13:00 and (c) 13:00-18:00……………………… 65 Figure 4.9: Seasonal time-course of (a) CWSI determined from canopy temperature measured with the infrared thermometers (IRTS) at 12.00 GMT, (b) CWSI derived from RPAS thermal imaging (12.00 GMT) and (c) REW. The NWSB shown in Table 4.3 for periods A, B & C were used in CWSI calculation. In (a), only clear-sky days were used for CWSI determination. Each point is the mean of two (a), four (b) and three (c) replicates per treatment. Down-facing arrows indicate the onset of water stress periods in RDI treatments; the up-facing arrow indicates the end of a water stress period. The error bars represent the standard error of the mean (SE). In (b) and (c), asterisks denote significant differences at P < 0.05…………………………………………………………………………………… 66 Figure 4.10: Relationship between CWSI determined from aerial thermal imaging and (a) midday stem water potential (ψst), (b) midday leaf water potential (ψl), (c) stomatal conductance (gsm) and (d) leaf transpiration rate (Em) for FI, 45RDICC and 45RDITP treatments. The straight lines represent the fitted regression lines to the data……………………………………………………………………………………….. 67 11 Introduction 12 I. Introduction 1.1 Background Agriculture has undergone yield-enhancing shifts in the past, including mechanisation before the second world war and the introduction of new crop varieties and agricultural chemicals in the 1950s and 1960s (Grassini et al.,2013). Currently, achieving the transformation to digital and sustainable agriculture is a major challenge that will need to be overcome without jeopardising the capacity of the agriculture sectors to meet the world’s food needs (FAO, 2016). In the coming decades, population increases will be concentrated in regions with the highest prevalence of undernourishment and high vulnerability to the impacts of climate change, while global food demand in 2050 is projected to increase by at least 60 per cent above today’s levels (FAO, 2016). Therefore, food production must be increased and enhanced by adopting different methods such as expanding arable areas where possible or intensifying yields and cropping intensity. Whatever the alternative, its necessary to achieve this goal with the least disturbance to environmental parameters while meeting food safety requirements and with the highest efficiency possible with regard to the use of natural and human resources. This context of food insecurity, similar to that proposed in Malthusian predictions, has been largely restrained by the advent of the so-called the “new green revolution”. This new paradigm seeks an information-guided plant production, relying on precision agriculture (PA) techniques that allow the study of fields at a much finer resolution. Farmers have always known that crop in certain areas of their fields grow and produce differently than others (Mulla, 2017), but until the arrival of these site-specific handling techniques, farmers lacked the technology to enhance their ability to make spatiallyprecise management decisions. Accurate crop management helps farmers tailor inputs such as agrochemicals, water or energy based on diverse information sources. Advances in data collection, mechatronics, yield monitoring, remote and proximal sensing, and controlled agricultural traffic using improved GPS-guided farm machinery or variable rate application directly affect the quantity and quality of the crop, the efficiency of the operations and their environmental impact (Zude-Sasse et al., 2016). The economic benefits of these site-specific management techniques are well documented (Heege, 2015; Smith et al., 2013; Zhang et al., 2012; Whipker et al., 2009) Introduction 13 and can lead to dramatically reduced environmental impacts while increasing profitability; however, a low adoption rate is still reported by experts and researchers (Lambert et al., 2015; Ellis et al., 2010; Tey et al., 2012). Among other factors, such as a lack of information and education on ICT or the access to bank credit (Aubert et al., 2012), wider adoption is also limited because of the high costs associated with the technology and because the development of new technology is designed for large-scale production, in which within-field variations are great enough to justify the cost of the technological equipment (Cassman, 1999). The adoption of PA by small-scale producers can be accomplished progressively by focusing on incorporating technologies that are less expensive, adapted to the conditions of their fields or machinery, easy to use and with standardised components and programming languages. This adoption through cost reduction is essential if smallholders in high-potential areas are to intensify their production and contribute to economic growth. It is also essential in semi-arid or remote areas to manage and augment the natural resource base to promote development (Poulton et al., 2010). Under this scenario, it is in horticultural or industrial crops (with higher added value) in which efforts must be made to incorporate more affordable systems, which allow data collection, field monitoring or variable applications to achieve the same yields but at a lower cost. Horticultural crops are typically grown in row patterns, forming a semi-structured environment in which there are different canopy structures for different growth stages (Yao et al., 2009). This configuration allows greater individual plant control than arable crops at initial stages and the automation of certain tasks such as inter-row weeding, automatic plant scouting or autonomous navigation between crop lines. Although drip or sprinkler irrigation systems are commonly used in the aforementioned row crops and intelligent weed control systems are already commercially viable (e.g. Weedseeker®) the heavy reliance of data for precision management still remains a barrier. Data is becoming ever more valuable, as agricultural business development and food policy decisions are being made based upon data (Ferris et al., 2016). Traditional data collection through sampling or hand measurements in these environments can be a daunting task, which contributes to the fact that manual labour is an important cost component (Bechar et al., 2016). One of the important requirements of agricultural automation systems is that sensors must obtain accurate data in a robust and field-deployable manner. Thus, automated sensor-data gathering fulfils the need to Introduction 14 measure a biosystem’s behaviour and allows the possibility of correction if measurements are not compliant with the proposed objectives. Methods for this data collection can be roughly divided into proximal measurements taken from ground-based platforms or vehicles and remote measurements taken from airborne sensors on unmanned aerial vehicles (UAVs), planes or satellites (Slaughter, 2014). On the one hand, sensing devices such as conductivity, optoelectronic, image or distance sensors mounted on terrestrial platforms have been widely applied in precision agricultural studies for mapping environmental and crop factors. In recent years, several studies on ground-based sensing have been conducted, including studies on soil characterisation from texture, salinity or ground-water content using electromagnetic induction sensors (Doolittle et al., 2014; Quebrajo et al., 2016; Niu et al., 2015; Gumiere et al., 2014), as well as on-the-go soil cutting resistance sensors (Agüera et al., 2013), which have also been designed and tested to generate soil strength profile maps. Ground-sensing vehicles or platforms have been developed for phenotyping and diagnosing crop conditions based on reflectance data through close visible, multispectral or hyperspectral sensing (Andrade-Sanchez et al., 2014; Busemeyer et al., 2013; Bu et al., 2017; Kitchen et al., 2010; Caturegli et al., 2015). Advanced ground image techniques have been employed for measuring structural crop parameters such as leaf area index (LAI) or seasonal canopy development (Yang et al., 2017). Weed detection and control have also been conducted using terrestrial optoelectronic sensors (Andújar et al., 2011) and using crop-plant precise spatial position (Sun et al., 2010), colour, shape or spectral signature features to discriminate between weeds and crops (Wendel et al., 2016; Zhang et al., 2012; Romeo et al., 2013; Peteinatos et al., 2014). Distance sensors such as ultrasonic and light detection and ranging (LiDAR) have been used mounted on terrestrial agricultural platforms not only as navigation and mapping elements but also for crop height and biomass measurements (Andújar et al., 2016), harvesting damage evaluation (Martinez-Guanter et al., 2017), characterisation of crop structures (Underwood et al., 2015; Escolá et al., 2015), and variable application of agrochemicals based on canopy dimensions (Llorens et al., 2010). On the other hand, aerial data collection has undergone a considerable change with the growth of UAVs, which have given birth to new, powerful sensor-bearing platforms for various agricultural applications. The ability of UAVs to fly at low altitudes carrying airborne sensors, allows data acquisition with both ultra-high spatial and spectral resolutions (Pajares, 2015), while UAVs also provide easy and fast mission design, Introduction 15 reusability, and cost-effectiveness, and open-source frameworks are specifically designed for UAVs (Urbahs et al., 2013). UAVs are a technology capable of obtaining and conveying agronomically relevant data: multispectral or hyperspectral airborne sensors allow tasks such as mapping vegetation cover (Torres-Sánchez et al., 2014), deriving vegetation index data (Agüera et al., 2015; Berni et al., 2009), field phenotyping and estimating nitrogen (Zaman-Allah et al., 2015), reconstructing 3D orchards (Nevalainen et al., 2017; Díaz-Varela et al. 2015), estimating topsoil moisture (Hassan-Esfahani et al., 2014) or early detection of diseases using chlorophyll fluorescence (Zarco-Tejada et al., 2012; West et al., 2017). As detailed in the subsections 1.2 and 1.3, multispectral airborne sensors have also been used in weed detection and mapping (Peña et al., 2013), while infrared thermal cameras have been employed to determine plant-water status and plant physiological conditions (Gonzalez- Dugo et al., 2015) through sensing approaches that must be assessed with plant-truth data, e.g., stomatal conductance (gs) and leaf water potential measurements (c) (Gago et al., 2015). However, the use of drones in agriculture is not limited to imaging or scouting of crop diseases, pests, weeds or water deficits. Recently, drones have been used to perform tasks such as selective ultra-low-volume herbicide application (Zhang et al., 2016; Giles et al., 2015; Huang et al., 2009) or remotely aerial controlled seeding and reforesting (Wired, 2015). In this thesis, the use of low-cost remote sensors mounted on aerial and terrestrial platforms for different crop management tasks in row crops is addressed. The suitability of implementing this type of sensors to carry out sustainable agricultural activities on relevant crops in the Andalusian region was tested in comparison to commercial sensors of higher cost. Thus, affordable systems that will aid the advancement of automated weed control in tomato crops (Solanum lycopersicum L.) were developed. In addition, the use of infrared sensors mounted on UAVs to monitor water stress in sugar beet (Beta vulgaris L.) and to assess the crop water stress index in super-high-density olive groves (Olea europaea L.) was explored. The following subsections present in detail the advances that have been made in weed control and in the use of thermography as a tool for efficient irrigation water management. 1.2 Advances in Automated Weed Control Infestations of crop weeds are a ubiquitous annual threat to productivity and must be minimised to ensure global food supply and food security. In today’s context, where Introduction 16 agricultural labour costs are rising, the use of a human workforce in heavy, repetitive and potentially health-damaging tasks such as manual weed control is bound to decline dramatically (Perez-Ruiz et al. 2014). Therefore, the emergence and advantages of synthetic herbicides over other control methods have made them a "technology" on which there is almost exclusive reliance for weed control; these compounds are highly effective but are not a completely suitable solution to the complex challenge that weeds present (Harker et al., 2013). Their overuse, intended to be addressed through the implementation of agri-environmental measures at the policy level, has led to a rapid onset of herbicide-resistant weeds, motivating researchers to propose a multifaceted approach in order to mitigate the evolution of herbicide resistance (Norsworthy et al, 2012), including diversification of weed-control techniques. Automatic weed control can contribute to a reduction in both hand weeding and herbicide use, allowing the field deployment of other implements that have lower development costs than new herbicide active ingredients (Kraehmer et al., 2014). 1.2.1 Automatic weed-control systems. Recent advances in mechatronics regarding electronic controllers and actuators, and their progressive implementation in agricultural fields, have enabled the development of effective labour-saving technologies for automatic weed control. These automatic systems can be divided into two major types: those that act by physically removing the weed, and those that apply herbicide accurate and selectively (Fennimore et al., 2016). Among the physical removal systems, we can distinguish between those involving mechanical means or those that use methods such as thermal control, abrasion or other types. On the one hand, traditional mechanical control systems such as cultivators have proven their effectiveness in eliminating weeds between rows of crop. However, intra-row mechanical weed removal without damaging the crop remains a challenge on which producers and researchers are still working. From these studies regarding intra-row weeding, the development of advanced intelligent cultivators (ICs) has emerged: Perez- Ruiz et al. (2012) designed a system based on two cutting blades travelling below the soil and between transplanted tomato plants along the row-centreline. An intra-row weeding roller mechanism was designed and developed by Saber et al. (2013). Other examples of ICs are the intra-row weeding hoe developed by Rasmussen et al. (2012), where rotating tines are moved within the crop row, or the rotating disks designed and tested by Dedousis et al. (2007). Introduction 17 At present, there are only a few commercially available ICs, such as the Robocrop InRow system (Garford Farm Machinery Ltd, England) that uses a variable-speed rotating, semicircle-shaped disc blade for hoeing, and has been tested in celery, lettuce, and radicchio crops in California (Fennimore et al., 2014). Other examples of commercial ICs are the Robovator System (F. Poulsen Engineering ApS, Denmark), which employs a set of electronically controlled tines and which has been tested in lettuce and broccoli cultivars (Lati et al. 2016), and the finger weeder system developed by Steketee and tested in lettuce (Hemming et al., 2011). Most of these commercial systems are actually being semi-experimentally used in horticultural row-grown crops, where their capabilities and benefits are still being tested. Other methods for physical weed control are in their earlier developmental stages but on their way to becoming commercially robust systems (Fennimore et al., 2016). The experimental use of on-the-go automated flamers such as the cross-flaming implement in the RHEA project (Frasconi et al. 2014) and the flaming implement mounted on the commercial IC Robovator System (F. Poulsen Engineering ApS, Denmark) are examples of reliable field-deployable machines that burn weeds where detected. Non-chemical weed control has also been tested using laser technology, where the laser beam must heats the growing point of the weed (apical meristem) to stop its development eliminating it (Blackmore, 2014). Kaierle et al. (2013) performed trials with lasers of different wavelengths, while other research groups have mounted weeding lasers in robotic systems (Shah et al., 2015) or on robotic arms (Zhenyang et al., 2013). The efficiency of this type of laser system depends to a large extent on the accuracy of locating the apical meristem and on the exposure time, which limits its use in mobile systems at normal working speeds (1-2 km.h-1). Effective control using these systems also depends on laser power and wavelength and the weed species (Fennimore et al., 2016). Another approach to physical weed control involves abrasion by the projection of sand or residues of agricultural products on the apical meristem using a stream of highpressure air (Wortman, 2014; Forcella, 2013; Forcella, 2012; Forcella, 2009). Currently, our research group is working on the feasibility of an implement capable of performing this abrasive weeding automatically. On the other hand, automatic precision spraying is a form of chemical control that seeks to apply herbicides accurately and selectively. In this type of spatially selective applications, targeted areas may be on the scale of square centimetres, many orders of magnitude smaller than conventional, traditional herbicide application. Since the early Introduction 18 2000s, several researchers have developed robotic implements or platforms for precision spraying, obtaining acceptable results closely linked to low travel speeds (Young and Giles, 2014; Fennimore et al., 2016). Precise application systems such as the LettuceBot (BlueRiver, California) use a series of sensors to detect unwanted plants (see section 1.2.2) and can achieve a success rate up to 90% in detection and application using an intelligent nozzle control system. The steps in assembling this type of equipment in a conventional tractor were provided as part of the framework of the European project RHEA, in which an intelligent implement was designed and constructed that is capable of performing herbicide micro-applications in previously detected weed patches (Pérez-Ruiz et al., 2015). In addition, one of the systems that are experiencing a major boom in recent years regarding the automatic application of herbicides is that based on aerial platforms performing ultra-low-volume treatments on selected weed patches (Krishna, 2016; Zhang et al., 2016). 1.2.2 Automatic weed-detection systems. Automatic weed-control systems such as those proposed in the previous subsection rely on an accurate detection of weeds, and the reliable discrimination between weeds and crop plants. Spatial distribution and plant spacing are considered key parameters for characterising a crop. Thus, the detection system is intended to collect information on target areas and make weed-control decisions. To this end, Sun et al. (2010) and Pérez- Ruiz (2012) proposed an approach using real-time kinematic (RTK) global navigation satellite systems (GNSS) to identify the exact position of each transplanted tomato plant in a field in a way that any subsequent task (including weed control) should avoid these coordinates. This approach has the advantages of requiring less computing resources or certain lighting conditions. However, the field tests carried out have shown differences between RTK-GNSS-based expected seed location versus actual plant position with an uncertainly of approximately 3 cm for seeds and tomato plants. Today, the predominant technology for plant detection is computer vision, which it represents a robust and powerful method to discriminate between crops and weeds. Differentiation under field conditions becomes difficult because highly variable natural objects must be discriminated from a background with its own highly random characteristics. A variety of visual characteristics can be used for plant species identification such as morphology, spectral reflectance and visual texture (Slaughter et al., 2008). Several steps are involved in the process of a computer vision-based sensing system, beginning with image acquisition and subsequent image segmentation, features extraction (shape, colour, texture…) and final decision making (Peteinatos et al., 2014). Introduction 19 Multispectral and hyperspectral cameras on land platforms have been widely used to detect and locate weeds in lettuce and tomato crops (Zhang and Slaughter, 2011; Piron et al., 2009), even on commercial platforms such as the aforementioned IC Robovator System (F. Poulsen Engineering ApS, Denmark), and have achieved 90% effective discrimination rates. As noted above, the use of these multispectral and hyperspectral sensors has also generated increasing interest in the growth of aerial platforms for remote sensing. Weed detection and location from UAV images have been extensively documented (Lopez- Granados et al., 2016; Torres-Sánchez et al., 2013) and demonstrate a greater spectral differentiation between crops and weeds. In addition to the use of 2D images, the price reduction of the 3D image sensors has motivated researchers to use them for the location and discrimination of weeds or crop plants (Gai et al., 2015). Considering the structure of weeds, in most computer vision-based systems developed in recent years, an approach based on the combined use of images and machine learning is used to extract the morphological characteristics of weeds and classify and discriminate them from the crop. In this application, Bayesian classifiers have been employed (De Rainville et al., 2014), and classification algorithms based on Random Forest (Haug et al., 2014) and pattern-matching algorithms have been developed (Priyadharsini et al., 2015). This is a promising field, and more accurate classifiers capable of working under field conditions will be developed in next years. Despite their advantages, computer vision systems still have some drawbacks: the use of delicate cameras in ground environments where there are external elements such as dust, mud, etc., may present difficulties for the acquisition of high-quality images. These systems also require complex hardware systems and environments with adequate lighting, which increases the costs of such systems. In addition, in agricultural environments, it is necessary to define suitable ground truth data, as machine learning systems have to be fed datasets that allow them to differentiate between different plant types. Other sensors used in proximal sensing for the location of the plants and their classification are LiDAR (Light Detection and Ranging) sensors. These devices measure the distance between the sensor and the target by the emission and reception of an infrared light pulse, also capturing the intensity of the reflected signal, allowing one to distinguish between different elements of vegetation and to differentiate vegetation from the soil. LiDAR sensors provide the morphological structure of the plant/weed, allowing them to be characterised (Llop et al., 2016), defining the geometrical structure of the crop Introduction 20 lines (Andújar et al., 2013) and performing tailored treatments on crops based on their structure (Miranda-Fuentes et al., 2016). In addition to the present work, other authors have explored the use of inexpensive sensors to perform scanning tasks similar to those of LiDAR for the purpose of using them in phenotyping systems of sugar beet varieties (Paulus et al., 2014). There are other sensors for weed detection that are not evaluated in the present work but have been studied in a relevant way by other authors, including fluorescence, X-ray or gamma sensors and ultrasonic sensors. Chapter 3 of this thesis explores the use of optical sensors such as inexpensive infrared light-beam sensors and affordable vision systems compared to the use of LiDAR sensors for the measurement of space between plants as a preliminary step to perform tasks such as intra-row physical weed removal among tomato crop plants. 1.3 Crop Water Stress, Aerial Thermography and Smart Water Management. Climate change has caused changes in seasonal patterns and triggered periods of more intense drought in recent years. Therefore, there is a special interest in drought prone areas in improving water management through new strategies, among which regulated deficit irrigation (RDI) has stood out in recent years. Through the use of new technologies based on remote sensing, it has been proposed to study the potential of thermal imaging and infrared thermometry to monitor the effects of this type of management in sugar beet crops and super high-density olive orchards. Leaf or canopy temperature (TC) has been defined as an indicator of the crop water status (Tanner, 1963; Jackson et al., 1977). A higher leaf temperature is a plant’s physiological response to water stress. Plant moisture stress occurs when the demand for water exceeds available soil moisture. When this occurs, the stomata close to minimise water loss through evapotranspiration. This stomatal closure generates a decrease in the latent heat flux, and because the energy flow must be balanced, there is an increase in the sensible heat flux (increase in leaf temperature). Therefore, the difference between Tc and Ta (air temperature) is linearly related to evapotranspiration under stable conditions (Monteith, 1981) and has been frequently used to detect drought stress and to evaluate the response of a plant to variable irrigation regimes (García - Tejero et al., 2011). In spite of this, Tc may not be completely adequate to evaluate plant-water relations accurately; temperature variations may be due to water stress, but they are also affected Introduction 27 Information-to-action decision processes for field operations will be largely conditioned by the accuracy of the initial data taken in situ (Kitchen 2008), as well as the ability of the DSS to transform this data into knowledge and include it in farm management information systems (FMIS) (Fountas et al., 2015). In the near future, FMIS will play an important role in the adoption of precision technologies, in order to establish the interrelationships between farm machinery or irrigation systems and their surroundings, through proximal or remote sensing. 28 29 II. Overall Summary of the Objectives This doctoral research focuses on the use of proximal sensors in row crops with the purpose of contributing to physical weed-control techniques and managing water resources for optimum crop development. The general objective of the thesis is the analysis, design, implementation and validation of the use of affordable sensors, mounted on aerial and terrestrial platforms, for i) for the study of the response to water stress in sugar beet and super-intensive olive orchards and ii) to accurately measure plant spacing in transplanted tomato row crops. To validate of these types of sensors, the systems developed and tested must meet following requirements; i) achieving acceptable accuracy (compared to commonly used higher cost sensors), ii) requiring less computing power, and iii) being generic enough to be adapted to work under different conditions. A series of specific objectives are proposed to achieve this global objective, which are: 1. To develop an affordable, robust system for infrared temperature measurement based on open-source software/hardware. 2. To determine, under the same field conditions, the effectiveness of the methodology by comparing sugar beet canopy temperature measurements taken by the developed system and those from a conventional thermal camera. 3. To evaluate the suitability of the crop water stress index (CWSI) derived from aerial thermal images as an indicator of seasonal dynamics on water status in super-high density olive orchards. 4. To design and evaluate a sensor platform with affordable components, for the accurate detection of tomato row plants and measurement of automated plant spacing. To compare, under the same conditions, the performance of the inexpensive optical sensors (light-beam and Kinect camera) on the platform to LiDAR sensor efficiency. 30 31 III. Publications: Chapter 1 A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet Martínez-Guanter, J.1; Egea Cegarra, G.1; Agüera Vega, J.2; Pérez-Ruiz, M.1 1Area of Agroforestry Engineering, Aerospace Engineering and Fluids Mechanic Department, School of Agricultural Engineering, Universidad de Sevilla, Ctra. Utrera km.1, 41013 Seville, Spain 2 Department of Rural Engineering University of Córdoba Córdoba Spain Published on: Precision Agriculture February 2017, Volume 18, Issue 1, pp 95–110 DOI: 10.1007/s11119-016-9470-9 Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 32 Abstract Increasing agricultural efficiency in a sustainable manner will contribute to feed a growing population under limited land, nutrient and water resources. Water scarcity and the increasing social concern for this resource are already requiring more sophisticated irrigation and decision-support systems. To address the heterogeneity in crop water status in a commercial field, precision irrigation requires accurate information about crops (e.g., crop water status), soil (e.g., moisture content) and weather (e.g., wind speed and vapor pressure deficit). Numerous studies have shown that plant canopy temperature can be used to derive reliable plant water stress indicators, thus making it a promising tool for irrigation water management. However, efficient and cost-effective measurement techniques are still lacking. This paper assesses the potential of infrared thermometry and thermal imaging for monitoring plant water stress in a commercial sugar beet field by comparing canopy temperature data acquired from a conventional thermal camera with an inexpensive infrared sensor, both mounted on a rotary-wing unmanned aerial vehicle (UAV). Measurements were taken at various phenological stages of the sugar beet growing season. Laboratory tests were performed to determine the key features for accurate temperature measurements and flight altitude. Experiments were conducted in 2014 and 2015 in experimental and commercial sugar beet fields in Southwestern Spain to (i) develop an affordable infrared temperature system suitable for mounting on a UAV to obtain thermal information, (ii) compare sugar beet canopy temperature measurements collected with the low-cost platform with those obtained from a conventional thermal camera, both mounted on a rotary-wing UAV, (iii) identify the factors that will limit the use of the low-cost system to derive temperature-based water stress indices. To accomplish these objectives, well-watered and deficit irrigated plots were established. Results indicated that the lightweight canopy temperature system was robust and reliable, although there were some constraints related to weather conditions and delimitation of the area covered by the infrared sensor. Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 33 3.1 Introduction Agricultural production must double by 2050 in order to meet the predicted food demands of the world population (Tilman et al. 2011). However, accomplishing this target will be a significant challenge for researchers and farmers because crop production would have to increase at a rate of 2.4 % per year; the average rate of increase is only 1.3 % (Ray et al. 2013). In the last five years, plant phenotyping techniques based on remote sensing have emerged and may be deployed for in situ screening with a wide range of breeding objectives, including yield potential, adaptation to abiotic (water stress, temperature, salinity) and biotic (susceptibility to pests and diseases) stresses, and even quality traits (Li et al. 2014). Thus, there is growing interest in adapting agricultural machinery and electronic sensors for field-based high-throughput phenotyping (White and Conley, 2013). Potential applications are mainly envisaged for genetic research (e.g., in the detection of quantitative trait loci, QTL) and crop improvement but also include monitoring of the crop response to soil and management variability (i.e. precision agriculture) (Andrade-Sanchez et al. 2014). While the tremendous potential for using agricultural and environmental unmanned aerial vehicles (UAV: fixed-wing and rotarywing platforms capable of carrying different measurement devices) has become evident through multiple applications during the last few years, the challenging requirement profile for both UAVs and imaging sensor units has emerged as more and more farmers are beginning to demand service (e.g., variable-rate N application, finding faults in irrigation systems), and private firms that offer this service for a fee are emerging (Zhang and Kovacs 2012). In contrast to satellite (largely fixed with rigid coverage areas and observation times) or ground-based measurements, the use of a UAV allows coverage of specific areas and obtaining spatially distributed and geometrically high-resolution information on the canopy/ground temperature. Satellites compensate their rigid flight schedule with extremely large coverage areas, while UAVs are capable of flying at any time and much lower altitude, hence collecting imagery at a much higher spatial and temporal resolution (Hunt et al. 2010; Berni et al. 2009a). Many researchers agree that the main disadvantage of many UAVs is the very limited payload. However, light-weight sensor systems are currently being investigated (Crawford et al. 2014; Colomina and Molina 2014). The use of appropriate sensor platforms can support site-specific management strategies by providing information with a very high spatial (centimetre level), temporal or even radiometric resolution (Zhang and Kovacs 2012). These platforms are excellent tools for crop monitoring and they could significantly reduce the Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 34 crop scouting costs for growers (Ehmke 2013). Most importantly, early detection of plant diseases could prevent their spread and minimize crop losses. Xu et al. (2014) and (Li et al. 2014) developed and used a rotary-wing aircraft sensing system that included a spectral imager to automate disease detection and an analysis process for strawberries (Botrytis fruit rot) and citrus (citrus greening disease-Huanglongbing), respectively. Calderón et al. (2013) used a fixed-wing aircraft UAV to acquire hyperspectral and thermal imagery as an indicator of Verticillium Wilt infection and severity in olive orchards. Another recent and growing use of this new technology is to generate suitable weed maps early in the growing season, a task that has not been possible with conventional aerial or satellite imagery due to the lack of sufficient image spatial resolution (Lopez-Granados 2011; Peña et al. 2014). In addition, UAVs can work on demand with great flexibility at critical times according to agronomic goals, which is crucial for detecting small weed seedlings in the majority of crops. The UAV industry has identified precision agricultural applications, including weed control and resource management, as the single largest market opportunity through the year 2015 (Jenkins and Vasigh 2013). Today, the increasing demand for water resources and increasing environmental concern render plant-sensing devices valuable for irrigation decision-making, such as the use of thermal sensors to generate accurate crop water stress index (CWSI) information (Jackson et al. 1981). This index has been widely used as a crop water status indicator and provides the crop stress level based on canopy-air temperature differences. Many authors have reported morpho-physiological responses of sugar beet to water stress conditions (Hoffmann 2014; Romano et al. 2013; Tsialtas and Maslaris 2012), which is widely considered to be the major limiting factor for yield (Pidgeon et al. 2006). The leaf temperature of sugar beet is considered to be a relevant indicator of drought stress (Shaw et al. 2002). Testi et al. (2008) concluded that more repeatable and effective CWSI values for evaluating tree water status for irrigation purposes were obtained from 1200 to 1500 h (local time), when the CWSI of stressed pistachio canopies reached their maximum diurnal value. To evaluate spatial variations in canopy temperature and relate them to the plant transpiration rate and drought stress, passive sensors (i.e., sensors that measure the characteristic radiation of an object) that provide detailed and highly refined information can be used (Mahlein et al. 2012). Thermography techniques can be applied to proximal or remote sensing. One major advantage of infrared thermal imaging is its non-invasive nature (Lenthe et al. 2007), which enables the use of airborne sensors, as was performed in this work. Thermal imagery used for this objective might provide a suitable tool for model evaluation under variable atmospheric conditions (Ahrens et al. 2014) and have been used for several crops, such Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 35 as cotton, soybean, wheat, citrus and grapes (O’Shaughnessy et al. 2011; Ballester et al. 2013; González-Dugo et al. 2013). However, the main concern with these thermal imagery techniques is their cost, especially when compared to the accuracy obtained, and the resolution is frequently insufficient for precision applications. For example, the system employed by (González-Dugo et al. 2013) is advertised to have a thermal image resolution of approximately 1 ºC, which may not be sufficient for measurements taken at times when there is a small range in canopy temperature differences (e.g., mornings). This hurdle can be overcome by using infrared point sensors, as in this work, which can provide an accuracy of 0.5 ºC at more affordable prices than conventional lightweight thermal cameras. The cost of an aerial thermal platform based on infrared point sensors of the type described in this work may represent 3–5 %, or even lower, of the cost of a conventional lightweight thermal camera, making this technology better suited for small to medium-sized farms that cannot afford costly equipment. The objective of this research was to design and validate a LIghtweight canopy TemperAture measurement System (LITAS) that can be mounted on a rotary-wing UAV with a limited payload of 2 kg maximum. The specific objectives were: (i) to develop an inexpensive infrared temperature system suitable for mounting on a UAV platform to obtain thermal information. (ii) to compare sugar beet canopy temperature measurements collected using LITAS with those obtained from a conventional thermal camera, both mounted on a rotary-wing UAV. (iii) to identify the factors that will limit the use of LITAS to derive temperaturebased water stress indices. 3.2 Materials and methods Canopy temperature measurement platform Canopy temperature measurements were collected by means of two remote sensing devices. These devices consisted of an inexpensive infrared sensor (Melexis, Ypres, Belgium) and a thermal camera (FLIR, Oregon, USA). The thermal camera was used as benchmark to determine whether the proposed low-cost sensing system may provide sufficiently accurate estimates of canopy temperature for sugar beet irrigation management. The inexpensive infrared sensor used an Arduino-based data logger. The infrared sensor, camera and data logger were mounted on a UAV and are described in detail below. The microcontroller board Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 36 The microcontroller (model ATmega32u4, Aduino, Duemilanove, Italy) is an open source board, which has 20 digital input/output pins (of which 7 can be used as pulse width modulation (PWM) outputs and 12 as analog inputs). This microcontroller has built-in USB communication, eliminating the need for a secondary processor and allowing the microcontroller to take control of the USB ports (using libraries). The small dimensions and weight of this microcontroller facilitated mounting on the UAV platform without affecting the navigation performance (Table 3.1). Table 3.1 Microcontroller specific features Arduino board features Microcontroller: ATmega32u4 Operating Voltage: 5 V Input Voltage (recommended): 7-12 V Input Voltage (limits) 6-20 V Analog input channels 12 DC current per I/O Pin: 40 mA Flash memory: 32 KB (ATmega 32a4) of which 4KB used by bootloader EEPROM: 1 KB Clock Speed: 16 MHz Length, width, weigth: 68.6 mm. 53.3 mm, 20 g Infrared sensor specifications An infrared thermometer, model MLX90614, was tested over a reference surface at constant temperature in the laboratory, before taking temperatures of the sugar beet canopy. Table 3.2 shows the specifications of the sensor. The sensor was selected because of its small size and its high accuracy-to-cost ratio. The sensor kit consists of a long-wave filter that allows infrared radiation (with wavelengths of 5.5–14 µm) to pass through, a thermopile detector chip and a unique proprietary signal-conditioning chip. The two voltages, Vs and Vo, correspond to the respective sensor body temperature (Ta) and the object or target temperature (To), which is internally compensated for by Ta. The voltage outputs were converted to temperatures using the following expression (Melexis Data Sheet 2009): 𝑇𝑖(𝐾)= 𝑉𝑗 𝑥 0.02 (1) where Ti represents either To or Ta in Kelvin, being later converted to Celsius degrees and Vj (measured in Volts) is either Vo or Vs. The advantages of this measurement method included the simple automation of data collection and the low-cost of the system components. The infrared sensor provides Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 43 weather conditions of the day of measurement. This is the base for CWSI calculation (Jackson et al. 1981), but this water stress indicator was not calculated in this study because it was outside the scope of the work. As stated in the objectives, the focus of the work was instead to demonstrate that canopy temperature is sensitive enough to water stress in this species and that the envisaged low-cost IR sensing platform (LITAS) is accurate enough to capture these differences. Figure III.4 Changes in canopy-to-air temperature differences (∆T = Tc-Ta) for FI and DI plants throughout the experimental period. Each point indicates the mean of six potted plants. The error bars denote the standard error. The difference in ∆T between the DI and FI plants (which is equivalent to the differences in Tc between the treatments) was related to the RET, a surrogate of soil water deficit (Fig. 3.5). The results indicate that ∆TDI-∆TFI increased with decreasing RET values, thereby confirming the sensitivity of this plant water status indicator to soil water deficit in sugar beets. Earlier work using infrared thermometry on sugar beets infected with Phytium aphanidermatum Edson (Fitz.) (Pinter et al. 1979), a root-rot disease of mature sugar beets, revealed that the midday leaf temperature of infected plants was 3–5 ºC warmer than that of adjacent plants with no sign of disease. Although their work was not focused on irrigation management, water stress is an early symptom of this disease and caused leaf temperature increments of a similar magnitude as those observed in this study for water-stressed plants (Fig. 3.5). The high scatter observed in the ∆TDI-∆TFI Vs RET relationship (Fig. 3.5) can be explained by the influence that other factors (e.g., atmospheric demand represented by air vapor pressure deficit), besides soil water availability, have on Tc (Maes and Steppe 2012). Days after experiment onset 0 2 4 6 8 10 12 14 16 18 20 Tc-Ta (ºC) -4 -2 0 2 4 FI DI Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 44 Figure III.5 Relationship between the difference among ∆T of FI and DI plants and the relative evapotranspiration (RET). LITAS vs. a conventional thermal camera on a UAV The mean and standard deviation values of the surface temperature of the irrigated and deficit-irrigated plots measured with the conventional thermal camera and LITAS during 7 UAV flights performed in the field site are shown in Table 4. The boxplot diagrams of these measurements are shown in Figs. 3.6 and 3.7. The mean canopy temperature (Tc) obtained with the thermal camera for the various flight altitudes varied between 25.60 ºC and 27.11 ºC in the DI plots and between 24.56 ºC and 26.86 ºC in the irrigated plots. The standard deviation (SD) values of the thermal camera measurements were between 1.26 ºC and 2.28 ºC for deficit irrigated plants and between 2.20 ºC and 3.54 ºC for those under normal irrigation regime, both being quite consistent for all flight levels (Table 3.4). The mean LITAS Tc for the various flight altitudes ranged within 26.34 and 27.3 ºC for the DI plots and between 26.30 ºC and 26.62 ºC for the irrigated plots. The LITAS SD values were between 0.47 ºC and 1.31 ºC for plots under deficit irrigation conditions, and between 1.06 ºC and 1.32 ºC for those under normal irrigation regime, both also consistent for all flight levels but lower than those obtained with the thermal camera (Table 4). The similarity in the mean Tc for both the irrigated and DI plots between the thermal camera and the inexpensive IR sensor confirms that LITAS is a promising lowcost alternative to conventional thermal cameras for remote estimations of surface temperature of crops covering the full ground surface. Moreover, the infrared sensor may be mounted on the aerial platform together with an RGB camera that both focus on the same target and record data simultaneously. This allows obtaining canopy temperature measurements from areas of clearly defined observation. Table III.4. Mean IR canopy temperature (Tc) and standard deviation (SD) of the measurements taken with the conventional thermal camera and LITAS. RET (unitless) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 (Tc-Ta)DI-(Tc-Ta)FI -0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 45 Flight Level, m Conventional Thermal Camera LITAS Deficit irrigated Irrigated Deficit irrigated Irrigated Tc (ºC) SD (ºC) Tc (ºC) SD (ºC) Tc (ºC) SD (ºC) Tc (ºC) SD (ºC) h1, 5 26.21 2.05 25.96 2.24 27.30 0.63 26.53 1.25 h2, 10 27.11 1.26 26.86 2.64 27.01 0.65 26.56 1.25 h3, 20 26.07 1.93 26.73 3.54 26.62 0.70 26.62 1.32 h4, 30 25.81 2.28 25.06 2.28 26.34 0.47 26.55 1.08 h5, 40 25.60 1.83 24.56 2.20 27.21 1.31 26.30 1.06 Figure III.6 Boxplot diagrams of mean canopy temperature (Tc) measured at various flight levels with LITAS. Measurements were taken throughout the irrigation season on both a) irrigated and b) deficitirrigated plots. The differences in Tc between the irrigated and DI plots were small (Tc ≈ 0.5 ºC higher in DI plots) and consistent between the two sensing systems employed. These relatively small differences in Tc, compared with the values obtained in the laboratory tests (Fig. 3.5), indicate that the water shortage under DI was not sufficiently severe to trigger more intensive Tc-based crop water stress signals. Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 46 Figure III.7 Boxplot diagrams of mean canopy temperature (Tc) measured at various flight levels with the thermal camera. Measurements were taken throughout the irrigation season on both a) irrigated and b) deficit-irrigated plots. Measurements of Tc with the conventional thermal camera decreased with flight altitude. The maximum mean Tc values were registered at a flight altitude of 10 m (26.86 and 27.11 ºC for the irrigated and DI plots, respectively), while minimum Tc values were obtained at a flight altitude of 40 m (24.56 and 25.60 ºC for the irrigated and DI plots, respectively). The fact that the sensor-target distance can influence remote estimations of surface temperatures is well documented (Jimenez-Munoz and Sobrino 2006; Meier et al. 2011). When remotely sensed from airborne or spaceborne sensors, an atmospheric correction of the thermal infrared imagery has to be performed for accurate determinations of surface temperatures. However, most studies that have focused on atmospheric corrections of remotely-sensed surface temperatures have been conducted for much larger sensor-target distances than those employed in this study (maximum flight altitude of 40 m). In fact, atmospheric effects for flight altitudes with less than ±50 m variation have normally been neglected (Berni et al. 2009b). The results confirm that thermal cameras using uncooled microbolometer sensors require atmospheric corrections to be performed even at flight altitudes as low as those used in this work. These findings agree with the simulated effects of flight altitude and atmospheric conditions on the estimated surface temperature obtained by (Berni et al. 2009b) using the MOTRAN radiative transfer code (Berk et al. 1999). Their simulation outputs showed that for a flight altitude of 40 m, a relative humidity of 60 % and an air temperature within 20/30 ºC, the target to sensor temperature difference may increase to approximately 1.5 ºC, which is in agreement with the Tc reductions of approx. 1.5 - 2 ºC observed in this study. Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 47 Unlike the conventional thermal camera, the surface temperatures estimated with LITAS remained relatively constant with flight altitude (Table 4). The fact that LITAS does not seem to require atmospheric corrections at low altitude of flight may be regarded as an advantage of this system versus thermal cameras, as long as the estimated surface temperatures are sufficiently accurate to derive crop water status indices and that the percentage of ground covered by the crop is almost 100 %. 3.4 Conclusions This work tested a low-cost sensing platform mounted on a small UAV for monitoring canopy temperature. The low-cost platform, called LITAS system, has been tested in a sugar beet field for irrigation management purposes. Laboratory tests showed that canopy temperature (Tc) measurements performed with the low-cost infrared sensor used in LITAS were correlated with Tc measurements obtained with a reference handheld infrared thermometer. Under controlled laboratory conditions, sugar beet plants registered significantly higher leaf-to-air temperature differences in deficit irrigated potted plants than in well-watered plants. Under field conditions, the LITAS system mounted on a small aerial platform captured smaller Tc differences among well-watered and deficit irrigated sugar beet plots, suggesting that the level of water stress developed by the crop was not severe enough to trigger more intense Tc signals. LITAS measurements were compared to Tc derived from thermal images that were recorded simultaneously during the flights. When compared, both systems reported similar Tc differences between wellwatered and deficit irrigated plots, indicating that the LITAS system is stable and a promising low-cost alternative to costlier thermal technologies. A properly timed RGB camera accurately aligned with the infrared sensor can be used to indicate the area where the sensor is monitoring the temperature. Unlike the thermal camera, surface temperature measured with the low-cost infrared sensor was not affected by flight altitude, at least for the atmospheric conditions prevailing within the 0– 40 m atmospheric layer of the experimental site. An important challenge for the inexpensive IR sensing platform is to validate its utility in crops with reduced percentages of ground cover, e.g. fruit tree orchards. Although further research is needed, the combination of low flight altitudes with narrower sensor field-of- views (e.g. there are sensor models with 10º view angle rather than the 90º view angle, as used in this study) could overcome this constraint. Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 48 3.5 References Ahrens, B., Hansson, K., Solly, E. F., & Schrumpf, M. (2014). Reconcilable differences: A joint calibration of fine-root turnover times with radiocarbon and minirhizotrons. New Phytologist, 204, 932–942. Andrade-Sanchez, P., Gore, M.A., Heun, J.T., Thorp, K.R., Carmo-Silva, A.E., French, A.N., Salvucci, M.E., White, J.W. (2014). Development and evaluation of a field-based highthroughput phenotyping platform. Functional Plant Biology 41(1) 68-79. Ballester, C., Jimenez-Bello, M., Castel, J., & Intrigliolo, D. (2013). Usefulness of thermography for plant water stress detection in citrus and persimmon trees. Agricultural and Forest Meteorology, 168, 120–129. Bellvert, J., Zarco-Tejada, P. J., Girona, J., & Fereres, E. (2014). Mapping crop water stress index in a ‘Pinot-noir’ vineyard: Comparing ground measurements with thermal remote sensing imagery from an unmanned aerial vehicle. Precision Agriculture, 15(4), 361–376. Berk, A., Anderson, G., Acharya, P., Chetwynd, J., Bernstein, L., Shettle, E., Matthew, M., Adler- Golden, S. (1999). MODTRAN4 User’s Manual. Hanscom AFB, MA: Air Force Res. Lab. Berni, J., Zarco-Tejada, P.J., Suárez, L., Fereres, E. (2009) Thermal and Narrowband Multispectral Remote Sensing for Vegetation Monitoring from an Unmanned Aerial Vehicle. IEEE transactions on geoscience and remote sensing, 47(3): 722-738. Berni, J., Zarco-Tejada, P.J. Suarez, L., González-Dugo, V., Fereres, E., (2009) Remote sensing of vegetation from UAV platforms using lightweight multispectral and thermal imaging sensors. ISPRS High Resolution Earth Imaging for Geospatial Information, Hannover. Leibniz Universität. June 2-5. Calderón, R., Navas-Cortés, J., Lucena, C., & Zarco-Tejada, P. (2013). High-resolution airborne hyperspectral and thermal imagery for early detection of verticillium wilt of olive using fluorescence, temperature and narrow-band spectral indices. Remote Sensing of Environment, 139, 231–245. Colomina, I., Molina, P. (2014) Unmanned aerial systems for photogrammetry and remote sensing: a review. ISPRS Journal of Photogrammetry and Remote Sensing. 92: 79-97. Crawford, K., Roach, J., Dhillon, R., Rojo, F., Upadhyaya, S. (2014). An inexpensive aerial platform for precise remote sensing of almond and walnut canopy temperature. 12th international conference on precision agriculture ISPA: Monticello, IL, USA https://www.ispag.org/media/ConferenceAbstracts2014_ FINAL.pdf. Ehmke, T. (2013). Unmanned aerial systems for field scouting and spraying. Crops and Soils, 46(6), 4-9. González-Dugo, V., Zarco-Tejada, P. J., Nicolás, E., Nortes, P. A., Alarcón, J. J., Intrigliolo, D. S., et al. (2013). Using high resolution UAV thermal imagery to assess the variability in the water status of five fruit tree species within a commercial orchard. Precision Agriculture, 14, 660– 678. Hoffmann, C. M. (2014). Adaptive responses of Beta Vulgaris L. and Cichorium Intybus L. root and leaf forms to drought stress. Journal of Agronomy and Crop Science, 200, 108–118. Hunt, J. E. R., W. D. Hively, S. Fujikawa, D. Linden, C. S. Daughtry and G. McCarty (2010). "Acquisition of NIRGreen-Blue Digital Photographs from Unmanned Aircraft for Crop Monitoring." Remote Sensing 2(1): 290-305. Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 49 Jackson, R. D., Idso, S. B., Reginato, R. J., & Pinter, P. J. (1981). Canopy temperature as a crop water stress indicator. Water Resources Research, 17(4), 1133–1138. Jenkins, D., Vasigh, B. (2013). The economic impact of unmanned aircraft systems integration in the United States. Association for Unmanned Vehicle Systems International. Arlington, 38pp. Jimenez-Munoz, J.C., Sobrino, J.A. (2006) Error sources on the land surface temperature retrieved from thermal infrared single channel remote sensing data. International Journal of Remote Sensing, 27(5-6): 999-1014. Lenthe, J. H., Oerke, E. C., & Dehne, H. W. (2007). Digital infrared thermography for monitoring canopy health of wheat. Precision Agriculture, 8(1–2), 15–26 Li, H., Lee, W. S., Wang, K., Ehsani, R., & Yang, C. (2014a). Extended spectral angle mapping (ESAM) for citrus greening disease detection using airborne hyperspectral imaging. Precision Agriculture, 15, 162–183. Li, L., Zhang, Q., & Huang, D. (2014b). A review of imaging techniques for plant phenotyping. Sensors, 14(11), 20078–20111. López, A., Molina-Aiz, F. D., Valera, D. L., & Peña, A. (2012). Determining the emissivity of the leaves of nine horticultural crops by means of infrared thermography. Scientia Horticulturae, 137, 49–58. Lopez-Granados, F. (2011) Weed detection for site-specific weed management: mapping and real-time approaches. Weed Research. 51(1):1-11. Maes, W. H., Steppe, K. (2012) Estimating evapotranspiration and drought stress with groundbased thermal remote sensing in agriculture: a review. Journal of Experimental Botany, 63 (13): 4671-4712. Mahlein, A. K., Steiner, U., Hillnhütter, C., Dehne, H. W., & Oerke, E. C. (2012). Hyperspectral imaging for small-scale analysis of symptoms caused by different sugar beet diseases. Plant Methods, 8, 3. Meier, F., Scherer, D., Richters, J., Christen, A. (2011). Atmospheric correction of thermal-infrared imagery of the 3-D urban environment acquired in oblique viewing geometry. Atmos. Meas. Tech., 4, 909–922, Melexis Data Sheet. (2009). MLX90614 family, single and dual zone infrared thermometer inTO- 39. Revised. 5, March 30. Melexis NV, Ieper. O’Shaughnessy, S. A., Evett, S. R., Colaizzi, P. D., & Howell, T. A. (2011). Using radiation thermometry to evaluate crop water stress in soybean and cotton. Agricultural Water Management, 98, 1523–1535. Peña, J. M., Torres-Sánchez, J., de Castro, A. I., López-Granados, F., & Dorado, J. (2014) The TOAS project: UAV technology for optimizing herbicide applications in weed-crop systems. 12th International Conference on Precision Agriculture, Sacramento, California, USA. July 20- 23, 2014. Pidgeon, J. D., Ober, E. S., Qi, A., Clark, C. J. A., Royal, A., & Jagard, K. W. (2006). Using multienvironment sugar beet variety trials to screen for drought tolerance. Field Crop Research, 95, 268–279. Pinter, P. J., Stanghellini, M. E., Reginato, R. J., Idso, S. B., Jenkins, A. D., & Jackson, R. D. (1979). Remote detection of biological stresses in plants with infrared thermometry. Science, 205, 585–587. Chapter 1. A cost-effective canopy temperature measurement system for precision agriculture: a case study on sugar beet 50 Ray, D.K., Mueller, N.D., West, P.C., Foley, J.A. (2013) Yields trends are insufficient to double global crop production by 2050. PLoS ONE 8, Romano, A., Sogona´, A., Lupini, A., Araniti, F., Stevanato, P., Cacco, G., et al. (2013). Morphophysiological response of sugar beet (Beta vulgaris L.) genotypes to drought stress. Acta Physiologiae Plantarum, 35(3), 853–869. Shaw, B., Thomas, T. H., & Cooke, D. T. (2002). Responses of sugar beet (Beta Vulgaris L.) to drought and nutrient deficiency stress. Plant Growth Regulation, 37, 77–83. Testi, L., Goldhamer, D. A., Iniesta, F., & Salinas, M. (2008). Crop water stress index is a sensitive water stress indicator in pistachio trees. Irrigation Science, 26, 395–405. Tilman, D., Balzer, C., Hill, J., Beford, B.L. (2011) Global food demand and the sustainable intensification of agriculture. Proceedings of the National Academy of Sciences. USA. 108, 20260-20264 Tipler, PA. (2000). Physik, 3rd edn. Spektrum Akademischer Verlag, Heidelberg, 1520 pp. Tsialtas, J. T., & Maslaris, N. (2012). Leaf physiological traits and their relation with sugar beet cultivar success in two contrasting environments. International Journal of Plant Production, 6(1), 15–36. Walker, J. (2008) Fundamentals of physics, 8th edn. Wiley, New York, 891pp. ISBN:9780471758013 White, J.W., Conley, M.M. (2013) A flexible, low-cost cart for proximal sensing. Crop Science 53, 1646-1649 Xu, Y., Ehsani, R., Kaplan, J., Ahmed, I., Kuzma, W., Orlandi, J., et al. (2014). An octo-rotor ground network for autonomous strawberry disease detection Year 1 Status Update. In: Proceedings of 2nd international conference on robotics and associated high-technologies and equipment for agriculture and forestry, Madrid. pp. 457-466. Zhang, C., Kovacs, J.M. (2012) The application of small unmanned aerial systems for precision agriculture: a review. Precision Agriculture 13(6) 693-712 51 52 III. Publications: Chapter 2 Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards Egea Cegarra, G.1; Padilla-Díaz, C.M. 2; Martínez-Guanter, J.1; Fernández, J.E.2; Pérez-Ruiz, M.1 1Area of Agroforestry Engineering, Aerospace Engineering and Fluids Mechanic Department, School of Agricultural Engineering, Universidad de Sevilla, Ctra. Utrera km.1, 41013 Seville, Spain 2 Irrigation and Crop Ecophysiology Group, Instituto de Recursos Naturales y Agrobiología de Sevilla (IRNAS, CSIC), Avenida Reina Mercedes 10, 41012, Seville, Spain Published on: Agricultural Water Management June 2017, Volume 187, Pages 210–221 DOI: 10.1016/j.agwat.2017.03.030 Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 59 Four Infrared Remote Temperature Sensors (IRTs) (model IR120, Campbell Scientific Ltd., Shepshed, UK) were mounted over two representative trees of one out of four plots in FI and 45RDITP treatments. The sensors had an angular field of view of 20º (half angle), and the accuracy over the calibrated range was ± 0.2 ºC. The IRTs were mounted on galvanized steel masts with a horizontal mounting arm (model IR1X0, Campbell Scientific Ltd., Shepshed, UK) ending with a white PVC solar shield (model IR-SS, Campbell Scientific Ltd., Shepshed, UK) to protect the sensor. The IRTs were mounted to aim vertically downward (nadir view), targeting the center of the canopy from a distance of approximately 1 m. The dense canopies typical of hedgerow olive orchards allowed the IRTs to view mostly foliage in a circular area of approximately 0.7 m diameter at the top of the canopy. The IRTs were connected to two dataloggers (model CR1000, Campbell Scientific Ltd., Shepshed, UK), which recorded the canopy temperatures (Tc) every minute and stored the 15-min averages. The canopy temperature measurements began on June 16th, 2015 (DOY 167) and continued with a sole interruption of 12 days due to power outage until November 5th (DOY 275). Values of Tc measured with the IRTs above the FI trees were used to derive the Non- Water-Stressed Baselines (NWSB) for CWSI calculation. Only clear-sky days were used for NWSB determination. Clear-sky days following a rainfall event were also discarded to avoid errors associated with wet foliage. Air temperature (Ta) along with VPD data recorded in the orchard every 30 min with a Campbell weather station (Campbell Scientific Ltd., Shepshed, UK) at the same time that Tc were used to derive the NWSB of the SHD olive orchard. 4.2.5. Image processing and CWSI calculation The thermal images taken with the RPAS (Fig. 4.1) were used to calculate the mean canopy temperature (Tc) of each experimental plot. Only the central 8 trees of each plot were used to calculate the mean Tc to avoid border effects. An image segmentation algorithm written in R (R Core Team, 2015) was used to extract pure vegetation pixels from the thermal image. At solar noon, the effects of tree shadow are minimized, and the thermal images are composed mainly of canopy, soil and mixed plant-soil pixels. Firstly, vegetation pixels from a bi-modal histogram (i.e. a histogram with two clearly differentiated peaks ascribed to soil and vegetation pixels in the point cloud) (Fig. 4.2) were selected with the mentioned algorithm. Then, the ‘full width at one-eighth maximum (FWEM)’ rule was used to distinguish the pixels with high probability of being pure vegetation from pixels that were likely to be mixed vegetation with soil and/or shadow effects. The FWEM rule is similar to the ‘full width at half maximum’ (FWHM) rule that Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 60 has been previously used to extract pure vegetation pixels from olive canopies (Rud et al., 2015), but differed in the amplitude of the selected histogram (one-eighth of the maximum in FWEM versus half the maximum in FWHM). FWHM demonstrated a lack of suitability for segmenting thermal images with multiple trees that differ greatly in their Tc, since vegetation pixels from trees with severe water stress can be erroneously discarded by the FWHM rule (Fig. 4.3). The selected segment resulting from the FWEM rule was then used to compute the mean Tc for each experimental plot. Mean Tc was used to calculate the CWSI for each experimental plot using Eq. 4.1. For each day of flight, (Tc-Ta)LL was calculated from the NWSB that was determined with the IRTs, as described in Section 4.2.4 and with the actual air VPD. The value of (Tc-Ta)UL was determined as Ta + 5ºC based on previous studies conducted in different crop species (Irmak et al., 2000; Möller et al., 2007), including olive trees (Agam et al., 2014, 2013; Ben-Gal et al., 2009; Rud et al., 2015). Figure 4.1 Thermal mosaic acquired with a FLIR Tau 2 324 thermal camera on board a RPAS model Phantom 2 observing (left) the experimental SHD olive orchard (white rectangle), and (right) details of the hot and cold reference surfaces. Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 61 Figure 4.2 Example of bi-modal histogram of temperatures obtained from a thermal image of an experimental plot. Figure 4.3 (left) Fraction of a thermal mosaic in which FI and 45RDICC trees have been delimited with white and blue dashed rectangles, respectively. The white circles show the hot reference surfaces. Date of flight: DOY 218; (upper right) thermal image plotted in the left panel with a superimposed layer denoting the vegetation pixels selected by the segmentation algorithm based on the full width at half maximum rule (FWHM); (middle right) thermal image with a superimposed layer showing the vegetation pixels selected by the segmentation algorithm based on the full width at one fifth maximum rule (FWFM); (lower right) thermal image with a superimposed layer showing the vegetation pixels selected by the segmentation Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 62 algorithm based on the full width at one-eighth maximum rule (FWEM). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) 4.2.6. Statistical analyses The relationships between Tc-Ta and VPD (NWSBs) as well as between CWSI and the physiological measurements were analyzed through linear regression analyses. The diurnal time-course of slopes and intercepts of the derived NWSBs was modeled through non-linear regression analysis. In all cases, the coefficient of determination (R2) was used to assess the goodness of fit of the associations among variables. Significant differences between slopes and non-zero intercepts of the NWSBs obtained diurnally and seasonally were evaluated with the Comparison of Regression Lines tool included in the statistical package Statgraphics (Statgraphics Centurion XV). 4.3 Results 4.3.1. Non-Water-Stressed Baselines The relationship between hourly ∆T (Tc – Ta) and VPD values derived for FI olive trees in clear-sky days throughout the study period, did not yield any significant relationship when all hours and days were pooled together (data not shown). The relationships became significant when ∆T and VPD were regressed for a given time of the day, as shown in Table 4.2. The coefficients of determination (R2) were notably affected by the daytime. The highest R2 were observed early in the morning (R2 = 0.74 at 8.00 GMT). These values decreased progressively down to 0.28 (14:00 GMT), recovering partially afterwards (R2 = 0.51 at 18.00 GMT). The intercepts and slopes of the fitted NWSBs also varied throughout the day. The intercepts were higher at midday (2.50 at 12:00 GMT) and lower in the morning and afternoon whereas the slopes showed an increasing trend throughout the day (from -0.77 ºC kPa-1 at 8.00 GMT to around -0.3 ºC kPa-1in the evening). Table 4.2 Fitted parameters for the non-water stressed baselines (Tc-Ta = a + b·VPD). Only clear-sky days from day of year 167–275 were used in the analyses. GMT: Greenwich Mean Time. GMT Intercept ºC Slope ºC kPa-1 R2 8:00 0.94 -0.77 0.74 9:00 0.97 -0.67 0.64 10:00 1.44 -0.61 0.67 Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 63 11:00 2.18 -0.55 0.7 12:00 2.5 -0.36 0.49 13:00 2.43 -0.3 0.32 14:00 2.05 -0.32 0.28 15:00 1.43 -0.3 0.35 16:00 0.83 -0.3 0.41 17:00 0.31 -0.31 0.42 18:00 -0.09 -0.32 0.51 In addition to the diurnal effect on the NWSBs, a marked seasonal effect in the ∆T vs VPD relationship was also observed (Fig. 4.4), as the NWSB shifted in August (Period B) and September (Period C), compared to the values derived in June-July (Period A). When ∆T and VPD of a given time of the day and for a given phenological period were regressed, the level of agreement of the NWSBs increased significantly, and the diurnal effect of R2 observed using pooled data (Table 4.2) was not noticed (Table 4.3). Seasonal variations in the NWSBs were mainly due to significant variations in the NWSB- intercepts, as the NWSB-slopes remained almost invariant throughout the irrigation season (Table 4.3). Table 4.3 Fitted parameters for the non-water-stressed baselines (Tc-Ta = a+b·VPD) determined for three representative periods: A (June-July, day of year −DOY- 167–212), B (August, DOY215–243), C (September, DOY 244–273). Only clear-sky days were used. GMT: Greenwich Mean Time. GMT Period Intercept ºC Slope ºC kPa-1 R2 8:00 A 1.02a −0.79a 0.72 8:00 B 1.15a −1.00a 0.68 8:00 C 1.32a −1.40a 0.75 9:00 A 1.28a −0.73a 0.74 9:00 B 0.97b −0.83ab 0.56 9:00 C 1.79b −1.50b 0.82 10:00 A 1.68a −0.64a 0.77 10:00 B 1.67b −0.89a 0.67 10:00 C 1.91b −1.02a 0.7 11:00 A 2.60a −0.62a 0.82 11:00 B 2.06b −0.58a 0.52 11:00 C 2.84b −0.98a 0.68 12:00 A 3.29a −0.50a 0.82 12:00 B 2.85b −0.54a 0.71 12:00 C 2.86c −0.70a 0.56 13:00 A 3.53a −0.47a 0.82 13:00 B 2.68b −0.45a 0.67 13:00 C 2.98c −0.71a 0.66 14:00 A 3.42a −0.53a 0.81 14:00 B 2.76b −0.59a 0.65 14:00 C 2.35c −0.71a 0.63 15:00 A 2.60a −0.48a 0.82 15:00 B 2.09b −0.54a 0.76 15:00 C 1.67c −0.62a 0.6 16:00 A 1.55a −0.41a 0.8 16:00 B 1.40b −0.49ab 0.65 16:00 C 1.33c −0.67b 0.66 17:00 A 1.11a −0.45a 0.69 17:00 B 0.88b −0.50a 0.77 17:00 C 0.64c −0.68a 0.79 Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 64 18:00 A 0.56a −0.45a 0.79 18:00 B 0.56a −0.64b 0.88 18:00 C 0.01b −0.64ab 0.9 As observed for the NWSBs obtained with pooled data (Table 4.2), the NWSBs derived during the three periods of study (A, B and C) also exhibited a strong diurnal variation (Fig. 4.5). Figure 4.4 Example of seasonal effect on the NWSB (∆T = a + b·VPD). The upper panel (a) shows the ∆T vs VPD relationship when data of the period DOY 167–275 are gathered together. The lower panel (b) shows the same relationship when data are split in three periods: A (June-July, DOY 167–212), B (August, DOY 215–243), C (September, DOY 244–273). Only clear-sky days were used in the calculations. In both panels, data for GMT = 14 h have been used. The straight lines represent the best fit to the data, whose fitted parameters and coefficients of determination are shown in Tables 2 (panel a) and 3 (panel b). Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 65 Figure 4.5 Diurnal variation of the NWSB (∆T = a + b·VPD) for the period June–July (DOY 167–212). The upper panel (a) shows the ∆T vs VPD relationship for GMT 8 to GMT 12. The lower panel (b) shows the same relationship for GMT 13 to GMT 18. Only clear-sky days were used in the calculations. The straight lines represent the best fit to the data, whose fitted parameters and coefficients of determination are shown in Table 4.3. The diurnal time-course of the NWSB-intercepts was successfully modeled with fourthorder polynomial equations during the three periods of study (A, B and C) (Fig. 4.6), whereas that of the NWSB-slopes was successfully modeled with second-order polynomial equations (Fig. 4.7). In order to reduce the empiricism and site specificity of these models, the parameters of the NWSBs obtained were regressed against the zenith solar angle (Testi et al., 2008). Figure 4.6 Diurnal evolution of the NWSB-intercept (◦C) for the three different periods of study: A (June– July, DOY 167–212), B (August, DOY 215–243), C (September, DOY 244–273). The lines represent the best-fit to the data. Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 66 Figure 4.7 Diurnal evolution of the NWSB-slope (◦C kPa−1) for the three different periods of study: A (June–July, DOY 167–212), B (August, DOY 215–243), C (September, DOY 244–273). The lines represent the best-fit to the data. While the relation between the NWSB-slopes and zenith solar angle was not significant (data not shown), the NWSB-intercepts showed a tight relationship with solar angle (Fig. 4.8). As depicted in Fig. 4.8a, this relationship showed a marked hysteresis during periods A and B but not in period C. When the dataset was split into morning and afternoon data, it was observed that the relationship NWSB-intercept vs solar angle was season-dependent during the morning hours only (period C differed from A and B) (Fig. 4.8b), whereas no seasonal effect on NWSB-intercept was observed during the afternoon (Fig. 4.8c). Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 67 Figure 4.8. Relationship between NWSB-intercepts and zenith solar angle for the period (a) 08:00-18:00, (b) 08:00-13:00 and (c) 13:00-18:00. 4.3.2. Crop Water Stress Index derived from IRTs and RPAS The canopy temperature measurements obtained with the IRTs and with the thermal camera mounted on the RPAS were used to derive the seasonal dynamics of CWSI in the irrigation treatments (Figs. 4.9a, b). The CWSI obtained from the IRT readings was markedly influenced by the irrigation regime (Fig. 4.9a). In FI, CWSI values ranged within the interval -0.17 to 0.15 throughout the irrigation season, with a mean value of 0.00. During the first water stress period depicted in Fig. 9 (DOY 174-236), the deficit irrigation treatment monitored with the IRTs (45RDITP) exhibited values of CWSI that were much higher than those found in FI trees, up to approximately 0.7 (Fig. 4.9a). During the last water stress period (DOY 258 onwards), the differences in CWSI between FI and 45RDITP were lower, with values close to 0 and 0.2, respectively. The CWSI values determined from aerial thermal imaging in FI, 45RDITP and 45RDICC treatments are depicted in Fig. 4.9b. The FI trees exhibited values of CWSI that ranged from -0.07 to 0.17, with mean seasonal values of 0.04. The 45RDITP and 45RDICC treatments showed similar trends, with maximum values close to 0.8 observed during the first water stress period. The differences in CWSI between FI and 45RDI treatments for DOY 239 and 281 were much smaller and not statistically significant, with mean values close to 0.3-0.4 in 45RDITP and 45RDICC and close to 0 in FI. The seasonal trend of CWSI derived from aerial thermal imaging was similar to that of REW (Fig. 4.9c), which also revealed significant differences between FI and the 45RDI treatments during the first water stress period but not afterwards. 4.3.3. Relationship between CWSI and other plant-based water status indicators The CWSI values derived from aerial thermal imaging for the five days of flight and the three irrigation treatments were plotted against stem water potential (  st), leaf water potential (  l), stomatal conductance (gsm) and leaf transpiration rate (Em) (Fig. 10). In all cases, significant (P <0.01) linear regressions were observed. The goodness of fit of the relationship between  st and  l with CWSI were similar, with coefficients of determination close to 0.7 (Figs. 4.9a, 4.9b). The relationship between leaf transpiration rate (Em) and CWSI was somewhat weaker than that previously described for  st and  l, with a coefficient of determination of 0.6 (Fig. 4.10d). Stomatal conductance (gsm) was the physiological variable that exhibited the tightest linear relationship with CWSI, with a coefficient of determination of 0.91 (Fig. 4.10c). Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 68 Figure 4.9 Seasonal time-course of (a) CWSI determined from canopy temperature measured with the infrared thermometers (IRTS) at 12.00 GMT, (b) CWSI derived from RPAS thermal imaging (12.00 GMT) and (c) REW. The NWSB shown in Table 4.3 for periods A, B & C were used in CWSI calculation. In (a), only clear-sky days were used for CWSI determination. Each point is the mean of two (a), four (b) and three (c) replicates per treatment. Down-facing arrows indicate the onset of water stress periods in RDI treatments; the up-facing arrow indicates the end of a water stress period. The error bars represent the standard error of the mean (SE). In (b) and (c), asterisks denote significant differences at P < 0.05. 4.4. Discussion The diurnal variation observed in the non-water-stressed baselines (NWSBs) derived for olive trees (Fig. 5) has already been described in other woody species, such as pistachio (Testi et al., 2008) and grapevine (Bellvert et al., 2014). Both of these studies found that diurnal shifts in the NWSBs were mainly due to variations in the NWSB-intercept, as the slope of the baselines was rather stable. As explained theoretically by Jackson et al. (1981), the intercept of the NWSB is expected to increase with solar radiation and to decrease with wind speed. decrease with wind speed. In our case study, diurnal variations of the NWSB-intercept due to wind speed are unlikely since the experimental area is not particularly windy (diurnal wind speed was below 2 m s-1 most of the days) and the effect of low to moderate wind speed on NWSB has been reported to be negligible (Testi et al., 2008), although it has to be noted that was in a different species with different aerodynamic conductance for sensible heat flux. Chapter 2. Assessing a crop water stress index derived from aerial thermal imaging and infrared thermometry in super-high density olive orchards 75 Jones, H.G., 2007. Monitoring plant and soil water status: Established and novel methods revisited and their relevance to studies of drought tolerance. J. Exp. Bot. 58, 119–130. doi:10.1093/jxb/erl118 Jones, H.G., 1992. Plants and microclimate: a quantitative approach to environmental plant physiology. Cambridge University Press. Maes, W.H., Steppe, K., 2012. Estimating evapotranspiration and drought stress with groundbased thermal remote sensing in agriculture: a review. Jorunal Exp. Bot. 63, 4671–4712. doi:10.1093/jxb/ers165 Möller, M., Alchanatis, V., Cohen, Y., Meron, M., Tsipris, J., Naor, A., Ostrovsky, V., Sprintsin, M., Cohen, S., 2007. Use of thermal and visible imagery for estimating crop water status of irrigated grapevine. J. Exp. Bot. 58, 827–838. doi:10.1093/jxb/erl115 Nielsen, D.C., Anderson, R.L., 1989. Infrared thermometry to measure single leaf temperatures for quantification of water-stress in sunflowe. Agron. J. 81, 840–842. Padilla-Díaz, C.M., Rodriguez-Dominguez, C.M., Hernandez-Santana, V., Perez-Martin, A., Fernández, J.E., 2016. Scheduling regulated deficit irrigation in a hedgerow olive orchard from leaf turgor pressure related measurements. Agric. Water Manag. 164, 28–37. doi:10.1016/j.agwat.2015.08.002 R Core Team, 2015. R: A language and environment for statistical computing. R Foundation for Statistical Computing. Vienna, Austria. Rius, X., Lacarte, J.M., 2010. La revolución del olivar. El cultivo en seto. Locator Maps PTY. LTD. Rud, R., Cohen, Y., Alchanatis, V., Beiersdorf, I., Klose, R., Presnov, E., Levi, A., Brikman, R., Agam, N., Dag, A., Ben-Gal, A., 2015. Characterization of salinity-induced effects in olive trees based on thermal imagery. Precis. Agric. 2015 - Pap. Present. 10th Eur. Conf. Precis. Agric. ECPA 2015 511–517. Scholander, P.F., Hammel, H.T., Bradstreet, E.D., Hemmingsen, E.A., 1965. Sap Pressure in Vascular Plants: Negative hydrostatic pressure can be measured in plants. Science (80-. ). 148, 339–346. doi:10.1126/science.148.3668.339 Sepulcre-Cantó, G., Zarco-Tejada, P.J., Jiménez-Muñoz, J.C., Sobrino, J.A., Miguel, E. De, Villalobos, F.J., 2006. Detection of water stress in an olive orchard with thermal remote sensing imagery. Agric. For. Meteorol. 136, 31–44. doi:10.1016/j.agrformet.2006.01.008 Testi, L., Goldhamer, D.A., Iniesta, F., Salinas, M., 2008. Crop water stress index is a sensitive water stress indicator in pistachio trees. Irrig. Sci. 26, 395–405. doi:10.1007/s00271-008-0104-5 Xie, Y., Sha, Z., Yu, M., 2008. Remote sensing imagery in vegetation mapping: a review. J. Plant Ecol. 1, 9–23. doi:10.1093/jpe/rtm005 Zimmermann, D., Reuss, R., Westhoff, M., Gessner, P., Bauer, W., Bamberg, E., Bentrup, F.-W., Zimmermann, U., 2008. A novel, non-invasive, online-monitoring, versatile and easy plant-based probe for measuring leaf water status. J. Exp. Bot. 59, 3157–67. doi:10.1093/jxb/ern171 76 77 III. Publications: Chapter 3 Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: Two Scenarios Martínez-Guanter, J.1; Garrido-Izard, M.2; Valero, C. 2; Slaughter, D.C.3 Pérez-Ruiz, M.1 1Area of Agroforestry Engineering, Aerospace Engineering and Fluids Mechanic Department, School of Agricultural Engineering, Universidad de Sevilla, Ctra. Utrera km.1, 41013 Seville, Spain 2 Laboratorio de Propiedades Físicas (LPF_TAGRALIA), Universidad Politécnica de Madrid (UPM), 28040 Madrid, Spain 3 Department of Biological and Agricultural Engineering, University of California, Davis, CA 95616, USA; Published on: Sensors May 2017, Sensors 2017, 17, 1096 DOI: 10.3390/s17051096 Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 78 Abstract: The feasibility of automated individual crop plant care in vegetable crop fields has increased, resulting in improved efficiency and economic benefits. A systems-based approach is a key feature in the engineering design of mechanization that incorporates precision sensing techniques. The objective of this study was to design new sensing capabilities to measure crop plant spacing under different test conditions (California, USA and Andalucía, Spain). For this study, three different types of optical sensors were used: an optical light-beam sensor (880 nm), a Light Detection and Ranging (LiDAR) sensor (905 nm), and an RGB camera. Field trials were conducted on newly transplanted tomato plants, using an encoder as a local reference system. Test results achieved a 98% accuracy in detection using light-beam sensors while a 96% accuracy on plant detections was achieved in the best of replications using LiDAR. These results can contribute to the decision-making regarding the use of these sensors by machinery manufacturers. This could lead to an advance in the physical or chemical weed control on row crops, allowing significant reductions or even elimination of hand-weeding tasks. Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 79 5.1. Introduction Precision agriculture requires accurate plant or seed distribution across a field. This distribution is to be optimized according to the size and shape of the area in which nutrients and light are provided to plant to obtain the maximum possible yield. These factors are controlled by the spacing between crop rows and the spacing of plants/seeds in a row (Klenin et al., 1985). For many crops, row spacing is determined as much by the physical characteristics of agricultural machinery used to work in the field as by the specific biological spacing requirements of the crop (Blas et al., 2013). According to the crop and machinery used, the accuracy of planting by the precision transplanter/seeder to the desired square grid pattern must be adequate for the operation of agricultural machinery in both longitudinal and transverse crop directions. The current designs of vegetable crop transplanters and seeders utilize several uncoordinated planting modules mounted to a common transport frame. These systems use sub-optimal open-loop methods that neglect the dynamic and kinematic effects of the mobile transport frame and of plant motion relative to the frame and the soil. The current designs also neglect to employ complete mechanical control of the transplant during the entire planting process, producing an error in the final planting position, due to the increased uncertainty of plant location as a result of natural variations in plant size, plant mass, soil traction and soil compaction (Prasanna Kumar et al., 2008). Accurately locating the crop plant, in addition to allowing automatic control of weeds, allows individualized treatment of each plant (e.g., spraying, nutrients). Seeking to ensure minimum physical interaction with plants (i.e., non-contact), different remote sensing techniques have been used for the precise localization of plants in fields. For these localization methods, some authors have decided to address automatic weed control by localizing crop plants with centimeter accuracy during seed drilling (Ehsani et al., 2004) or transplanting (Sun et al., 2010; Pérez-Ruiz et al., 2012) using a global positioning system in real time (RTK-GNSS). These studies, conducted at UC Davis, have shown differences between RTK-GNSS-based expected seed location versus actual plant position. The position uncertainly ranged from 3.0 to 3.8 cm for seeds, and tomato transplants, the mean system RMS was 2.67 cm in the along-track direction. Nakarmi and Tang used an image acquisition platform after planting to estimate the interplant distance along the crop rows (Nakarmi and Tang, 2012). This system could measure inter-plant distance with a minimum error of ±30 cm and a maximum error of ±60 cm. Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 80 Today, one of the biggest challenges to agricultural row crop production in industrialized countries is non-chemical control of intra-row (within the crop row) weed plants. Systems such as those developed by Pérez-Ruiz et al. (2014) or the commercial platforms based on computer-controlled hoes developed by Dedousis et al. (2007) are relevant examples of innovative mechanical weeding systems. However, the current effectiveness of mechanical weed removal is constrained by plant spacing, the proximity of the weeds to the plant, the plant height and the operation timing. Other methods for non-chemical weed control, such as the robotic platform developed by Blasco et al. (2002) (capable of killing weeds using a 15-kV electrical discharge), the laser weeding system developed by Shah et al. (2015) or the cross-flaming weed control machine designed for the RHEA project by Frasconi et al. (2014), demonstrate that research to create a robust and efficient system is ongoing. A common feature of all these technological developments is the need for accurate measurement of the distance between plants. Spatial distribution and plant spacing are considered key parameters for characterizing a crop. The current trend is towards the use of optical sensors or image-based devices for measurements, despite the possible limitations of such systems under uncontrolled conditions such as those in agricultural fields. These image-based tools aim to determine and accurately correlate several quantitative aspects of crops to enable plant phenotypes to be estimated (Li et al., 2014; Fahlgren et al., 2015). Dworak et al. (2011) categorized research studying inter-plant location measurements into two types: airborne and ground-based. Research on plant location and weed detection using airborne sensors has increased due to the increasing potential of unmanned aerial systems in agriculture, which have been used in multiple applications in recent years (López-Granados, 2011). For ground-based research, one of the most widely accepted techniques for plant location and classification is the use of Light Detection and Ranging (LiDAR) sensors (Garrido-Izard et al., 2015). These sensors provide distance measurements along a line scan at a very fast scanning rate and have been widely used for various applications in agriculture, including 3D tree representation for precise chemical applications (Rosell et al., 2009; Garrido-Izard et al., 2012) or infield plant location (Shi et al., 2013). This research continues the approach developed by Garrido-Izard et al (2014), in which a combination of LiDAR + IR sensors mounted on a mobile platform was used for the detection and classification of tree stems in nurseries. Based on the premise that accurate localization of the plant is key for precision chemical or physical removal of weeds, we propose in this paper a new methodology to precisely estimate tomato plant spacing. In this work, non-invasive methods using optical sensors such as LiDAR, infrared (IR) light-beam sensors and RGB-D cameras have been employed. For this purpose, a platform was developed on which different sensor Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 81 configurations have been tested in two scenarios: North America (UC Davis, CA, USA) and Europe (University of Seville, Andalucia, Spain). The specific objectives, given this approach, were: -To design and evaluate the performance of multi-sensor platforms attached to a tractor (a UC Davis platform mounted on the rear of the tractor and a University of Seville platform mounted on the front of the tractor). -To refine the data-processing algorithm to select the most reliable sensor for the detection and localization of each tomato plant. 5.2. Materials and Methods To develop a new sensor platform to measure the space between plants in the same crop row accurately, laboratory and field tests were conducted in Andalucia (Spain) and in California (USA). This allowed researchers to obtain more data under different field conditions and to implement the system improvements required, considering the plant spacing objective. These tests are described below, characterizing the sensors used and the parameters measured. 5.2.1. Plant Location Sensors 5.2.1.1. Light-Beam Sensor Specifications IR light-beam sensors (Banner SM31 EL/RL, Banner Engineering Co., Minneapolis, MN, USA) were used in two configurations: first as a light curtain (with three pairs of sensors set vertically, Figure 5.1 central and Figure5. 4) and later a simpler setup, using only one pair of sensors (Figure 5.2), which simplifies the system while still allowing the objective (plant spacing measurement) to be attained. In the light curtain, light-beam sensors were placed transversely in the middle of the platform to detect and discriminate the plant stem in a cross configuration to prevent crossing signals between adjacent sensors. Due to the short range and focus required in laboratory tests, it was necessary to reduce the field of view and the strength of the light signal by masking the emitter and receiver lens with a 3D-printed conical element. In laboratory tests, the height of the first emitter and receiver pair above the platform was 4 cm, and the height of 3D plants (artificial plants were used in laboratory tests; see Section 5.2.2) was 13 cm. In the field tests, the sensor was placed 12 cm from the soil (the average height measured manually for real plants in outdoor tests was 19.5 cm) to avoid obstacles in the field (e.g., dirt clods, slight surface undulations). In both cases, the receiver was set to obtain a TTL output pulse each time Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 82 the IR light-beam was blocked by any part of the plant. The signals generated by the sensors were collected and time-stamped by a microcontroller in real time and stored for off-line analysis. Technical features of the IR light-beam sensors are presented in Table 5.1. Table 5.1. IR light-beam sensor features. Operational Voltage (V) 10–30 V Detection range (m) 30 m Response time (milliseconds) 1 ms Sinking and sourcing outputs (mA) 150 mA 5.2.2.2 Laser Scanner A LMS 111 LiDAR laser scanner (SICK AG, Waldkirch, Germany), was used in the laboratory and field testing platforms to generate a high-density point cloud on which to perform the localization measurements. Its main characteristics are summarized in Table 5.2. The basic operating principle of the LiDAR sensor is the projection of an optical signal onto the surface of an object at a certain angle and range. Processing the corresponding reflected signal allows the sensor to determine the distance to the plant. The LiDAR sensor was interfaced with a computer through an RJ 45 Ethernet port for data recording. Data resolution was greatly affected by the speed of the platform’s movement; thus, maintenance of a constant speed was of key importance for accurate measurements. During data acquisition, two digital filters were activated for optimizing the measured distance values: a fog filter (becoming less sensitive in the near range (up to approximately 4 m)); and an N-pulse-to-1-pulse filter, which filters out the first reflected pulse in case that two pulses are reflected by two objects during a measurement (LMS100 Product Family Operating Instructions, Waldkirch SA). Different LiDAR scan orientations were evaluated: scanning vertically with the sensor looking downwards (Figure 5.1), scanning with a 45° inclination (push-broom) and a lateral-scanning orientation (side-view). Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 83 Figure 5.1. Details of the sensors on the laboratory platform (vertical LiDAR and Light-beam sensors) for the detection and structure of the modular 3D plant. Table 5.2. LMS 111 technical data. Operational Range From 0.5 to 20 m Scanning field of view 270° Scanning Frequency 50 Hz Angular resolution 0.5° Light source 905 nm Enclosure rating IP 67 5.2.1.3. RGB-D Camera A Kinect V2 commercial sensor (Microsoft, Redmond, WA, USA), originally designed for indoor video games, was mounted sideways on the research platform during field trials. This sensor captured RGB, NIR and depth images (based on time-of-flight) of tomato plants, although for further analysis, only RGB images were used for the validation of stick/tomato locations obtained from the LiDAR scans, as detailed in Section 5.2.4.3. Kinect RGB-captured images have a resolution of 1920 × 1080 pixels and a field of view (FOV) of 84.1 × 53.8°, resulting in an average of approximately 22 × 20 pixels per degree. NIR images and depth camera have a resolution of 512 × 424 pixels, with an FOV of 70 × 60° and a depth-sensing maximum distance of 4.5–5 m. Although systems such as the Kinect sensor were primarily designed for use under controlled light conditions, the second version of this sensor has higher RGB resolution (640 × 480 in v1) and its infrared sensing capabilities were also improved, enabling a more lighting-independent view and supporting its use outdoors under high-illumination conditions. Despite this improvement, we observed that the quality of the RGB images were somewhat affected by luminosity and direct incident light, and therefore, the image must be post-processed to obtain usable results. The images taken by the Kinect sensor were simultaneously acquired and synchronized with the LiDAR scans and the encoder pulses. Because the LabVIEW software (National Instruments, Austin, TX, USA) used for obtaining the scan data was developed to collect three items (the scans themselves, the encoder pulses and the timestamp), a specific Kinect recording software had to be developed to embed the timestamp value in the image data. With the same timestamp for the LiDAR and the image, the data could be matched and the images used to provide information about the forward movement of the platform. 5.2.2. Lab Platform Design and Tests Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 84 To maximize the accuracy of the distance measurements obtained by the sensors, an experimental platform was designed to avoid the seasonal limitations of testing outdoors. Instead of working in a laboratory with real plants, the team designed and created model plants (see Figure 5.1) using a 3D printer (Prusa I3, BQ, Madrid, Spain). These plants were mounted on a conveyor chain at a predetermined distance. This conveyor chain system, similar to that of a bicycle, was driven by a small electric motor able to move the belt at a constant speed of 1.35 km·h−1. For the odometry system, the shaft of an incremental optical encoder (63R256, Grayhill Inc., Chicago, IL, USA) was mounted so that it was attached directly to the gear shaft and used to measure the distance travelled, thus serving as a localization reference system. Each channel in this encoder generates 256 pulses per revolution, providing a 3-mm resolution in the direction of travel. The data generated by the light-beam sensors and the cumulative odometer pulse count were collected using a low-cost open-hardware Arduino Leonardo microcontroller (Arduino Project, Ivrea, Italy) programmed in a simple integrated development environment (IDE). This device enabled recording of data that were stored in a text file for further computer analysis. Several repetitions of the tests were made on the platform to optimize the functions of both light-beam and LiDAR sensors. From the three possible LiDAR orientations, lateral scanning was selected for the field trials because it provided the best information on the structure of the plant, as concluded in Garrido-Izard et al., (2015). In lab tests, two arrangements of light-beam sensors were assessed: one in a light curtain assembly with three sensor pairs at different heights and another using only one emitterreceiver pair. 5.2.3. Field Tests The initial tests, performed in Davis, CA (USA), were used to assess the setup of the light-beam sensor system and detected only the stem of the plants rather than locating it within a local reference system. Once the tomato plants were placed in the field, tests were conducted at the Western Center for Agriculture Equipment (WCAE) at the University of California, Davis campus farm to evaluate the performance of the sensor platform for measuring row crop spacing. For this test, an implement was designed to house the sensors as follows. The same IR light-beam sensor and encoder, both described in Section 5.2.1, were used (Figure 5.2). The output signals of the sensors were connected to a bidirectional digital module (NI 9403, National Instruments Co.), while the signal encoder was connected to a digital input module (NI 9411, National Instruments Co.). Both modules were integrated into an NI cRIO-9004 (NI 9411, National Instruments Co.), and all data were recorded using LabVIEW (National Instruments Co.). Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 91 ▪ The squared Euclidean distance measure and the k-means++ algorithm were used for cluster centre initialization. ▪ The clustering was repeated five times using the initial cluster centroid positions from the previous iteration. ▪ Method for choosing initial cluster centroid positions: Select k seeds by implementing the k-means++ algorithm for cluster centre initialization. A reduction in the number of clusters was determined directly by evaluating the cluster centroids. If pair of centroids were closer than the min_distance_between_plants, the process was repeated by reducing the number of clusters by one (Table 5.5). A clustering size evaluation was performed, excluding clusters smaller than min_cluster_size. Table 5.5. Plant identification and stem identification parameters. Minimum Distance between Plants Minimum Cluster Size Histogram Jumps (mm) Distance_between_plants_theoretical×0.2 5 4 (iii) Plant Location Three different plant locations were considered: • Centre of the cluster • Location of the lowest point at each plant cluster • Intersection of the estimated stem line and ground line o By dividing the aerial plant data in slices defined by “histogram jumps”, the x limits on the maximal number of counts were obtained. For the z limits, data belonging to the bottom half of the aerial point cloud were considered. o The remaining data inside these limits were used to obtain a line of best fit, which was considered the stem line. o Plant location was defined as the intersection between the stem line and the corresponding ground line obtained previously from the MSAC function. 5.2.4.3. Validation of Plant Location Using RGB Kinect Images To obtain the distance between the stems of two consecutive plants using the Kinect camera, it is necessary to characterize the plants correctly and then locate the stems with RGB images from the Kinect camera. This characterization and location of the stem was conducted as follows: a sequence of images of the entire path was obtained (~250 Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 92 images in each repetition), where the camera’s shooting frequency was established steadily in at 1-s intervals. Obtaining the string with the timestamp of each image was a key aspect of the routine developed in Matlab, as this string would later be used for integration with the LiDAR measurement. The relationship between the timestamp and its corresponding encoder value was used to spatially locate each image (x-axis, corresponding to the tractor advance). Image processing techniques applied in the characterization of tomato plants from Kinect images generally followed these steps: (i) According to Hamuda et al. (2015), the first step in most works regarding image analysis is the pre-processing of the image. In this work, an automatic cropping of the original images was performed (Figure 5.4a), defining a ROI. Because the test was conducted under unstructured light conditions, the white balance and the saturation of the cropped image (Figure 5.4b) were modified; (ii) Next, the image segmentation step of an object-based image analysis (OBIA) was performed to generate boundaries around pixel groups based on their colour. In this analysis, only the green channel was evaluated to retain most of the pixels that define the plant, generating a mask that isolates them (Figure 5.4c) and classifying the pixels as plant or soil pixels. In addition, morphological image processing (erosion and rebuild actions) was performed to eliminate green pixels that were not part of the plant; (iii) Each pair of consecutive images was processed by a routine developed in Matlab using the Computer Vision System Toolbox. This routine performs feature detection and extraction based on the Speeded-Up Robust Features “SURF” algorithm (Bay et al., 2008) to generate the key points and obtain the pairings between characteristics of the images. Once the plant was identified in each image, the location of the stem of each plant was defined according to the following subroutine: (1) Calculating the distance in pixels between two consecutive images, as well as the encoder distance between these two images, the “forward speed” in mm/pixel was obtained for each image; (2) For each image, and considering that the value of the encoder corresponds to the centre of the image, the distance in pixels from the stem to the centre of the image was calculated. Considering whether it was to the left or to the right of the centre, this distance was designated positive or negative, respectively; (3) The stem location was calculated for each image using the relation shown in Equation (5.2) below; Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 93 (4) Plant location was obtained for each image in which the plant appeared; the average value of these locations was used to calculate the distance between plants: Stem Location = Encoder Value ± Distance (from stem to centre) × motion relation ( mm pixel ) (5.2) Figure 5.4. (a) raw image captured by the Kinect sensor with the ROI indicated; (b) ROI with white balance and saturation adjustment; (c) OBIA resulting image with isolated plant pixels. 5.3. Results 5.3.1. Laboratory Test Results Several laboratory tests were conducted to properly adjust and test the sensors. As an example, Figure 5.5 shows various profiles representing detections, comparing the pair of light-beam sensors placed on the bottom (stem detection) with those placed in the middle, detecting the aerial part of the 3D plant. Detections of the third pair of sensors, placed above the plants, were not considered relevant because most of their beam was above the plants. Figure 5 also shows that the detection of the stem using a single pair of light-beam located at a lower height with respect to the soil was more effective and robust than trying distinguish between different plants based on the detection of the aerial part of the plant. From our point of view and in agreement with the results in Garrido et al., (2014), this justifies the use of a single pair of light-beam sensors for the field tests rather than the use of the curtain mode (3 pairs as originally tested in the laboratory). The algorithm to analyze data from a single light-beam is also simpler and faster, being more adequate for real-time usage. Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 94 Figure 5.5. (Left) Stem detection of the 11 artificial plants through the lower pair of light-beam sensors. (Right) Detection of the aerial part (leaves and branches) of the 11 artificial plants through the centre pair of light-beam sensors. Figure 5.6 shows the positions of the stem detections and the estimated distances between them. In all laboratory tests, 100% of the stems were detected using the lightbeam sensor. Notably, under laboratory conditions, there were no obstacles in the simulated crop line, which does not accurately represent field conditions. Figure 5.6 (right) shows the stem diameter measured from a test, with an estimated average diameter of 11.2 mm for an actual value of 10 mm. (a) (b) Figure 5.6. (a) Positions of the stems of the 11 plants in the laboratory test; (b) Average plant position for a distance of 11 mm. Three laboratory tests corresponding to three different stages of the light-beam sensor adjustment process are shown in Figure 5.7. Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 95 Figure 5.7. Histogram of measured distances between 3d plant stems during the adjusting process of sensors on the detection platform in the laboratory. The histogram in this figure represents the estimated distances between plant stems, where the average distance is 102.5 mm when the real distance between plants was 100.0 mm. The average standard deviation for the three trials was 2.2 mm. For test 3 (last histogram in Figure 5.7), all distance values between plants were estimated between 100.1 mm and 103.8 mm. For the results obtained by the LiDAR in the laboratory, as described in Section 5.2.2, multiple scanning configurations and orientations were tested. After a visual comparison of the results of these tests, lateral-scanning orientation gave the best point clouds, and its plant representation and spacing measurements were therefore more accurate. Figure 5.8 shows the point cloud representation of one of the laboratory scans made using the LiDAR sensor with lateral orientation. Figure 5.8. Point cloud representation of artificial 3D plants obtained using a lateral-scanning LiDAR sensor during laboratory tests. Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 96 5.3.2. Field Tests Results Preliminary tests results using the light-beam sensors are presented in Figure 5.9a. Unlike laboratory detections, in the field data, the determined distances between stems were more variable. This variability is mainly due to crop plants missing from the line or the presence of a weed very close to the stem of the detected plant. Figure 5.9b reveals that several positions for potential plants (marked with a cross) were established, while only one plant was present (marked with a circle). a) b) Figure 5.9. (a) Detections and positions of the 41 tomato plants in first field test with IR sensors; and (b) detail of the positions of potential plants. During the first field test, 41 detections occurred when there were 32 real plants growing in the line; thus, for this trial, the error in the number of detections was 22% (Table 5.6). We speculate three possible causes of this error in the initial test: the plants were very close together, soil clods were detected between plants, or the evaluation of the distance between plants for the selection of potential plants was not optimal. However, in trial two, 32 stem detections occurred when 34 real plants were in the line, resulting in a 94% success rate. For test three, the accuracy was 98%. Table 5.6. Plant detection ratio in field trials. Test5 Real Plants Detected Plants % Accuracy First test 32 41 78 Second test 34 32 94 Third test 48 49 98 Regarding the experiments conducted to estimate the response of the LiDAR system in real field conditions, the detection system was used on three lines (one with wooden Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 97 sticks and two with tomatoes). The system was intended to detect the presence of plant stems correctly and to distinguish between the foliage and the stem. In addition, this system was used to measure the distances between plants accurately, providing valuable information for future herbicide treatments. Because the use of the laser scanner for detection implies the generation of a very dense cloud of points, in which not all data are relevant, an initial filtering was performed as explained above. Table 5.7 presents the number of original points obtained by eliminating the seedbed, and those considered to be representative of the aerial part of the plant, which comprised only 4.5% of the total. Table 5.7. Point cloud reduction during plant point cloud extraction. Test Seedbed delimit Gridding Plant Points Stick: Test 1 44,708 (100%) 40,903 (91.5%) 1667 (3.7%) Tomatoes 1: Test 1 374,963 (100%) 328,593 (87.6%) 14,720 (3.9%) Tomatoes 1: Test 2 237,396 (100%) 220,714 (93%) 13,098 (5.5%) Tomatoes 2: Test 2 3,807,858 (100%) 340,051 (89.3%) 18,706 (4.9%) Regarding plant detection, Table 5.8 summarizes the results of some of the tests, implementing the methodology explained previously. Based on these data, it is observed that 100% of the sticks have been correctly detected, without obtaining false positives or negatives. When this analysis was performed on tomato plants, about 3%–21% of false positive or negative detections were found. A plant detected by the LiDAR was defined as a false positive when the estimated plant centre was more than half of the plant spacing from the real plant centre. Table 5.8. Plant and sticks detection results. Test Plant Location Method Correctly Detected False Positive False Negative Stick with filter Test 1 Centre of the cluster 19 0 0 Lowest point 19 0 0 Intersection of stem line and ground line 19 0 0 Tomatoes 1 with filter Test 1 Centre of the cluster 49 2 6 Lowest point 48 3 7 Intersection of stem line and ground line 46 5 9 Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 98 Tomatoes 1 with filter Test 2 Centre of the cluster 53 6 2 Lowest point 52 7 3 Intersection of stem line and ground line 47 12 8 Tomatoes 2 with filter Test 2 Centre of the cluster 51 5 0 Lowest point 48 8 3 Intersection of stem line and ground line 42 14 9 Related to the data presented in Table 5.8, Figure 5.10 below shows stick and plant LiDAR detection results in three rows. Platform advance information (x-axis on the plot) provided by the encoder is given in mm. Each plant/stick detected is marked with different colour (note the overlapping on some plants, explaining false negative and false positive detections). (a) (b) Figure 5.10. Detection results on three tested lines. Green dotted lines represent the cluster centre and black dotted lines the real plant interval obtained by the Kinect image (location ± Std). (a) 19 stick detections using LiDAR; (b) 51 plants detected during Test 1 in row 1. As explained in Section 5.2.4.2.2, the plant location method based on the point-to-ground intersection has been evaluated (in addition to the lowest-point and centre-of-cluster methods). Figure 5.11 shows an example of the insertion point obtained from the intersection of the two lines (aerial part and soil line) in tomato plants and sticks. Plant locations obtained from the LiDAR and Kinect data are shown in Table 5.9. As explained in Section 5.2.4.3, when processing the Kinect images, a value of the encoder was automatically selected for each image. Mean values are obtained from the difference between the actual location and the location obtained with the LiDAR. The negative mean value obtained for Tomatoes 2 with filter Test 2 for the intersection of the stem and ground line method means that the LiDAR detected the plant at a distance greater than the actual distance (obtained from the Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 99 Kinect). High standard deviation values can be explained due to the high variability found in the encoder values assigned to each plant or stick during the image processing. Figure 5.11. Plant location method based on the intersection of stem line and ground line. (a) Intersection point between the average aerial part of the plant line (green line) and the average ground line (red line); (b) Intersection point between the stick aerial part line (green line) and the ground line (red line). Both histograms of aerial part of the plant points are shown on the bottom. Table 5.9. Mean and standard deviation (Std.) of the plant and stick locations during the tests. Test Plant Location Method Mean (mm) Std. (mm) Stick—Test 1 Centre of the cluster 8.32 10.09 Lowest point 7.25 8.47 Intersection of stem line and ground line 6.15 9.45 Tomatoes 1—Test 1 Centre of the cluster 20.74 40.37 Lowest point 10.06 51.72 Intersection of stem line and ground line 5.34 62.65 Tomatoes—Test 2 Centre of the cluster 30.42 37.03 Lowest point 35.48 52.90 Intersection of stem line and ground line 27.60 50.36 Tomatoes 2—Test 2 Centre of the cluster 1.33 41.49 Lowest point 10.00 61.86 Intersection of stem line and ground line −5.95 53.32 Figure 5.12 shows the plant identifications in each line during the different tests. For each detection previously shown in Figure 5.10, tags for each of the localization methods are shown by colored lines. Chapter 3. Optical Sensing to Determine Tomato Plant Spacing for Precise Agrochemical Application: two scenarios 100 (a) (b) (c) Figure 5.12. Plant and sticks location results obtained by the three different methods: Centre of the cluster; lowest point; and ground-plant intersection. (a) Sticks row data using LiDAR; (b) Tomato plants detected during Test 1 on row 1; (c) Detail of the aerial point cloud of tomato plants generated by the LiDAR during Test 1 on row 1. Each plant location method is marked with different colored dotted line. The centre of the cluster is marked in green, the lowest point in blue, and the intersection in red. Encoder stick/plant locations by the Kinect image (location ± Std.) are both marked in black lines, and their distances are represented by black doted. 5.4. Conclusions A combination of optical sensors mounted on a frame on a tractor, which are capable of detecting and locating plants and utilizing ground-wheel odometry to determine a local reference system, was successfully developed and tested. Combining the use of affordable sensors (light-beam and Kinect) with more expensive ones (LiDAR), a laboratory setup platform and two field test platforms have been created. The following conclusions were drawn based upon the results of this research: • The results obtained from the precise locations of the plants allow us to consider the viability of further methods capable of performing individualized treatments for each of the plants, or accurate agrochemical treatments to remove existing weeds between crop plants in the same row. • A light-beam detection system was improved by reduction of the number of sensor pairs used. The reduction from three pairs to one pair had no effect on the desired plant detection results. In the field tests, 98% precision of detection was obtained, similar to that obtained by Garrido et al. (2014), showing that this 107 General Discussion of the Results 108 V. General Discussion of Results and Future Work Irrigation water management and accurate weed control are currently two of the main concerns in the agricultural sector, as they directly affect both crop quality and yield. Social and political demands show that it is necessary to move towards sustainable management levels, which in turn has motivated the growing interest in precision agriculture in the last two decades. This agricultural approach aims to take advantage of the current technological potential to reduce the efforts from producers, managing to increase the productivity and reduce the impact on the environment. The main contribution of this thesis is the study and validation of the use of affordable sensors, mounted on both aerial and terrestrial platforms, which has been addressed in order to contribute generating management strategies regarding the two major lines of work for this research: irrigation water management and automated weed detection and control, whose overall results based on initial hypotheses are summarized hereunder. Due to the inherent spatial and temporal variability of crop fields, the adoption of technologies capable of conduct site-specific management is necessary as a first step, in order to be able to deal with this type of issues. A large number of sensors and electronic solutions are available for automated crop sensing, monitoring, and collecting information, but several constraints exist which can limit their application in research work and acceptance by farmers. Features and prices of these commercial sensing devices typically range from low-cost sensors with low resolution, limited inputs or weak manufacturing to very expensive, full-featured devices with high resolution measurement capacity. Usually, this type of expensive sensors needs specialized maintenance, while offers the customer a closed and proprietary environment, dependent on a particular brand both for data collection and analysis, building ecosystems that result incompatible with other manufacturers, which could be one of the causes of the low adoption of these technologies. In recent years, rapid advances in consumer electronics, coupled with the growth of the do-it-yourself (DIY) and maker movement, has made fairly inexpensive sensors readily General Discussion of the Results 109 available. At a very low cost (few euros/dollars), standardized sensors can be acquired for any type of monitoring, automation or data logging tasks, interfacing directly with microcontrollers, which simplifies electronic designs and reach people with limited previous knowledge in electronics. In addition to the sensors, this low-cost sensing approach needs that the microcontroller should also be affordable and easy to programme. Arduino is an open-source electronics prototyping platform based on flexible, easy-to-use hardware and software which could be considered a very good alternative to develop precision agriculture oriented sensors, having a big community of users that employs this platform in multidisciplinary projects (Mesas-Carrascosa et al., 2015; USAID, 2016; Fisher and Gould, 2012). The characteristics of irrigated crops in arid and semi-arid zones demonstrate the need to develop techniques that allow the efficient use of water resources, providing the right amount of water at the right time and place. One possibility that offers the use of remote sensing is the development of methodologies that allow to delimit uniform management zones, attending to spatial variability of the soil and crops, while also allows irrigation scheduling based on them. Data obtained through these sensors can be used as an accurate indicator for irrigation management, since through it water status of the plants can be determined; this status is commonly measured in terms of water potential (Jones, 1992), for which the pressure chamber has traditionally been used (Scholander et al., 1965). In this way, the water deficit can be determined from the water potential of the leaves (ψL). Although this method is effective for measuring the water status of the plant, it is very slow and laborious and requires continuous analysis that makes it ineffective (Cohen et al., 2005). For more than 30 years, there are evidences on the usefulness of the measurement of crop temperature to monitor the water status of the crop. When water stress is induced, leaf stomata close, the rate of transpiration is reduced and its cooling effect decreases, which causes leaf temperature to rise. Crop canopy temperature (Tc) is the result of the energy balance between energy gains (air temperature and incident radiation) and losses due to evapotranspiration. Chapter 1 of this thesis addresses the use of a lowcost sensing device for infrared thermometry, in order to determine the temperature of the crop canopy as an alternative method to water potential, for measuring the plant water status. Firstly, laboratory validation of the measurements obtained with the sensor was conducted, comparing them with a conventional thermometer, and an open source hardware/software-based temperature recording system capable of being mounted on board a small multi-rotor UAV was successfully developed. General Discussion of the Results 110 Under field conditions, the work carried out on sugar beet during a complete growing season had the aim of determine the responsiveness of the crop to induced water stress under different soil and irrigation conditions, dividing the plots according to the texture (sandy or clay texture). The differences between the canopy and the air temperature (∆T =Tc- Ta) were measured in both controlled deficit irrigation (CDI) and full irrigated (FI) treatments. The results indicate that the difference in ∆T between DI and FI plants increased as the relative ET values decreased, although a large dispersion was found in the relation between both, possibly due to the influence of other factors such as air vapor pressure deficit, as explained in Maes et al. (2012). After the validation against a conventional infrared thermometer, the developed system was compared to the canopy temperature measurements performed by a commercial thermal camera several orders of magnitude more expensive. Seven flights at different flight heights were made during the period in which the crop completely covered the soil, mounting the two sensors together on board the UAV. The similarity in the Tc measurements in both treatment zones using the thermal camera and the low-cost sensors (view Table 3.4) indicates that the developed system can be a promising alternative to expensive thermal cameras. The percentage of good data collected by the sensor averaged the 95%, similar to that obtained by Fisher et al. (2010), with the added challenge of having been gathered aboard an aerial platform. Despite the system demonstrated its robustness in terms of operation and acceptable accuracy, this type of sensor has some drawbacks: In the study of the operational parameters with respect to the measurements, it was revealed that the absence of wind and clouds was a relevant factor in the measurement. This limit the areas in which the Tc can be a reliable indicator, since as stated Jones et al., (1997), infrared thermometry has had several application difficulties in humid regions due to the lower vapour pressure deficit than in semi-arid areas. Another drawback is the wide FOV (around 90º), doesn’t allow knowing which physical element is being targeted by the sensor. This limits its use to measurements that seek to know the temperature of a surface as a whole and not a particular point, especially when mounted on a platform for in-height measurements. In addition, this fact also limits its use in crops that completely cover the soil, to be sure that the measurements correspond to the Tc and not to the soil temperature. However, this work has demonstrated that the implementation of this type of affordable remote sensors allows the estimation of the water status in large areas using infrared General Discussion of the Results 111 thermometry and taking advantage of aerial platforms. Therefore, this technology opens up a range of possibilities while at the same time reduces costs, something that will certainly allow its mass adoption even in developing areas. A relevant aspect to accurately know the crop water status (that was already observed in the first chapter of this thesis) is that unlike other physiological parameters such as leaf water potential (ψL) or stomatal conductance (gs), the canopy temperature is strongly affected by the climatic conditions at the time of measurement. In order to minimize the effect of the environmental variation and to be able to adequately estimate the water status of the crop, it is necessary to develop indexes that have reference temperature values. In the Chapter 2 of this thesis the Crop Water Stress Index (CWSI) has been developed according to the empirical model proposed by Idso et al. (1981), as a reliable estimator of the water status variability in super high-density olive orchards during irrigation season. For this, the non-water-stressed-baseline (NWSB) and its daily and seasonal evolution have been determined. In addition, the CWSI has been calculated based on high resolution aerial thermal images and the suitability of the CWSI to estimate the variation of water status in super high-density olive orchards has been determined. Three treatments were defined in this olive grove: a full irrigation (FI) treatment that covered the complete demand for water, and two treatments of regulated deficit irrigation (CDI) covering 45% of the demand (one determined by the method of the crop coefficient 45RDICC and another one based on the leaf turgor 45RDILT). In order to determine the NWSB, the temperature of the canopy was measured continuously for one year using infrared thermometers, permanently placed at a height of one meter above the canopy. From the results obtained regarding the NWSB calculation, a significant correlation between ∆T and VPD was found, with a clear effect of diurnal variation, in which this correlation was greater during the central hours of the day. A seasonal variation was also found, being the period from June - July (in which the angle of solar incidence is higher) in which the relationship between ΔT and VPD was more significant. These variations were previously reported by other authors (Bellvert et, 2014; Testi et al., 2008), and explained by the tight relationship between the solar radiation and NWSB-intercept. The canopy measurements obtained with the infrared thermometers and the thermal camera mounted on the UAV were used to derive seasonal CWSI in the three irrigation treatments. This CWSI was highly determined by the treatment: while the FI treatment maintained a CWSI near to 0 during all the irrigation season, values for the 45RDICC and 45RDILT ranged from 0.8 to 0.2 during the stress seasons. A seasonal trend was also General Discussion of the Results 112 observed in the data obtained, besides that CWSI showed a linear correlation between stem water potential (ψst), leaf water potential (ψL) and stomatal conductance (gs). This work also allowed to corroborate the suitability of aerial thermography to map water stress in discontinuous canopy structures such as the super high-density olive grove, as previously reported in Sepulcre-Cantó et al. (2006). Because the flights were made using a thermal camera with a resolution that can be considered low (324x256 pixels), low altitude flights (20 meters) had to be performed, which, considering the limited autonomy of these equipment, does not allow to cover large areas. As reported in Bellvert et al. (2014), camera resolution plays a determining role in the acquisition of thermal images, and it was recommended not to perform flights whose height generated pixels greater than 30 cm of ground sampling distance (GSD). In our case, the flights were conducted at a height in which the GSD was around 13 cm. In addition, it was necessary to develop an algorithm of image analysis, which by segmentation techniques was able to extract the pixels that actually corresponded to the vegetation and separate them from those corresponding to the ground or shadows, which have different temperatures. Chapter 3 of this thesis is devoted to the study of automated weed control, focusing on the plant/weed detection and intra-row plant spacing measurement. This line of work is considered to be of great interest for producers (and also for the author), since sustainable weed management presents a major challenge in the coming years, even more if it is approached from the perspective of automation (Fennimore et al., 2016). The objective defined in this chapter was to design a system capable of detecting the presence of tomato plants of the same crop line, and to determine the distance between them. For this purpose, a system based on three types of optical sensors was developed and tested: i) three pairs of light-beam sensors, in curtain arrangement, that allowed to detect the presence/absence of plants with no ability to discern its morphological structure, only distinguishing between weeds and crops based on its height; (ii) Kinect sensor (camera), a commercial RGB-D sensor capable of obtaining RGB, NIR images and depth images (based on Time-of-Flight); iii) LiDAR sensor, a commercial laser scanner that provide distance measurements along a line scan at a very fast scanning rate, generating 3D point clouds. The first two aforementioned sensors are considered low-cost sensors, since they did not exceed a hundred euros in both cases, while the LiDAR scanner used to perform the validation was much more expensive. The different General Discussion of the Results 113 sensors mounted on a platform in the front of the tractor were tested under field conditions in California and Seville, assisted by an encoder-based odometry system. The results obtained with light-beam sensors indicated a very high detection capacity (~ 98% accuracy in California test). The LiDAR sensor was able to detect 100% of sticks that simulated the stem of the tomato plant, whereas when tested on real plants, the developed algorithm obtained between 3-21% of false positives or negatives when detecting the plant. The insertion point of the stem in the ground was demonstrated to be the most difficult to detect. On the other hand, the image analysis from those captured by the Kinect sensor (combined with the odometry system), provided very good results, detecting the 100% of the plants and being able to accurately measure the distance between stems. These detection results with LiDAR and light-beam sensors are in agreement with the obtained by Garrido-Izard et al., (2014), whereas the precise detection using the Kinect sensor demonstrates the validity of what Fennimore et al. (2016) stated before: computer vision is one of the most promising fields in the challenge of detecting and discriminating weeds mounted on modern implements and farm machinery. The results obtained from the precise locations of the plants demonstrate the viability the low-cost detection system, and contributes to the development of further methods capable of performing individualized treatments to remove existing weeds between crop plants in the same row. Different considerations can be established regarding the location of the same plant. According to the structure and morphology of the plant, it is assumed that the aerial part will not always be vertically in line with the stem. For this reason, three locations can be established as the location of the plant: the aerial part, the cluster centroid of the filtered point cloud or the insertion of the stem. Depending on the type of treatment to be made, one of these locations could be more interesting to evaluate than the others. This work showed the importance of the robustness of these systems, despite being low cost, since they are mounted on the tractor and are subject to vibrations and an unstructured environment. Future work The contributions of this thesis open new lines of future research that will be interesting to explore. On the one hand, as already mentioned in the introduction, the integration of affordable sensors is open to help irrigation management systems and complete automation of them. It will be interesting to achieve this integration in these systems for its later General Discussion of the Results 114 inclusion on the modern FMIS. As has been pointed out, one of the barriers that these FMIS still present, is the lack of adaptation to the local conditions of the crops. Affordable sensors can achieve greater adoption of technology, and generate a cycle of 1) greater data acquisition, 2) algorithm refinement, 3) more accurate decision-making and 4) greater yields in crops. Based on the response of these sensors and its ability to measure temperature in the field and its relationship with water stress indexes, it will be interesting to see these responses in other crops, as well as to evaluate the suitability of carrying out developments with other affordable and commercial sensors in order to compare the results. On the other hand, the improvement and price-lowering of detection and localization systems in crops with row-growing patterns will enable the research group to develop agricultural implements that can act accurately on weeds among crop plants. One of the objectives is to develop an intelligent implement capable of eliminating weeds by abrasion using solid waste from other crops. In addition, the implementation of these detection and measurement methods on robotic platforms opens a very interesting future research route. Author envisions these robotic platforms being commonplace on farms in the next decades, with fully automated operations such as weed control made by smart machinery and implements. It will also be interesting to continue the development of algorithms for the identification, classification and training of systems based on computer vision. The author finds in this branch a promising way of working in agriculture regarding its use for example for navigation, weed detection, pest identification or yield predictions. 115 Conclusions 116 VI. Conclusions As a summary, the conclusions and final considerations that can be established from the present dissertation are presented below: - Affordable sensors validated on this thesis have proven their reliability and precision when developing innovative solutions for specific field contexts within data driven agriculture. - The use of infrared thermometry using low cost sensors to estimate the water status of sugar beet crops can be performed satisfactorily. In addition, it has been shown that the canopy temperature is a reliable indicator of water status, while it is highly conditioned by climatic variables (air temperature, wind speed, presence of clouds, etc.), which limit its use to non-humid regions. - The variability in water status and sustainability in water management presents a considerable challenge when it comes to evaluating highly technified crops such as super high-density olive groves. Due to the influence of climatic variables in the determination of water status, it is important to use indexes that take into account these environmental conditions, such as CWSI. - The feasibility of obtaining CWSI maps through the acquisition of thermal imagery using unmanned aerial platforms has been demonstrated. The determination of the NWSBs has allowed to quantify the diurnal and seasonal variability due to the incidence angle of solar radiation. - Infrared sensors and aerial thermography allows estimating the water status in industrial and woody crops, studying the spatial variability as a function of the temperature of the crop canopy and effectively delimiting areas of homogeneous management for precision irrigation. - The feasibility of the use of optical sensors for the detection of crop plants and the measurement of the space between tomato stems has been demonstrated. Combining cameras, infrared sensors and LiDAR laser scanner, high detection rates and high localization accuracy have been achieved. - The use of optical sensors as a detection system can lead to a new era in which precise weed control can be carried out in an affordable manner and be robust in field conditions. - The most significant technology barriers these affordable sensor presents are: 1) the need of enabling efficient data gathering and transmission, 2) the need of testing and refining analysis to accurately reflect local conditions, and 3) reaching a robustness scale that supports the challenges of field conditions. References 123 Kaierle S, Marx C, Rath T, Hustedt M (2013) Find and irradiate—lasers used for weed control. Laser Technology Journal 10:44–47 Khanal, S., Fulton, J., Shearer, S., (2017) An overview of current and potential applications of thermal remote sensing in precision agriculture. Computers and Electronics in Agriculture 139, 22–32. Kitchen, N.R., (2008) Emerging technologies for real-time and integrated agriculture decisions. Computers and Electrics in Agriculture 61, 1–3. Kitchen, N. R., Sudduth, K. A., Drummond, S. T., Scharf, P. C., Palm, H. l., Roberts, D. F., and Vories, E. D. (2010) Ground-based Canopy reflectance sensing for Variable Rate Corn fertilization. Agronomy Journal, 102, 71–82 Kraehmer, H., Laber, B., Rosinger, C., Schulz, A., (2014) Herbicides as Weed Control Agents: State of the Art: I. Weed Control Research and Safener Technology: The Path to Modern Agriculture. Plant Physiology 166, 1119–1131 Krishna, K. R. (2016). Push Button Agriculture Robotics, Drones, Satellite-Guided Soil and Crop Management. CRC Press. Lambert, D. M.; Paudel, K. P.; Larson, J. A. (2015). Bundled Adoption of Precision Agriculture Technologies by Cotton Producers. Journal of Agricultural and Resource Economics, 40(2), 325- 345. Lati, R.N., Siemens, M.C., Rachuy, J.S., Fennimore, S.A., (2016) Intrarow Weed Removal in Broccoli and Transplanted Lettuce with an Intelligent Cultivator. Weed Technology 30, 655–663. Li, Z.L., Tang, R., Wan, Z., Bi, Y., Zhou, C., Tang, B., Yan, G., Zhang, X., (2009) A review of current methodologies for regional Evapotranspiration estimation from remotely sensed data. Sensors 9, 3801–3853. Llop, J., Gil, E., Llorens, J., Miranda-Fuentes, A., & Gallart, M. (2016). Testing the Suitability of a Terrestrial 2D LiDAR Scanner for Canopy Characterization of Greenhouse Tomato Crops. Sensors, 16(9), 1435. Llorens, J., Gil, E., Llop, J., Escolà, A., (2010). Variable rate dosing in precision viticulture: Use of electronic devices to improve application efficiency. Crop Protection 29, 239–248. López-Granados, F., Torres-Sánchez, J., De Castro, A.I., Serrano-Pérez, A., Mesas-Carrascosa, F.J., Peña, J.M., (2016). Object-based early monitoring of a grass weed in a grass crop using high resolution UAV imagery. Agronomy for Sustainable Development 36 Maes, W. H., Steppe, K. (2012). Estimating evapotranspiration and drought stress with groundbased thermal remote sensing in agriculture: a review. Journal of Experimental Botany, 63(13), 4671-4712. Maes, W., Huete, A., Steppe, K., (2017) Optimizing the Processing of UAV-Based Thermal Imagery. Remote Sensing 9, 476. Maes, W.H., Achten, W.M.J., Reubens, B., Muys, B., (2011) Monitoring stomatal conductance of Jatropha curcas seedlings under different levels of water shortage with infrared thermography. Agricultural and Forest Meteorology 151, 554–564. Martínez-Casasnovas, J.A.; Bigorda, D.V.; Ramos, M.C. (2009) Irrigation management zones for precision viticulture according to intra-field variability. In Proceedings of the EFITA Conference,Wageningen, The 6–8 July 2009 References 124 Martinez-Guanter, J., Garrido-Izard, M., Agüera, J., Valero, C., Pérez-Ruiz, M., 2017. Over-the- row harvester damage evaluation in super-high-density olive orchard by on-board sensing techniques. Advances in Animal Bioscience 8, 487–491. Mesas-Carrascosa, F.J., Verdú Santano, D., Meroño, J.E., Sánchez de la Orden, M., García- Ferrer, A., 2015. Open source hardware to monitor environmental parameters in precision agriculture. Biosystems Engineering 137, 73–83. Miranda-Fuentes, A., Llorens, J., Rodriguez-Lizana, A., Cuenca, A., Gil, E., Blanco-Roldán, G. L.; Gil-Ribes, J. A. (2016). Assessing the optimal liquid volume to be sprayed on isolated olive trees according to their canopy volumes. Science of The Total Environment, 568, 296-305. Möller, M., Alchanatis, V., Cohen, Y., Meron, M., Tsipris, J., Naor, A., ... & Cohen, S. (2007). Use of thermal and visible imagery for estimating crop water status of irrigated grapevine. Journal of experimental botany, 58(4), 827-838. Monteith, J. L. (1981). Evaporation and surface temperature. Quarterly Journal of the Royal Meteorological Society, 107(451), 1-27. Moriana, A., Villalobos, F. J., & Fereres, E. (2002). Stomatal and photosynthetic responses of olive (Olea europaea L.) leaves to water deficits. Plant, Cell & Environment, 25(3), 395-405. Mulla, D.J., (2017) Encyclopedia of GIS. Chapter 1 pp 1-8 DOI:10.1007/978-3-319-23519- 6_1652-1. Nevalainen, O., Honkavaara, E., Tuominen, S., Viljanen, N., Hakala, T., Yu, X., Hyyppä, J., Saari, H., Pölönen, I., Imai, N., Tommaselli, A., (2017). Individual Tree Detection and Classification with UAV-Based Photogrammetric Point Clouds and Hyperspectral Imaging. Remote Sensing 9, 185. Nielsen, D. C. (1990). Scheduling irrigations for soybeans with the crop water stress index (CWSI). Field Crops Research, 23(2), 103-116. Niu, Q.; Fratta, D.; Wang, Y.-H. (2015). The use of electrical conductivity measurements in the prediction of hydraulic conductivity of unsaturated soils. Journal of Hydrology 522, 475–487. Norsworthy, J.K., Ward, S.M., Shaw, D.R., Llewellyn, R.S., Nichols, R.L., Webster, T.M., Bradley, K.W., Frisvold, G., Powles, S.B., Burgos, N.R., Witt, W.W., Barrett, M., (2012) Reducing the Risks of Herbicide Resistance: Best Management Practices and Recommendations. Weed Science 60, 31–62 O’Shaughnessy, S. A., Evett, S. R., & Colaizzi, P. D. (2014). Infrared thermometry as a tool for site-specific irrigation scheduling. In Proceedings of the Central Plains Irrigation Conference. O’Shaughnessy, S.A., Hebel, M.A., Evett, S.R., Colaizzi, P.D., (2011) Evaluation of a wireless infrared thermometer with a narrow field of view. Computers and Electronics in Agriculture 76, 59–68. Oerke E-C., Steiner, U. (2010) Potential of digital thermography for disease control. In: Oerke E- C, Gerhards R, Menz G, Sikora RA (eds) Precision crop protection – the challenge and use of heterogeneity, 1, vol Chapter 11. Springer Science, pp 167–182 Oerke, E-C., Steiner, U. H-W Dehne, M Lindenthal; (2006) Thermal imaging of cucumber leaves affected by downy mildew and environmental conditions. Journal Experimental Botany; 57 (9): 2121-2132 P.J. Zarco-Tejada, V. González-Dugo, J.A.J. Berni, (2012) Fluorescence, temperature and narrow-band indices acquired from a UAV platform for water stress detection using a microhyperspectral imager and a thermal camera. Remote Sensing of Environment. Volume 117, Pages 322-337. References 125 Pajares, G., (2015) Overview and Current Status of Remote Sensing Applications Based on Unmanned Aerial Vehicles (UAVs). Photogrammetric Engineering & Remote Sensing. 81, 281– 330. Paulus, S., Behmann, J., Mahlein, A.K., Plümer, L., Kuhlmann, H., (2014) Low-cost 3D systems: Suitable tools for plant phenotyping. Sensors 14, 3001–3018 Peña, J.M., Torres-Sánchez, J., de Castro, A.I., Kelly, M., López-Granados, F.; (2013) Weed Mapping in Early-Season Maize Fields Using Object-Based Analysis of Unmanned Aerial Vehicle (UAV) Images. PLoS One 8, 1–11. Pereira, Luis S., Ian Cordery, and Iacovos Iacovides. (2012) Improved Indicators of Water Use Performance and Productivity for Sustainable Water Conservation and Saving. Agricultural Water Management 108: 39–51. Perez-Ruiz, M., Gonzalez-de-Santos, P., Ribeiro, A., Fernandez-Quintanilla, C., Peruzzi, A., Vieri, M., Tomic, S., Agüera, J.; (2015) Highlights and preliminary results for autonomous crop protection. Computers and Electronics in Agriculture 110, 150–161. Perez-Ruiz, M., Slaughter, D.C., Fathallah, F.A., Gliever, C.J., Miller, B.J., (2014) Co-robotic intrarow weed control system. Biosystems Engineering 126, 45–55. Perez-Ruiz, M., Slaughter, D.C., Gliever, C., Upadhyaya, S.K., (2012). Tractor-based Real-time Kinematic-Global Positioning System (RTK-GPS) guidance system for geospatial mapping of row crop transplant. Biosystems Engineering 111, 64–71. Pérez-Ruiz, M., Slaughter, D.C., Gliever, C.J., Upadhyaya, S.K., (2012) Automatic GPS-based intra-row weed knife control system for transplanted row crops. Computers and Electronics in Agriculture 80, 41–49. Peteinatos, G.G., Weis, M., Andújar, D., Rueda Ayala, V., Gerhards, R., (2014) Potential use of ground-based sensor technologies for weed detection. Pest Management Science 70, 190–199. Piron, A.; Leemans, V.; Lebeau, F.; Destain, M.F. (2009) Improving in-row weed detection in multispectral stereoscopic images, Computers and Electronics in Agriculture, Volume 69, Issue 1, Pages 73-79, Poulton, C.; Dorward, A.; Kydd, J.; (2010). The Future of Small Farms: New Directions for Services, Institutions, and Intermediation. World Development. 38, 1413–1428. Priyadharsini, S., Sathishkumar, B.S., (2015) Developing a Real Time Smart Herbicide Sprayer Robot and Automatic Weed Detection System. International Journal of Emerging Technology and Advanced Engineering 5, 323–326. Quebrajo, L.; Pérez-Ruiz, M.; Rodriguez-Lizana A.; Agüera, J.; (2015) An approach to precise nitrogen management using hand-held crop sensor measurements and winter wheat yield mapping in a Mediterranean environment. Sensors 15, 5504–5517. Rasmussen, J., Griepentrog, H.W., Nielsen, J., Henriksen, C.B., (2012). Automated intelligent rotor tine cultivation and punch planting to improve the selectivity of mechanical intra-row weed control. Weed Research. 52, 327–337 Romeo, J., Guerrero, J.M., Montalvo, M., Emmi, L., Guijarro, M., Gonzalez-de-Santos, P., Pajares, G., (2013) Camera sensor arrangement for crop/weed detection accuracy in agronomic images. Sensors 13, 4348–4366. Rud, R., Cohen, Y., Alchanatis, V., Levi, A., Brikman, R., Shenderey, C., Heuer, B., Markovitch, T., Dar, Z., Rosen, C., Mulla, D., Nigon, T., (2014) Crop water stress index derived from multi- References 126 year ground and aerial thermal images as an indicator of potato water status. Precision Agriculture 15, 273–289. Saber, M.N.; Lee, W.S.; Burks, T.F.; MacDonald, G.E.; Salvador, G.A.; (2013) An Automated Mechanical Weed Control System for Organic Row Crop Production. Transactions of 2013 ASABE Annual International Meeting Scholander, P.F., Hammel, H.T., Bradstreet, E.D., Hemmingsen, E.A., (1965). SapPressure in Vascular Plants: negative hydrostatic pressure can be measured inplants. Science 148, 339–346. Sepaskhah, A.R., Kashefipour, S.M., (1994) Relationships between leaf water potential, CWSI, yield and fruit quality of sweet lime under drip irrigation. Agricultural Water Management 25, 13– 21. Sepulcre-Cantó, G., Zarco-Tejada, P.J., Jiménez-Muñoz, J.C., Sobrino, J.A., Miguel, E. De, Villalobos, F.J., (2006) Detection of water stress in an olive orchard with thermal remote sensing imagery. Agricultural and Forest Meteorology 136, 31–44 Shah, N.G.; Das, I. (2012) Precision Irrigation: Sensor Network Based Irrigation. In Problems, Perspectives and Challenges of Agricultural Water Management; InTech pp. 217–232. Shah, R.; Lee, W.S. (2015) An approach to a laser weeding system for elimination of in-row weeds. Proceedings of the X European Conference on Precision Agriculture, Tel Aviv, Israel, 12– 16 July 2015; pp. 309–312. Siegfried, J., Khosla, R., & Longchamps, L. (2017). Infrared thermometry to quantify in-field soil moisture variability. Journal of Crop Improvement, 31(1), 72–90. Slaughter, D. (2014). The Biological Engineer: Sensing the Difference Between Crops and Weeds. In Stephen L.Young & Francis J.Pierce (Eds.) Automation: The Future of Weed Control in Cropping Systems (pp 71-95) Slaughter, D.C., Giles, D.K., Downey, D., (2008) Autonomous robotic weed control systems: A review. Computers and Electronics in Agriculture. 61, 63–78. Smith, B.C.M., Dhuyvetter, K.C., Kastens, T.L., Dietrich, L., Smith, L.M., (2013) Economics of Precision Agricultural Technologies Across the Great Plains. Journal of the American Society of Farm Managers and Rural Appraisers. 76, 185–206. Sun, H., Slaughter, D.C., Ruiz, M.P., Gliever, C., Upadhyaya, S.K., Smith, R.F., (2010) RTK GPS mapping of transplanted row crops. Computers and Electronics in Agriculture. 71, 32–37. Sun, H., Slaughter, D.C., Ruiz, M.P., Gliever, C., Upadhyaya, S.K., Smith, R.F., (2010) RTK GPS mapping of transplanted row crops. Computers and Electronics in Agriculture 71, 32–37 Tanner, C. B. (1963). Plant temperatures. Agronomy Journal, 55(2), 210-211 Tattaris, M., Reynolds, M.P., Chapman, S.C., (2016) A Direct Comparison of Remote Sensing Approaches for High-Throughput Phenotyping in Plant Breeding. Frontiers in Plant Science 7, 1– 9 Testi L., Goldhamer D., Iniesta F., Salinas M. (2008) Crop water stress index is a sensitive water stress indicator in pistachio trees. Irrigation Science 26:395–405. Tey, Y. S.; Brindal, M. (2012). Factors influencing the adoption of precision agricultural technologies: a review for policy implications. Precision Agriculture, 13(6), 713-730. References 127 Torres-Sánchez, J., López-Granados, F., De Castro, A.I., Peña-Barragán, J.M., 2013. Configuration and Specifications of an Unmanned Aerial Vehicle (UAV) for Early Site-Specific Weed Management. PLoS One 8. Turner, N. C., Schulze, E. D., & Gollan, T. (1984). The responses of stomata and leaf gas exchange to vapour pressure deficits and soil water content. Oecologia, 63(3), 338-342. Underwood, J.P.; Jagbrant, G.; Nieto, J.I.; Sukkarieh, S.; (2015) Lidar-Based Tree Recognition and Platform Localization in Orchards. Journal of Field Robotics 32(8), 1056–1074. Urbahs, A., Jonaite, I., (2013) Features of the use of unmanned aerial vehicles for agriculture applications. Aviation 17, 170–175. USAID (2016) Low-Cost Sensors for Agriculture. Key Findings Report. June 2-3, 2016 Available Online:https://static.globalinnovationexchange.org/s3fs- ublic/asset/document/USAID%20Sensors4Ag%20Key%20Findings%20FINAL_FOR%20DISTRI BUTION.pdf Accesed on 13th June 2017. VAROS (Variable Rate Operations in Orchards) project. (2016) ICT-Agri and robotics for sustainable agriculture Program http://ict-agri.eu/node/36325 Accessed on 10 June 2017 Wang, D.; Gartung, J. (2010) Infrared canopy temperature of early-ripening peach trees under postharvest deficit irrigation. Agricultural Water Management, 97, 1787–1794. Wang, L., Qiu, G. Y., Zhang, X., & Chen, S. (2005). Application of a new method to evaluate crop water stress index. Irrigation Science, 24(1), 49-54. Wang, X., Yang, W., Wheaton, A., Cooley, N., & Moran, B. (2010). Automated canopy temperature estimation via infrared thermography: A first step towards automated plant water stress monitoring. Computers and Electronics in Agriculture, 73(1), 74-83. Wendel, A., Underwood, J., 2016. Self-supervised weed detection in vegetable crops using ground based hyperspectral imaging. 2016 IEEE International Conference on Robotics and Automation (ICRA) 5128–5135. West, J.S., Canning, G.G.M., Perryman, S.A., King, K., 2017. Novel Technologies for the detection of Fusarium head blight disease and airborne inoculum. Tropical Plant Pathology doi:10.1007/s40858-017-0138-4 Whipker, L.D., and J.T. Akridge. (2009) Precision Agricultural Services: Dealership Survey Results. West Lafayette, IN: Center for Food and Agricultural Business, Purdue University. Wired Magazine "Re-Planting a Forest, One Drone at a Time.". (2015) Accessed via Web on June 2017. https://www.wired.com/brandlab/2015/07/re-planting-forest-one-drone-time/ Wortman, S. (2014). Integrating weed and vegetable crop management with multifunctional airpropelled abrasive grits. Weed Technology 28:243–252 Yang, H., Yang, X., Heskel, M., Sun, S., Tang, J., (2017) Seasonal variations of leaf and canopy properties tracked by ground-based NDVI imagery in a temperate forest. Scientific Reports 7, 1267. Yao, Y., Liu, Q., Liu, Q., (2009) Canopy modeling and validation for row planted crops of key growth stages. International Geoscience Remote Sensing Symposium. 2, 686–689. Young, S.L. and Giles, D.K. (2014) Targeted and Microdose Chemical Applications. In Stephen L.Young & Francis J.Pierce (Eds.) Automation: The Future of Weed Control in Cropping Systems (pp 139-148) References 128 Yu, X., Wu, P., Han, W., & Zhang, Z. (2013). A survey on wireless sensor network infrastructure for agriculture. Computer Standards & Interfaces, 35(1), 59-64. Zaman-Allah, M., Vergara, O., Araus, J.L., Tarekegne, A., Magorokosho, C., Zarco-Tejada, P.J., Hornero, A., Albà, A.H., Das, B., Craufurd, P., Olsen, M., Prasanna, B.M., Cairns, J., (2015) Unmanned aerial platform-based multi-spectral imaging for field phenotyping of maize. Plant Methods 11, 35. Zhang P, Deng L, Lyu Q, He S L, Yi S L, Liu Y D. (2016) Effects of citrus tree-shape and spraying height of small unmanned aerial vehicle on droplet distribution. International Journal of Agricultural and Biological Engineering; 9(4): 45－52 Zhang, C., Kovacs, J.M., (2012). The application of small unmanned aerial systems for precision agriculture: A review. Precision Agriculture 13, 693–712. Zhang, H., Wang, D., Gartung, J., (2017) Influence of Irrigation Scheduling Using Thermometry on Peach Tree Water Status and Yield under Different Irrigation Systems. Agronomy 7, 12. Zhang, Y.; Slaughter, D.C.; (2011) Hyperspectral species mapping for automatic weed control in tomato under thermal environmental stress, Computers and Electronics in Agriculture, Volume 77, Issue 1, Pages 95-104 Zhang, Y.; Staab, E.S.; Slaughter,D.C.; Giles, D.K.; Downey,D. (2012) Automated weed control in organic row crops using hyperspectral species identification and thermal micro-dosing, Crop Protection, Volume 41, Pages 96-105, ISSN 0261-2194 Zhenyang, G.; Wangwang, W., Yingjie, Y., Ruiqing, Z. (2013). Design of Mechanical Arm for Laser Weeding Robot. Proceedings of the 2nd International Conference on Computer Science and Electronic Engineering (ICCSEE 2013) 2340–2343. Zude-Sasse, M.; Fountas, S.; Gemtos, T. A.; Abu-Khalaf, N.; (2016). Applications of precision agriculture in horticultural crops. European Journal of Horticultural Science. 81(2), 78–90. 129 Derechos de autor ©2017 por Jorge Martínez Guanter. Todos los derechos reservados. La Universidad de Sevilla, España, podrá incluir en su catálogo digital esta Tesis, solo para usos no comerciales. Se permite el uso personal de este material. No se permite reimprimir/republicar este material con fines lucrativos, publicitarios o promocionales o para la creación de nuevos trabajos colectivos para la reventa o redistribución a servidores o listas, la reutilización de cualquiera de los componentes con derechos de autor de esta obra en otros trabajos se debe obtener de su autor. TODOS LOS DERECHOS PERTENECEN AL AUTOR.