Task-based agricultural mobile robots in arable farming: A review

REVIEW ARTICLE OPEN ACCESS Task-based agricultural mobile robots in arable farming: A review Krishnaswamy R. Aravind1, Purushothaman Raja1 and Manuel Pérez-Ruiz2 1SASTRA University, School of Mechanical Engineering, Thanjavur-613401, Tamil Nadu, India. 2University of Seville, Area of Agroforestry Engineering, Dept. Aerospace Engineering and Fluid Mechanics, Ctra. Sevilla-Utrera km1, Sevilla 41013, Spain. Spanish Journal of Agricultural Research Instituto Nacional de Investigación y Tecnología Agraria y Alimentaria (INIA) Abstract In agriculture (in the context of this paper, the terms “agriculture” and “farming” refer to only the farming of crops and exclude the farming of animals), smart farming and automated agricultural technology have emerged as promising methodologies for increasing the crop productivity without sacrificing produce quality. The emergence of various robotics technologies has facilitated the application of these techniques in agricultural processes. However, incorporating this technology in farms has proven to be challenging because of the large variations in shape, size, rate and type of growth, type of produce, and environmental requirements for different types of crops. Agricultural processes are chains of systematic, repetitive, and time-dependent tasks. However, some agricultural processes differ based on the type of farming, namely permanent crop farming and arable farming. Permanent crop farming includes permanent crops or woody plants such as orchards and vineyards whereas arable farming includes temporary crops such as wheat and rice. Major operations in open arable farming include tilling, soil analysis, seeding, transplanting, crop scouting, pest control, weed removal and harvesting where robots can assist in performing all of these tasks. Each specific operation requires axillary devices and sensors with specific functions. This article reviews the latest advances in the application of mobile robots in these agricultural operations for open arable farming and provide an overview of the systems and techniques that are used. This article also discusses various challenges for future improvements in using reliable mobile robots for arable farming. Additional key words: precision agriculture; task-based agricultural robots; soil analysis; seeding; weed detection; harvesting; crop scouting robot Correspondence should be addressed to P. Raja: [email protected] Abbreviations used: AMS (Autonomous Mechanization System); EC (Electrical Conductivity); GEONET (Global positioning system Earth Observation NETwork); GNSS (Global Navigation Satellite System); GPS (Global Positioning System); IMU (Inertial Measurement Unit); LAI (Leaf Area Index); LAIC (Leaf Area Index Calculation); LIDAR (LIght Detection And Ranging); LPG (Liquefied Petroleum Gas); MC (Moisture Content); MRE (Mean Relative Error); NIR (Near Infra-Red); NIRS (Near Infra-Red Spectroscopy); PTO (Power Take-Off); RGB (Red Green Blue); RMSE (Root Mean Square Error); RTK (Real Time Kinematic); RTSS (Real Time Soil Sensing); SOM (Soil Organic Matter); TOF (Time Of Flight); UAV (Unmanned Aerial Vehicle); VOC (Volatile Organic Compound). Funding: The authors received no specific funding for this work Authors’ contributions: Conceived and identified the outline: KRA and PR. Drafting of the manuscript: KRA. Critical revision of the manuscript: PR and MPR. Competing interests: The authors have declared that no competing interests exist. Copyright © 2017 INIA. This is an open access article distributed under the terms of the Creative Commons Attribution (CC-by) Spain 3.0 License. Citation: Aravind, K. R.; Raja, P.; Pérez-Ruiz, M. (2017). Task-based agricultural mobile robots in arable farming: A review. Spanish Journal of Agricultural Research, Volume 15, Issue 1, e02R01. https://doi.org/10.5424/sjar/2017151-9573 Received: 3 March 2016. Accepted: 3 March 2017. 15(1), e02R01, 16 pages (2017) eISSN: 2171-9292 https://doi.org/10.5424/sjar/2017151-9573 Introduction Historically, agriculture has consumed a large amount of energy, and it continues to do so currently. In the past, energy has been extremely inexpensive, and agricultural products have consumed large amount of energy to develop rapidly. The world’s population has surpassed 7 billion and is expected to continue to grow in the coming decades, reaching 9 billion in 2050 (Ocampo, 2014). Therefore, agricultural production must continue to increase while consuming the minimum possible amount of resources. Moreover, conventional, imprecise mechanized farming demands relatively more petrochemical energy, the majority of which is already consumed by automobiles and other applications. To address the increasing energy demand, part of the harvest from agricultural crops such as corn and soya beans has been used to produce biofuels; this portion has doubled since 2007 according to the survey “Biofuels Impact on Crop and Food Prices”, which was conducted in 2009 (Baier et al., 2009). In light of these problems, precision Krishnaswamy R.Aravind, Purushothaman Raja and Manuel Pérez-Ruiz Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 2 in the agricultural production processes, from soil tilling to harvesting, is required for the efficient and quality production of crops with the minimum use of resources. Precision agriculture is a farm management practice that uses modern technologies to observe and respond to farm variability depending on computational analysis of observations (Mandal & Maity, 2013). Technology such as automation can be used in these applications; however, the agricultural operational environment is dynamic, and complex infrastructure and facilities with costly machines are required to completely automate an agricultural process. As an alternative to spending considerable money on infrastructure, a few intelligent mobile robots that each possess specific task capabilities and that are able to move and adapt in the current field can be developed to reduce production costs. A study by Pedersen et al. (2008) on the economic feasibility of using robot applications in agriculture indicated a significant reduction in production costs. Introducing robots into agriculture improves sustainability and consistency in agricultural tasks in addition to reducing costs. The precision of robots in tasks such as applying chemicals reduces environmental problems and their harmful effects on humans (Comba et al., 2010). Agricultural robots can include modified tractors, small ground robots (Blackmore et al., 2004a), and aerial robots. Several reviews have examined agricultural robots, which have been categorized based on several features. Table 1 lists articles in agricultural robots, mainly focusing on the latest trends in the development and use of robots with various features and their descriptions. Commercially available tractors can be modified into autonomous vehicles by adding the electronics and communication devices necessary for autonomous operation in agricultural fields. Blackmore et al. (2004b) implemented an automatic steering system with an initially defined route plan to a modified tractor. Small robots include small modified tractors or completely new robots. Small ground-based robots can assist a human in, for example, harvesting strawberries. These robots can transport strawberries from a worker to an unloading station; workers can spend up to 20% of their time simply walking back and forth from the strawberry field to the unloading station, making automation of this task very useful (Arikapudi et al., 2014). A group of small robots can work together by communicating with each other and the main coordination stations that Table 1. List of articles on recent trends in agricultural robots Reference Description Research in autonomous agricultural vehicles in Japan (Torii, 2000) Briefly explains the various applications of image sensors, soft computing, fuzzy control, and cooperative behavior in agricultural robots. Various researches by manufacturers such as Kubota and Mitsubishi are briefly explained. Agricultural robotssystem analysis and economic feasibility (Pedersen et al., 2006) A brief review of various studies on crop scouting, weeding and grass cutting robots. Analyses the economic benefits of using robots over conventional grass cutting machines. Robotics in crop production (Grift, 2007) Discusses robots in scouting operations, and other robotic operations and the advantages of using multi-robot systems. Overview of research on agricultural robots in China (Libin et al., 2008) Discusses various outdoor and indoor robots in various agricultural operations such as grafting, transplanting, spraying, mowing and harvesting in China. Robotics and automation for crop management: trends and perspective (Comba et al., 2010) A review on robots based on field type, navigation, agricultural operations, navigation and sensor systems along with their future possibilities. Autonomous robots for agricultural tasks and farm assignment and future trends in agro robots (Yaghoubi et al., 2013) Reviews the use of agricultural robots in fungicide and herbicide applications. Discusses various robots such as Ecobot I (sugar eating robot), Ecobot II (fly eating robot), and Ag Ant (legged robot to attack weeds). Mobile sensor platforms: categorisation and research application in precision farming (Zecha et al., 2013) Categorizes agricultural robots based on several factors such as systematic concept, type and method of sensing, size, mobility, propulsion, degree of automation, architecture and information fusion; suggests future directions for such robots. A brief overview and systematic approach for using agricultural robot in developing countries (Tarannum et al., 2015) Discusses various robot operations; land preparation, soil observation, seeding & planting, plant observation, harvesting and picking fruit. Task-based agricultural mobile robots in arable farming: A review Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 3 are monitored by humans (Blackmore et al., 2004a). Aerial robots and unmanned aerial vehicles (UAVs) have limited capacity to carry chemicals and batteries and cannot be used for certain ground-based tasks, such as soil preparation and plant-specific pesticide applications. However, aerial robots are excellent tools for data collection over large fields. Farming can be categorized based on the crop type. i.e., arable or permanent crops. Farming based on arable crops (as shown in Table 2) is known as arable farming. Arable farming is an important farming type that provides staple foods, medicines, and aromatic plants for the world population. This article mainly addresses the categorization of robots based on the agricultural tasks in open arable farming. This paper is organized as follows. Five major agricultural tasks in open arable farming are discussed in the “Task-based agricultural robots” section. Research regarding the use of mobile robots in each agricultural task is documented in the following sections: “Tilling robots”, “Soil analysis robots”, “Seeding and transplanting robots”, “Crop scouting and pest, weed, and disease control robots”, and “Harvesting robots”. Based on these studies, the scope for using agricultural robots and the challenges that need to be overcome to achieve reliable, autonomous agricultural mobile robots are discussed in the “Scope and challenges” section. The “Conclusion” section provides insights into future problems faced by agricultural robots. Task-based agricultural robots In this article, agricultural tasks are classified into five important operations for open arable farming based on the study by Blackmore et al. (2007). The five major operations are shown in Fig. 1; some of the suboperations shown as part of the major operations might be omitted depending on the crop and the farming type of the country. Tilling is the first step in the agricultural process; this step involves manipulating the soil by mixing the soil above and below the surface, loosening the soil particles, blending nutrients and making the bed suitable for crop growth. Tilling also benefits crop development by destroying weeds and insect pests (Kladivko, 2001). In some countries, irrigation is included as the step after tilling. Additional irrigation is required at specific intervals throughout the crop growth. Generally, farm irrigation is performed with widely used and efficient automation methods such as central pivot systems, automated sprinklers and drip irrigation. Otherwise, tilling is followed by soil analysis, which is the process of measuring various physical and chemical properties of the soil to evaluate its fertility and physical conditions. In some occasions, soil analysis is also performed during different stages of plant growth to determine the correct nutrients to apply. After tilling and soil analysis, crops can be grown either by placing seeds in the soil or by transplanting seedlings that were grown in a nursery into the field. Planting requires precision as each plant has specific space requirements and should therefore be placed at specific coordinates relative to the field. This precise positioning helps to generate a map containing the location of each plant, which aids the robots in performing subsequent operations. Crop scouting is the process of continuously monitoring the field to acquire information on the plant status, disease incidence, and weed and pest infestations, which affect crop growth. Based on the acquired information, a precise control methodology such as herbicide or pesticide applications can be used Table 2. List of arable crops Cereals Industrial crops Other crops Maize Cotton Vegetables Wheat Tobacco Sugarbeet Oats Hops Melons Sorghum Soya Pulses Rice Rape and turnip Potatoes Other cereals Sunflower Strawberries Other oil-seed or fiber crops Flowers and ornamental plants Aromatic plants Fodder crops Medicinal plants Culinary plants Figure 1. Task-based agricultural robots Krishnaswamy R.Aravind, Purushothaman Raja and Manuel Pérez-Ruiz Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 4 to improve the plant growth. Finally, the crops need to be harvested to obtain the produce; this process depends on the type of crop. Mobile robots that are categorized based on their ability to perform these five important agricultural processes are known as task-based agricultural robots, as shown in Fig. 1. Research on the individual categories of robots is presented in the following sections. Tilling robots Tilling operations consist of primary and secondary tilling. Primary tilling is deep agitation of the soil that produces a rough surface, while secondary tilling is the opposite of primary tilling. Plowing is an example of a primary tilling operation; harrowing, rototilling and cultivating are a few examples of secondary tilling. Primary tilling is followed by secondary tilling to produce a smooth surface. Tilling is a tedious and labor intensive process, making this task a clear target for automation (Sahay, 2006; Micheal & Ojha, 2008). Traditional tractors have sufficient power to pull up the soil because they are heavy and have high torque, but small robots cannot perform this task. Tractors also tend to damage soil through compaction. One of the earliest studies on automatic plowing was conducted by Harries & Ambler (1981) using a conventional tractor with a furrow following transducer and steering servo. This tractor used an opto-electronic sensing technique for turning around the corners; however, the system was not well suited for inclined slopes due to its large displacement errors. Trials conducted using the implemented guidance system and automatic turning control were monitored over 200 runs. The tractor returned to its target position within ± 30 cm over 50% of the time and within ± 60 cm over 75% of the time. In related research, an existing commercial tractor was modified into an autonomous robot, and additional equipment was attached to it in order to perform agricultural operations. Nagasaka et al. (2011) modified a commercial tractor (Yanmar EG65) by fixing a global navigation satellite system (GNSS) antenna for navigation and a DC motor for steering control. The GNSS of the tractor obtains the real-time reference position data from the global positioning systems (GPS) earth observation network system (GEONET) through the mobile phone network; the data then must be corrected for the inclination of the vehicle. The inertial measurement unit (IMU) in the tractor provides information regarding roll, pitch and heading angle. The data from the azimuth sensor, GNSS and IMU are compared to estimate the control parameter that is converted into an actuator command. This control parameter corrects the error in following a particular trajectory. Path planning was performed manually by considering the coordinates of the four corners of a square field and was provided to the tractor as an input. Matsuo et al. (2012) modified a commercially available tractor into a robotic vehicle known as Robotra, as shown in Fig. 2. Robotra was able to perform unmanned tilling operations using its pathplanning algorithm. This robot used the real-time kinematic-global navigation satellite system (RTKGNSS) to provide position information for navigation. RTK (Bakker et al., 2011) is a technique that is used to improve the accuracy of satellite-based positioning systems, such as GNSS, by approximately a few centimeters. Three methods of navigation operation were analyzed: basic, diagonal and round operation methods. Each operation method used a two-step process consisting of path planning and vehicle guidance to follow the path. The evaluation of these three operation methods concluded that the basic operation method can be conveniently performed by a human operator but that the other two methods can be efficiently and accurately performed only by robots since human visual judgment is limited for conducting precise operations in the field. Soil analysis robots Soil is the main source of nutrients for plants; therefore, various tests are manually performed in the field by taking samples across the field and then performing statistical analysis to estimate the soil properties. The results of laboratory tests depend on the number and density of the measurement locations. This process costs significant time and money to determine several soil properties. A study by Rossel & McBratney (1998) analyzed and compared the costs of estimating soil properties in the US and Australia. The average cost per sample for analyzing soil pH, carbon, nitratenitrogen, phosphorus and potassium were A$18.4, A$22.2, A$29.9, A$22.5 and A$19.4, respectively, in Australia. The costs in Australia were significantly Figure 2. Robotra performing a tilling operation Task-based agricultural mobile robots in arable farming: A review Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 5 higher than in the US, and precision agriculture requires more soil samples, resulting in economically inefficient farming. Therefore, an automated real-time measurement system for measuring soil properties can greatly benefit farmers. Scholz et al. (2014) developed an automatic soil penetrometer, which was integrated into an autonomous mobile robot named Bonirob, as shown in Fig. 3. The soil penetrometer has a probing rod with a force sensor that penetrates into the soil to a depth of 80 cm via a linear actuator. This robot is also equipped with surface moisture and temperature sensors and can measure the physical properties of the soil. Their study showed a strong correlation to the data in the commercial penetrologger with a root mean square error (RMSE) of 0.185, 0.145 and 0.120 MPa for soil textures of loamy sand, sand and silt, respectively. Pobkrut & Kercharoen (2014) developed a soilsensing survey robot based on an electronic nose to determine certain chemical properties of soil. The robot had six wheels and was equipped with six gas sensors: TGS 825 for hydrogen sulfide; MQ2 for combustible gas; MQ5 for LPG and natural gas; MQ135 for ammonia, benzene and carbon dioxide; TGS 2600 for air contaminants and TGS 2602 for volatile organic compound (VOCs) and odorous gases. The Arduino Mega 256 controller was used to obtain data from the sensors and to control the entire system. The robot sent the data to the system by using a Zigbee-based wireless network. The robot was tested under real field conditions with different soils, such as sandy soil, sandy soil with fertilizers, loamy soil and loamy soil with fertilizers. The responses of these sensors to various soil conditions were recorded, although further detailed study is needed to correlate the responses with the soil. A study by Baharom et al. (2015) on real-time soil sensing (RTSS) used a visible and near-infrared spectrophotometer to detect the various chemical properties of soil, such as the total carbon, organic matter, total nitrogen, available phosphorus, and moisture content in cultivated paddies. This RTSS process included a halogen lamp as a light source; these lights were guided by an optical fiber to illuminate a 50 mm-diameter area at depths of 10, 15 and 20 cm below the soil surface. The reflected spectra were then guided to the spectrophotometer by the optical fiber and analyzed. A calibration models was built, and the soil was mapped at all three depths. The highest accuracy of the combined data for the three depths had correlation coefficient (R2) values of 0.88, 0.83, 0.88, 0.85 and RMSE values of 1.38, 0.26, 0.15, 0.01% for moisture content, organic matter, total carbon and total nitrogen, respectively. The results from this study suggest that combining the data from all three depths provides better prediction accuracy. This RTSS configuration is connected to a commercial tractor and has not yet been tested while attached to an autonomous robot. A few notable studies on RTSS using commercial, non-autonomous tractors are listed in Table 3. Because these systems are automated, they can potentially interface with robots. Seeding and transplanting robots Seeding is the process of planting seeds in the soil so that they are successfully able to germinate. Transplanting involves placing a small plant seedling that has germinated in a particular position in the field based on the specific space requirements of each crop in the field. Food grains, such as rice and wheat, represent a major types of food that is consumed by people around the world. Haibo et al. (2010) developed a precise wheat seeder robot that uses an air suction precision seeding mechanism to accurately drop seeds using an RTK -GNSS module. The precision of the seeding mechanism was ensured by considering the geometric characteristics of wheat seeds. The study analyzed the influence of design and suction speed for precise seeding and identified the process and structures necessary to optimize this precision. Based on the analysis results, the optimum diameter of the suction hole in the metering plate was in the range of approximately 2.0 to 2.8 mm, and the optimum vacuum in air chamber was between approximately 1.8 to 2.8 kPa to ensures the precise placement of the seeds in the soil. Nagasaka et al. (2011) automated the rice transplanter Kubota SPU650 by fixing a GNSS antenna approximately 2 m from the ground surface and by controlling the steering using a servomotor that was connected to the steering axle. Path planning was Figure 3. Bonirob with a soil penetrometer Krishnaswamy R.Aravind, Purushothaman Raja and Manuel Pérez-Ruiz Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 6 Table 3. Related works on automated real-time soil sensing studies Reference Sensing equipment Detection parameters[1] Results[2] On-line real time soil sensor (Shibusawa, 2003) Spectrophotometer Prediction model for MC, SOM, NO3-N, pH, EC and soil maps Evaluated at 860 locations MC (R2 = 0.95) SOM (R2 = 0.93 NO3-N (R2 = 0.94) pH (R2 = 0.99) EC (R2 = 0.93) Direct measurement of soil chemical properties on the go using ion-selective electrodes (Adamchuk et al., 2005) Ion-selective electrodes Soil pH, available potassium, NO3-N, and sodium RMSE varied from 0.11 to 0.26 pX in the order pH< pK< pNO3 <pNa (precision) & 0.20 to 0.37 pX (accuracy) Development of soil pH and lime requirement maps using on the go soil sensors (Lund et al., 2005) Ion-selective pH electrodes, electrical conductivity sensor, near infra-red spectrometer pH, and lime requirements of the soil Overall RMSE=0.38 pH. RMSE ranged from 0.28 to 0.55 pH for different fields in USA. The predicted R2 value was 0.52 kg/ha on the pH map, and R2 value was 0.83 kg/ha on the pH, EC & NIRS combined map An automated system for rapid in-field soil nutrient testing (Lobsey et al., 2010) Ion-selective electrodes Proximal sensing of soil nitrate, potassium and sodium The predicted R2 values for nitrate, sodium and potassium were 0.92, 0.99, and 0.99, respectively. Soil pH mapping with an on the go sensor (Schirrmann et al., 2011) Ion-selective antimony elec trodes Soil pH The soil pH values from the pH sensor were well correlated with the lab pH (CaCl2) values. After calibration, the mean absolute error varied from 0.28 to 0.48 pH units. Integrated sensing of soil moisture at the field scale: measuring, modeling and sharing for improve agricultural decision support (Phillips et al., 2014) Hydraprobe II – coaxial impedance dielectric reflectometry Soil moisture Using the Penman-Monteith method, the water stress conditions were estimated. Data of the potential evaporation, evapotranspiration & precipitation presented the inputs and outputs throughout the plant growth stages until harvest. [1]MC, moisture content; SOM, soil organic matter; EC, electrical conductivity. [2]pX: negative base 10 logarithm of specific ion activity. NIRS: near infrared spectroscopy Table 4. Various works on agricultural robots for seeding and transplanting operation Reference Description Development of a seed-planting robot for the creation of large scale growing flower images (Riesen & Rohrer, 2011) 67% of seed ejected was within a1-cm radius, and 85% fell within 1.5 cm. Automated three-wheel rice seeding robot operating in dry paddy fields (Ruangurai et al., 2015) GNSS-based positioning was used along with extended Kalman filter-based localization; the average seeding accuracy was 91%. Command-based selfguided digging and seed sowing rover (Priyadarshini & Sheela, 2015) Commands were given manually by a mobile phone for navigation and sowing seeds. Initial field-testing of Thorvald, a versatile robotic platform for agricultural application (Grimstad et al., 2015) Used RTK-GNSS for navigation. Seeding experiments were conducted using different seeding patterns and were analyzed for weed suppression. Task-based agricultural mobile robots in arable farming: A review Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 7 performed manually by considering the four corners of a square field, and the path was corrected by measuring the deviation from the desired path. The same authors used another rice transplanter (Iseki PZ60), as shown in Fig. 4, with similar automation and path planning procedures (Nagasaka et al., 2011). Oksanen (2013) modified a tractor named the APU–Module, as shown in Fig. 5, and successfully conducted trials for the autonomous sowing of spring wheat. This vehicle had the capacity to store seeds for sowing 0.85 ha without any refills. Navigation was performed by using a GNSS-based guidance system. This study indicated many practical problems, such as drops in the GNSS signal because of natural obstacles, errors in satellite communication, and frequent manual intervention. Manual intervention was required to check the seed and the drill settings as the tractor was not able to perform automatic detection and adjustments preventing the complete autonomy of the robot. Recovering the GNSS signal quality required a period of 1 h 27 min and 11-20 interruptions occurred per hectare. The longest continuous operation without intervention was approximately 20 min and the shortest was 2.8 s. The spatial accuracy deviated by approximately 12.5 mm from the mean with respect to the selected path. The mean lateral error was -0.15 cm, and the mean angular error was 0.06°. Griepentrog et al. (2013) retrofitted a Hakotrac 3000 with GNSS for navigation and an electro-hydraulic valve for steering to create an autonomous mechanization system (AMS). Crops were established by interfacing with the data logging system that stored maps for seeding with a grid seeder and punch plater. GNSS was used for the precise placement of seeds in the field. The experimental results showed a mean standard deviation of 2.53 mm; and based on a normal distribution 95% of the data were within 5.1 mm. Table 4 provides insight regarding some of the other notable studies conducted in the area of seeding and transplanting using agricultural mobile robots. Crop scouting and pest, weed and disease control robots In general, crop scouting is the process of assessing an agricultural field through monitoring of factors such as pests, weed growth and diseases, which can restrict the crop growth. Crop status monitoring robots Crops exhibit various characteristic features because of genetic factors, aging, responses to environment, pests, and soil fertility. Each characteristic provides information regarding complex traits such as morphology, physiology, growth, ecology, and yield. These complex traits can be assessed by measuring basic quantitative parameters such as leaf length, the leaf area index (LAI), canopy volume, and shoot biomass; and this process is known as plant phenotyping (Li et al., 2014). These parameters can be estimated by using various sensors, including light detection and ranging (LIDAR) or image-capturing devices such as nearinfrared (NIR), visible, hyperspectral and multispectral cameras (Fahlgren et al., 2015). More detailed monitoring of the plant growth status and other characteristics can be achieved using multisensor data fusion technology. This technology was tested by developing a non-autonomous phenotyping multi-sensor platform attached to a tractor as a trailer (Busemeyer et al., 2013). The platform was equipped with 3D time-of-flight (3D-TOF) cameras, a color camera, a laser distance sensor, a hyperspectral imaging system and a light curtain imaging system. The repeatability of these sensors for measuring height and coverage density were analyzed and were very high except for the 3D-TOF and laser distance sensor. The accuracy of the light curtain imaging system for determining the height was denoted by an R 2 value of 0.97 and a mean relative error (MRE) of Figure 4. Rice transplanter lseki PZ60 Figure 5. Wheat-sowing APU-Module Krishnaswamy R.Aravind, Purushothaman Raja and Manuel Pérez-Ruiz Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 8 0.024 with a repeatability that had an R 2 =0.99 and an MRE=0.11. This platform was also used in a plant breeding trial of 25 different genotypes of triticale. Several characteristics of these plants were deduced by processing the multi-sensor data. The author concluded that this device could be further integrated with an autonomous robot in future studies. Bonirob, which was introduced in the “Soil analysis robot” section, was studied using this phenotyping system (Ruckelshausen et al., 2009). This robot used sophisticated sensors, such as a 3D-TOF camera, light curtains and hyper spectral imaging devices, to obtain information that can be extracted from the sensory data; and the generalized health status of the plants was then computed using an algorithm. To obtain repeated phenotypic information of each plant in the field, positional accuracy is an important parameter. Weiss & Biber (2011) used Bonirob with 3D LIDAR to detect the plants and to create a 3D point cloud map of an experimental field with model paper plants, plastic plants and maize plants. The detection rate of the plants was 60-70% in outdoor environments with an average positional accuracy of 0.03 m. Chapman et al. (2014) developed a pheno-copter, which is an autonomous robotic helicopter used for plant phenotyping. This device was equipped with two digital cameras and one far infrared camera to analyze images in multiple spectra. One of the experiments estimated the ground cover of hybrid sorghum using the pheno-copter at an elevation of 60 m to determine the correlation between the number of plants per plot and the green cover for 100 plots. The canopy temperature and relative transpiration index in sugarcane under different irrigation conditions were also estimated using the data obtained from visible and thermal cameras. The potential transpiration index for 40 sugarcane clones based on green cover and relative crop temperature were determined by approximation. To quantify the crop lodging of wheat, images from an NIR filtered camera with information on longitude, latitude, elevation and flight log were used to generate a point cloud elevation model, from which the canopy height was estimated. Córcoles et al. (2013) used a quadrotor UAV to estimate LAI in onion crops using a digital photography camera. They also developed LAI calculation (LAIC) software using the artificial neural network. First, the software tool converts red, green, and blue (RGB) into L*a*b image format and use the K-means algorithm for cluster analysis. The selected clusters are provided as inputs to the artificial neural network, which calculates the area covered by leaves. Statistical analysis was conducted using the minimum value, maximum value, standard deviation, mean and coefficient of variation to establish a relationship between LAI and canopy cover. The maximum value of LAI was 56% in the onion crops. The study found that the polynomial model showed better results than the exponential model and that the coefficient of determination was approximately 84%. Crop pest and disease monitoring and control robots Incidences of disease from pests and microorganisms affect the output of agricultural products wordwide. The majority of these diseases induce visible symptoms in plants; however, farmers can identify a disease in a large field only when a significant number of plants are infected. Polder et al. (2014) developed a fully enclosed, manually propelled platform that is equipped with a diffused fluorescent lamp and a multi-spectral camera (RGB & NIR). The platform is manually moved over each tulip plant as an image of the plant is obtained. Images in the NIR spectrum help segment the image to differentiate the plant from the soil. Diseased plants among healthy plants are identified using Fisher’s linear discriminant classification algorithms. The result is then compared with the enzyme-linked immunosorbent assay (ELISA) score and expert survey results. The results of this study showed that crop experts identified 80% of the diseased plant and misclassified healthy plants as diseased plants 10% of the time. However, the machine vision system correctly identified over 90% of the diseased plant and misclassified 10% of the healthy plants as diseased plants. The author also offered ideas for improving this platform for a robotic system. Other research has discussed using robots to monitor and identify diseases in the field at an early stage. Pilli et al. (2014) developed eAGROBOT to detect diseases in cotton and ground nut plants at an early stage using image processing techniques. A grid data acquisition was used, and sample images were provided as input to a K-means clustering algorithm with varying numbers of nodes for different disease types. A neural network using a single hidden layer with a back propagation technique was used to classify the diseases with highly variable Table 5. Percentage losses from weed competition in India Crop Critical period of crop weed competition Loss (%) Rice (transplanted) 4-6 weeks after planting 15-35 Wheat 30-45 days 6-35 Sorghum 40 days 6-40 Groundnut 30-45 days 30-50 Cotton 15-60 days 47.5 Sugarcane 90 days 15-72 Jute 35-42 days 56-87 Task-based agricultural mobile robots in arable farming: A review Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 9 neurons quantity based on the disease type. The results showed that the Pearson’s r and p values had a poor correlation, while the “contrast of hue” and “correlation of saturation” features showed negative correlation with disease type. The above features along with other textural features (e.g., energy and homogeneity of hue) were used to classify the disease with an accuracy range of 83-96%. Rieder et al. (2014) described the preliminary stage of development of a virtual realitybased system for identifying diseases in strawberry plants using a drone. This system includes a 3D user interface to control and monitor the drone, which will assist in studying the phenology of strawberry plants and disease incidence. Pest infestation and other microbial-based diseases can be controlled by the appropriate application of pesticides. However, misuse of these chemicals can cause pesticide-resistant diseases, and excess application pollutes the environment. Furthermore, the long-term exposure of farmers to these chemicals has harmful effects on the health of farmers. These problems suggest a need for precise applications that are assessed by studying variable rate technology and the site-specific application of pesticides (Bongiovanni & Lowenberg- Deboer, 2004); this precise application can only be achieved by automating the process. Further automation of pesticide application can be achieved by electronic monitors, which measure the application rate and provide a command signal, and rate controllers, which control the rate of application (Zhang & Pierce, 2013). A study on the precise application of pesticides using an UAV was performed by Faiçal et al. (2014). They mainly focused on developing a system architecture using an UAV through simulations. They adjusted the UAV route plan depending on the relative concentration of pesticides in a given area measured using wireless sensor networks for precise application of pesticides. Simulated wireless sensor nodes were positioned at various locations in an experimental field, and the concentration was measured with a color scale varying from green (most concentrated) to red (no pesticide) using the heat map of the field. The behavior of the system was best at a constant light wind of 10 km/h and random gusts of light wind at 10 km/h with no significant difference in measurement at heights of 5 m, 10 m and 20 m. Furthermore, the results show that the chemical dispersion was approximately 14% better for messages sent (from the wireless sensor network to the UAV) every 10 s compared to every 30 s. Sheng (2014) developed a robotic system controlled via wireless network using a mobile phone. The robot was equipped with infrared sensors for obstacle avoidance, a video camera module to obtain visible information, and a sprayer module with a spray head that was adjusted according to the height. The robot was tested on different floors, such as concrete, mud, gravel and grass. The system was able to spray up to a maximum distance of 130 cm with a spray angle of 30–50° and a spray area of 0.88 m2. Weed detection and control robots A weed is an unwanted plant that affects the production and quality of produce on a farm. Weeds compete with crops for resources, which is known as weed competition. According to a survey by Parker & Fryer (1975), production loss from weed competition is estimated to be 11.5% globally, although the loss in individual countries varies because of differing environmental conditions. Table 5 shows the loss of produce due to weeds in India according to the Rice Knowledge Management Portal (http://14.139.94.101/ wisy/prsentation/yield_losses.aspx). Weeds can be classified according to their location in the field; inter-row or intra-row weeds. Inter-row weeds develop between rows of crops, while intra-row weeds develop among crops in the same row. One of the important challenges of using robots for weed control is identifying weeds from crops. Bonirob was equipped with a multispectral monocular camera, known as JAI AD-130 GE, which captured images with 1.3 million pixels and an image size of 1296 × 966 pixels. This robot was tested on a commercially viable carrot field and had an accuracy of approximately 93.8% when differentiating carrot plants from weeds (Haug et al., 2014). Jensen et al. (2012) developed a crop scouting robot known as Armadillo. The robot prototype was developed and tested in a maize field for weed detection, and the researchers. are currently developing a mechanical weeder for the weed removal. Physically removing weeds is a labor-intensive job that requires constant monitoring during the early stage of crop development. Weeds can also be removed by applying chemical herbicides or by using other techniques, such as mechanical force or heat. Blasco et al. (2002) developed a mobile platform with a manipulator that was equipped with an electrode and a secondary vision system. The primary vision system, which is a color camera with a resolution of 768 × 576 pixels, was attached to the front frame of the robot. The images from the secondary vision system were compared with those of the primary vision to locate weeds and to send spatial coordinates to the electrode for weed removal. The electrical discharge from the electrode destroyed the weed tissues, and the system was able to eliminate 100% of the detected weeds. The system was able to locate 84% of the weeds and 99% of the lettuce plants using its vision system. Krishnaswamy R.Aravind, Purushothaman Raja and Manuel Pérez-Ruiz Spanish Journal of Agricultural Research March 2017 • Volume 15 • Issue 1 • e02R01 16 Sahay J, 2006. Elements of agricultural engineering. Standard Publ & Distrib, New Delhi. 462 pp. Sankaran S, Mishra A, Ehsani R, Davis C, 2010. A review of advanced techniques for detecting plant diseases. Comput Electron Agric 72: 1-13. https://doi.org/10.1016/j. compag.2010.02.007 Schirrmann M, Gebbers R, Kramer E, Seidel J, 2011. Soil pH mapping with an on-the-go sensor. Sensors 11: 573-598. https://doi.org/10.3390/s110100573 Scholz C, Moeller K, Ruckelshausen A, Hinck S, Goettinger M, 2014. Automatic soil penetrometer measurements and GIS based documentation with the autonomous field robot platform bonirob. Int Conf on Precis Agr; July. Sheng P, 2014. An intelligent robot system for spraying pesticides. Open Electr Electron Eng J 8: 435-444. https:// doi.org/10.2174/1874129001408010435 Shibusawa S, 2003. On line real time sensing. IEEE Int Conf on Advanced Intelligent Mechatronics; July. pp: 1006- 1066. Slaughter DC, Giles DK, Downey D, 2008. Autonomous robotic weed control systems: a review. Comput Electron Agric 61: 63-78. https://doi.org/10.1016/j. compag.2007.05.008 Søgaard HT, Lund I, 2007. Application accuracy of a machine vision-controlled robotic micro-dosing system. Biosyst Eng 96: 315-322. https://doi.org/10.1016/j. biosystemseng.2006.11.009 Tarannum N, Rhaman MK, Khan SA, Shakil SR, 2015. A brief overview and systematic approach for using agricultural robot in developing countries. J Mod Sci Tech 3: 88-101. Torii T, 2000. Research in autonomous agriculture vehicles in Japan. Comput Electron Agric 25: 133-153. https://doi. org/10.1016/S0168-1699(99)00060-5 Weiss U, Biber P, 2011. Plant detection and mapping for agricultural robots using a 3D LIDAR sensor. Robot Auton Syst 59: 265-273. https://doi.org/10.1016/j. robot.2011.02.011 Yaghoubi S, Akbarzadeh NA, Bazargani SS, Bazargani SS, Bamizan M, Asl MI, 2013. Autonomous robots for agricultural tasks and farm assignment and future trends in agro robots. Int J Mech Mechatron Eng 13: 1-6. Yoon B, Kim S, 2013. Design of Paddy weeding robot. IEEE Int Symp on Robotics; October: 1-2. Zecha CW, Link J, Claupein W, 2013. Mobile sensor platforms: categorisation and research applications in precision farming. J Sens Sens Syst 2: 51-72. https://doi. org/10.5194/jsss-2-51-2013 Zhang Q, Pierce FJ, 2013. Agricultural automation: fundamentals and practices. CRC Press, London. 411 pp. Zhang Z, Noguchi N, Ishii K, Yang L, Zhang C, 2013. Development of a robot combine harvester for wheat and paddy harvesting. IFAC Proc 46: 45-48. https://doi. org/10.3182/20130327-3-jp-3017.00013