Prediction of cow calving in extensive livestock using a new neck-mounted sensorized wearable device: a pilot study

sensors Article Prediction of Cow Calving in Extensive Livestock Using a New Neck-Mounted Sensorized Wearable Device: A Pilot Study Carlos González-Sánchez , Guillermo Sánchez-Brizuela , Ana Cisnal , Juan-Carlos Fraile * , Javier Pérez-Turiel and Eusebio de la Fuente-López   Citation: González-Sánchez, C.; Sánchez-Brizuela, G.; Cisnal, A.; Fraile, J.-C.; Pérez-Turiel, J.; Fuente-López, E.d.l. Prediction of Cow Calving in Extensive Livestock Using a New Neck-Mounted Sensorized Wearable Device: A Pilot Study. Sensors 2021,21, 8060. https://doi.org/10.3390/s21238060 Received: 19 October 2021 Accepted: 30 November 2021 Published: 2 December 2021 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. Copyright: © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). ITAP (Instituto de las Tecnologías Avanzadas de la Producción), Universidad de Valladolid, Paseo del Cauce 59, 47011 Valladolid, Spain; [email protected] (C.G.-S.); [email protected] (G.S.-B.); [email protected] (A.C.); [email protected] (J.P.-T.); [email protected] (E.d.l.F.-L.) *Correspondence: [email protected] Abstract: In this study, new low-cost neck-mounted sensorized wearable device is presented to help farmers detect the onset of calving in extensive livestock farming by continuously monitoring cow data. The device incorporates three sensors: an inertial measurement unit (IMU), a global navigation satellite system (GNSS) receiver, and a thermometer. The hypothesis of this study was that onset calving is detectable through the analyses of the number of transitions between lying and standing of the animal (lying bouts). A new algorithm was developed to detect calving, analysing the frequency and duration of lying and standing postures. An important novelty is that the proposed algorithm has been designed with the aim of being executed in the embedded microcontroller housed in the cow’s collar and, therefore, it requires minimal computational resources while allowing for real time data processing. In this preliminary study, six cows were monitored during different stages of gestation (before, during, and after calving), both with the sensorized wearable device and by human observers. It was carried out on an extensive livestock farm in Salamanca (Spain), during the period from August 2020 to July 2021. The preliminary results obtained indicate that lying-standing animal states and transitions may be useful to predict calving. Further research, with data obtained in future calving of cows, is required to refine the algorithm. Keywords: cow; extensive livestock; sensorized wearable device; monitoring; parturition prediction 1. Introduction The study and monitoring of livestock has always been a subject of great interest. Indeed, quantitative measurement of animal behaviour is an important tool for understanding their reproduction, survival, welfare, and interaction with other animals [ 1 ]. Animal activity is one of the most important indicators associated with animal health and welfare [ 2 ], and animal behaviour is an indicator of the well-being and health of cows [ 3 ]. Detecting changes in the behaviour and activity of cows is a good preventive tool to determine the animal’s health status [4]. Every year an average of 8.5% of perinatal calves are lost due to natural abortions, stillbirths, and complications during parturition (calving) [ 5 ], which translates into higher economical costs and reduced animal wellbeing. Ideally the calving process should be carefully overviewed by experts to avoid or correct any problems that may arise (e.g., dystocia). Even today, the analysis of cow behaviour and calving detection is mainly carried out by experienced workers through unaided monitoring. These approaches are however expensive and time consuming, and not effective for extensive livestock farming, where many animals are kept under grazing in the open air on large areas of surface. An automated solution based on cow data collection from sensors could provide better calving predictions. This will allow the farmer to better identify those cows that require intensive supervision and to focus on caring for cows with upcoming calving, reducing possible risks and improving the health and wellbeing of the animals. Sensors 2021,21, 8060. https://doi.org/10.3390/s21238060 https://www.mdpi.com/journal/sensors Sensors 2021,21, 8060 2 of 15 Internet of things (IoT), an already mature and effective technology, can help improve the efficiency and productivity in agriculture and livestock production systems [ 6 ]. IoT has initially spread into the agriculture and farming industry, and mainly aims to supervise the well-being of animals, thus enhancing the profitability of farms by increasing productivity [ 7 ]. Most connected livestock solutions are developed for cattle, especially cows, due to the valuable price of such animals [8]. Precision livestock farming (PLF) aims to manage individual animals by continuously monitoring their health, welfare, production/reproduction and/or environmental impact in real time [ 9 ]. This is achieved through real-time image [ 10 , 11 ] and/or sound analyses or by wearable devices with sensors that monitor physical (position, direction of movement, speed . . . ) [ 12 – 15 ], and physiological variables (heart rate, breathing rate, temperature . . . ) [16–18] of each animal. Different physiological and behavioural parameters associated with calving can be monitored through sensors. The analysis of the internal temperature and its evolution, usually measured in the vulva, the rectum, or the rumen of the animal, is one of the most accurate calving predictors [ 19 ]. It has been demonstrated that a decrease in vaginal temperature equal or greater than 0.3 ◦ C in cows bearing singletons can predict calving within the next 36 h in 83.3% of cases and up to 100% within 60 h [ 20 ]). However, the core temperature of the animal is difficult to be measured in a non-intrusive way, and the available commercial solutions require intensive veterinary care to install and to check the correct location of the measuring device. This approach is therefore not preferred for use in PLF. An exhaustive meta-analysis of the different publications related to calving detection in cows is showed in [ 21 ]. It concludes that automated monitoring and detection of calving, as well as of dystocia incidents, is possible. However, behavioural changes associated with calving vary between individual animals. Behaviour associated with feeding and rumination descent gradually in the two weeks leading up to calving and is drastically reduced during calving [ 22 ]. The duration of rumination descends up to 33% the day when calving takes place in comparison with the previous day. This behaviour could be successfully measured using ear or neck-mounted devices [23–26]. Another indicator of calving is the increase in lying bouts (LB). This behaviour is associated with the restlessness that the animal feels due to the imminency of the calving. The frequency of lying bouts and their mean duration increase greatly as the calving event approaches, starting already 48 h before and being maximized on the day of calving [ 26 ]. This increase [ 6 ] (from 9.3 ± 1.31 LB/day four days before calving to 13.0 ± 1.02 LB/day the day of calving) is especially important in heifers, but also multiparous dam show more activity prior to calving, and can be observed on average 6 h before calf birth [27]. Lying bouts can be easily identified using a leg-mounted accelerometer [ 28 ], but detection is much more challenging using ear or neck-based sensors due to the similarities in the signals from the accelerometer when the animal is standing and lying. A recent study [ 29 ] reported the use of a neck-based accelerometer to distinguish between those states by detecting the characteristic movement associated with the transition between states. After calving, the number of steps per hour stays elevated, whereas lying bouts tend to gradually decrease as the animal transitions between pre and postpartum states. Tail-raising patterns have been observed to change in the 24 h prior to calving [ 30 , 31 ] and can be monitored using tail-mounted accelerometers. This new approach is however not viable for long-term monitoring due to the weight limitation of these devices, overall reduced stability, and possible damage to the skin of the animal [32]. In calving prediction by traditional methods, the farmer makes a visual inspection of the cow to know its status. This is an error-prone task in which even experts may fail to provide an accurate prediction of calving date. In extensive farming, animals move freely in a wide area, which makes it more difficult for the farmer to properly monitor and manage pregnant cows. An automated solution based on cow data collected from sensors and processed by algorithms, can provide better delivery predictions than visual Sensors 2021,21, 8060 3 of 15 observation. This will allow the farmer to have a more accurate estimation of the expected calving date of the cow and to identify those cows that require intensive supervision due to the proximity of calving. It will make possible both reducing the workload of farmers, who can focus on caring for cows with upcoming calving and improving the health of the cow. In this paper a low-cost neck-mounted sensorized wearable device was designed and an algorithm was developed to detect the onset of calving of cows in extensive livestock farming. This work was divided in three main phases: (1) development of a wearable solution for data collection based on different sensors; (2) data collection in extensive livestock farming by using the aforementioned solution and human-based observations; and (3) development of algorithms to detect the onset of calving and creation of a decision function based on the frequency and duration of lying and standing behaviour (lying bouts). This paper is organized as follows. Section 2presents the sensorized wearable device, the software developed to collect data from different cows by human observers, gives an overview of the collected data, as well as the methodology followed during this study and introduces the proposed algorithm for parturition detection. Results and discussion of this algorithm are presented in Section 3. Finally, Section 4shows the conclusions and future work. 2. Materials and Methods The goal of the first phase of the study is the development of a solution for recording large quantities of behavioural information, which will later be combined with human observation of the animals. The captured information will be used for the development, validation, and quantification of algorithms for calving prediction. A new low-cost sensorized wearable device was developed and integrated into a collar, which can be placed around the cow’s neck. The developed device incorporates three sensors: an Inertial Measurement Unit (IMU), a GNSS and a thermometer. Human observers helped with monitoring, labelling, and recording the animals’ state using our own development PC software. The designed collars were tested on cows from an extensive livestock farm in Salamanca (Spain), during the period August 2020–July 2021. 2.1. Sensorised Wearable Device The sensorized wearable device is a collar which is placed on the animal’s neck. The collar houses a nRF52840-dongle (Nordic Semiconductor, Trondheim, Norway) microcontroller, three sensors (thermometer, 9-axis IMU—3-axis accelerometer, 3-axis gyroscope and 3-axis magnetometer—and GNSS), a microSD card breakout board (AdaFruit, NY, EEUU) for data storage and lithium batteries. Figure 1shows the overall architecture of the collar. Sensors 2021, 21, x FOR PEER REVIEW 4 of 16 Figure 1. Overall architecture of sensorized collar. The microcontroller nRF52840 is a small, low-cost device built around the 32-bit ARM ® Cortex™-M4 CPU which supports short-range wireless standards including BLE. For measuring the cow’s body temperature, a DS18B20 digital thermometer with a temperature range of −55 to 125 °C (accuracy ± 0.5 °C) is placed in such a way that contact between the skin of the cow and the sensor occurs. The microcontroller acquires the 9–12 bits configurable Celsius temperature measurements using a unique 1-wire interface, which only requires one port pin for the communication. The IMU integrated in the collar is the ICM-20948 (InvenSense, Berkeley, CA, USA, EEUU), which is a low-power 9-axis motion tracking device embedding a 3-axis gyroscope, a 3-axis accelerometer, and a 3-axis compass. A SAM-M8Q (U-blox, Thalwil, Switzerland) receiver is used for precise geographic positioning of the animal and is configured to work in PSM (Power Save Management) mode to minimize power consumption. Both the IMU and the GNSS communicate with the nRF52840 microcontroller using I 2 C at 400 KHz. The microSD card, which allows the storage of data from the sensors, communicates with the microcontroller using SPI. Four lithium ion NCR18650GA (Sanyo, Osaka, Japan) batteries of 3.7 V and 3350 mA were used to power the device. Figure 2 shows the developed collar. To avoid damage to the electronic components and the batteries, a protective box was designed and produced using additive manufacturing in a 3D printer (Ultimaker ® 2+). The polymer selected was ABS, a low-cost plastic material with good mechanical properties. The cover box has two areas (Figure 2a,b). The lower part contains the electronics, whereas the upper area houses the batteries. This design facilitates the replacement of the batteries in the collar without affecting the electronic components. To waterproof the container, a 2 mm diameter nitrile rubber O-ring was fitted in the junction between the two covers. The box was attached to the neck of the cows using an adjustable leather belt (Figure 2c), which allows for a good fit on animals of different sizes. The material of the belt is soft, which maximises the animal’s comfort. (a) (b) Figure 1. Overall architecture of sensorized collar. The microcontroller nRF52840 is a small, low-cost device built around the 32-bit ARM ® Cortex ™ -M4 CPU which supports short-range wireless standards including BLE. For measuring the cow’s body temperature, a DS18B20 digital thermometer with a temperature Sensors 2021,21, 8060 4 of 15 range of − 55 to 125 ◦ C (accuracy ± 0.5 ◦ C) is placed in such a way that contact between the skin of the cow and the sensor occurs. The microcontroller acquires the 9–12 bits configurable Celsius temperature measurements using a unique 1-wire interface, which only requires one port pin for the communication. The IMU integrated in the collar is the ICM-20948 (InvenSense, Berkeley, CA, USA, EEUU), which is a low-power 9-axis motion tracking device embedding a 3-axis gyroscope, a 3-axis accelerometer, and a 3- axis compass. A SAM-M8Q (U-blox, Thalwil, Switzerland) receiver is used for precise geographic positioning of the animal and is configured to work in PSM (Power Save Management) mode to minimize power consumption. Both the IMU and the GNSS communicate with the nRF52840 microcontroller using I 2 C at 400 KHz. The microSD card, which allows the storage of data from the sensors, communicates with the microcontroller using SPI. Four lithium ion NCR18650GA (Sanyo, Osaka, Japan) batteries of 3.7 V and 3350 mA were used to power the device. Figure 2shows the developed collar. To avoid damage to the electronic components and the batteries, a protective box was designed and produced using additive manufacturing in a 3D printer (Ultimaker ® 2+). The polymer selected was ABS, a low-cost plastic material with good mechanical properties. The cover box has two areas (Figure 2a,b). The lower part contains the electronics, whereas the upper area houses the batteries. This design facilitates the replacement of the batteries in the collar without affecting the electronic components. To waterproof the container, a 2 mm diameter nitrile rubber O-ring was fitted in the junction between the two covers. The box was attached to the neck of the cows using an adjustable leather belt (Figure 2c), which allows for a good fit on animals of different sizes. The material of the belt is soft, which maximises the animal’s comfort. Sensors 2021, 21, x FOR PEER REVIEW 4 of 16 Figure 1. Overall architecture of sensorized collar. The microcontroller nRF52840 is a small, low-cost device built around the 32-bit ARM ® Cortex™-M4 CPU which supports short-range wireless standards including BLE. For measuring the cow’s body temperature, a DS18B20 digital thermometer with a temperature range of −55 to 125 °C (accuracy ± 0.5 °C) is placed in such a way that contact between the skin of the cow and the sensor occurs. The microcontroller acquires the 9–12 bits configurable Celsius temperature measurements using a unique 1-wire interface, which only requires one port pin for the communication. The IMU integrated in the collar is the ICM-20948 (InvenSense, Berkeley, CA, USA, EEUU), which is a low-power 9-axis motion tracking device embedding a 3-axis gyroscope, a 3-axis accelerometer, and a 3-axis compass. A SAM-M8Q (U-blox, Thalwil, Switzerland) receiver is used for precise geographic positioning of the animal and is configured to work in PSM (Power Save Management) mode to minimize power consumption. Both the IMU and the GNSS communicate with the nRF52840 microcontroller using I 2 C at 400 KHz. The microSD card, which allows the storage of data from the sensors, communicates with the microcontroller using SPI. Four lithium ion NCR18650GA (Sanyo, Osaka, Japan) batteries of 3.7 V and 3350 mA were used to power the device. Figure 2 shows the developed collar. To avoid damage to the electronic components and the batteries, a protective box was designed and produced using additive manufacturing in a 3D printer (Ultimaker ® 2+). The polymer selected was ABS, a low-cost plastic material with good mechanical properties. The cover box has two areas (Figure 2a,b). The lower part contains the electronics, whereas the upper area houses the batteries. This design facilitates the replacement of the batteries in the collar without affecting the electronic components. To waterproof the container, a 2 mm diameter nitrile rubber O-ring was fitted in the junction between the two covers. The box was attached to the neck of the cows using an adjustable leather belt (Figure 2c), which allows for a good fit on animals of different sizes. The material of the belt is soft, which maximises the animal’s comfort. (a) (b) Sensors 2021, 21, x FOR PEER REVIEW 5 of 16 (c) (d) Figure 2. Collar developed for data collection. (a) PCB with electronic components; (b) collar batteries and temperature sensor; (c) collar with belt placed on the cow’s neck; (d) IMU axis orientation in the collar. All electronic components of the collar were configured to work in low power mode, to minimize power consumption, and to extend battery life. The data from the accelerometer and the gyroscope embedded in the IMU were sampled at a frequency of 17.6 Hz. The temperature and geolocation information (longitude, latitude, altitude, and speed) were collected at 1 Hz. The orientation of the IMU axis is shown in Figure 2d. Initial tests were performed using BLE 4.2 for the communication between the board and a computer where data was registered. Due to high power consumption, however, it was decided to store the information onboard instead, using a microSD card. This approach increased the average collar battery life to up to 15 days. Each datapoint includes the information from the sensors and a timestamp, using the format described in Table 1. The data are then saved to the microSD card using binary format. To minimize data loss if the collar runs out of batteries or fails, a new file is created every hour. Although the data transmission is no longer performed wirelessly, the collar still makes use of BLE: during collar initialization a timestamp will be exchanged between the collar and the laptop, which allows for later data synchronization. A keepalive message is also sent periodically (every minute), from collar to laptop, using BLE, to allow the human observer monitoring the cow to detect if the collar is still operative or another action is required (e.g., replacement of batteries). Table 1. Raw data collected by the sensorized collar and saved in the microSD card. Variable Datatype Units Timestamp int32 ms since Unix Epoch Temperature IMU float32 degrees C Temperature DB18B20 float32 degrees C Longitude (Lon) int32 degrees (×10 −7 ) Latitude (Lat) int32 degrees (×10 −7 ) Altitude above sea level (Alt) int32 m (×10 −3 ) Speed int32 m s −1 (×10 −3 ) Acceleration axis X (ax) float32 ×g Acceleration axis Y (ay) float32 ×g Acceleration axis Z (az) float32 ×g Rotation X axis (gx) float32 degrees s −1 Rotation Y axis (gy) float32 degrees s −1 Rotation Z axis (gz) float32 degrees s −1 2.2. Data Annotation with Computer Software Direct visual observations were used to collect states and actions of cows in their natural environment at the cattle farm. A PC-based software (Figure 3a) was designed Figure 2. Collar developed for data collection. ( a ) PCB with electronic components; ( b ) collar batteries and temperature sensor; ( c ) collar with belt placed on the cow’s neck; ( d ) IMU axis orientation in the collar. All electronic components of the collar were configured to work in low power mode, to minimize power consumption, and to extend battery life. The data from the accelerometer and the gyroscope embedded in the IMU were sampled at a frequency of 17.6 Hz. The Sensors 2021,21, 8060 5 of 15 temperature and geolocation information (longitude, latitude, altitude, and speed) were collected at 1 Hz. The orientation of the IMU axis is shown in Figure 2d. Initial tests were performed using BLE 4.2 for the communication between the board and a computer where data was registered. Due to high power consumption, however, it was decided to store the information onboard instead, using a microSD card. This approach increased the average collar battery life to up to 15 days. Each datapoint includes the information from the sensors and a timestamp, using the format described in Table 1. The data are then saved to the microSD card using binary format. To minimize data loss if the collar runs out of batteries or fails, a new file is created every hour. Table 1. Raw data collected by the sensorized collar and saved in the microSD card. Variable Datatype Units Timestamp int32 ms since Unix Epoch Temperature IMU float32 degrees C Temperature DB18B20 float32 degrees C Longitude (Lon) int32 degrees (×10−7) Latitude (Lat) int32 degrees (×10−7) Altitude above sea level (Alt) int32 m (×10−3) Speed int32 m s−1(×10−3) Acceleration axis X (ax) float32 ×g Acceleration axis Y (ay) float32 ×g Acceleration axis Z (az) float32 ×g Rotation X axis (gx) float32 degrees s−1 Rotation Y axis (gy) float32 degrees s−1 Rotation Z axis (gz) float32 degrees s−1 Although the data transmission is no longer performed wirelessly, the collar still makes use of BLE: during collar initialization a timestamp will be exchanged between the collar and the laptop, which allows for later data synchronization. A keepalive message is also sent periodically (every minute), from collar to laptop, using BLE, to allow the human observer monitoring the cow to detect if the collar is still operative or another action is required (e.g., replacement of batteries). 2.2. Data Annotation with Computer Software Direct visual observations were used to collect states and actions of cows in their natural environment at the cattle farm. A PC-based software (Figure 3a) was designed and developed to help the user register the observed state of multiple cows. The program includes the following functionalities: Collar initialization During collar initialization, a timestamp is sent to the collar using BLE. This timestamp will be used for the synchronization of the sensor data collected by the collar and the states and actions of the cows recorded by the human observer. Collar management (Figure 3b) According to the European Commission, individual identification and registration of bovine livestock is mandatory to ensure full traceability and, consequently, enhance food safety and better safeguard animal health. The application uses an alpha-numeric code as the cowl’s ID. Similarly, each device is identified by a unique 64-bit collar ID, which corresponds to the serial number of the microcontroller housed in the collar. Sensors 2021,21, 8060 6 of 15 Sensors 2021, 21, x FOR PEER REVIEW 7 of 16 (a) (b) (c) (d) Figure 3. PC Software for collar management and data collection based on visual observation. (a) Startup screen; (b) collar management; (c) collar scanner; (d) data annotation. 2.3. Calculation 2.3.1. Animals, Facility and Data Collection For this study, ten collars were developed. However, cow monitoring was limited to a maximum of three animals simultaneously in order to improve the quality of the annotated data. The rest of the produced collars were saved as replacements, to minimize dead times and data loss in case a collar stopped working. The monitored beef cattle belong to an extensive livestock farm located in the municipality of Carrascal de Barregas, in the province of Salamanca, Spain. During the study, collars were fitted to six cows (see Table 2). For eleven months (August 2020–July 2021) two experienced observers (one working in the morning shift and the other during the afternoon shift) annotated the actions of the animals for a total of 6 h every day. It is to be noted that although the observers were following the animals continuously, some situations introduced uncertainty in the data, for example, when cows stampede from one location to another. To mitigate the data deterioration, observers left the annotations blank when detecting these situations. To record the different states a laptop running the application previously presented was used while maintaining a clear line of sight with the animals. Every week, the data stored in the microSD card of the collar was downloaded, and the batteries were replaced. A detailed overview of the cow-wise distribution of the data is presented in Table 2. As datapoints were dumped in the SD card on an hourly basis, it is straightforward to know the number of hours we got data from. The total quantities shown in Table 2 accumulate the number of hours the collar was recording data (raw data) and the number of hours the observer labelled behaviours for every studied cow. Table 2. Cow-wise distribution of the data. Cow Raw Data Collection Period (dd/mm/yyyy) Calving Date and Hour Hours of Raw Data Hours of Labelled Data 01 24/08/2020–17/02/2021 01/12/2020 13 h:30′ 1.634 212 02 24/08/2020–25/05/2021 24/02/2021 08 h:30′ 3.417 510 03 01/03/2021–15/06/2021 05/05/2021 16 h:45′ 1.720 279 Figure 3. PC Software for collar management and data collection based on visual observation. ( a ) Startup screen; ( b ) collar management; (c) collar scanner; (d) data annotation. During operation, it was noted that human observers sometimes had difficulty identifying cows when recording actions, especially over long distances or when animals were close to each other. To solve this, each manufactured collar was printed with a different colour (Figure 2c shows a red collar). Collar management allows registering and assigning a new collar to an animal, unregistering, or reassigning an existing one and updating the colour of the collar. Collar scanner (Figure 3c) This feature allows to check active collars within the BLE’s range, by listening to a keepalive message containing the ID that the devices send every minute. The discovered collars are listed on an overview and their status is changed to active (Figure 3c). The application prevents the annotation of data from not active collars. Data annotation (Figure 3d) It allows the annotation, by a human observer, of the status and the action the cow is performing. Additional observations can also be recorded. All annotated data and observations are stored on the PC. Once the annotation activity is finished, the software automatically generates a file containing the recorded information. Two predetermined sets of states have been considered: • General behaviour. The tags considered are: “Grazing/Eating”, “Ruminating”, “Neutral” and “Walking”. •Standing behaviour. The tags considered are: “Standing” and “Lying” position. This reduction was necessary to identify transitions between lying and standing positions (lying bouts) which would not be registered in the general behaviour tag set, where the neutral state can happen both standing and lying (if the cow is doing nothing else). Import data: Once the operator recovers the collar and obtains the sensor data files stored in the microSD, the application allows the generation of a final file, using the timestamp in both files to combine and synchronize the data from the collar with the data from the file containing the visual observation data (stored in the PC). This synchronized data set is used to develop our own parturition prediction algorithm described in Section 2.3.2. Sensors 2021,21, 8060 7 of 15 2.3. Calculation 2.3.1. Animals, Facility and Data Collection For this study, ten collars were developed. However, cow monitoring was limited to a maximum of three animals simultaneously in order to improve the quality of the annotated data. The rest of the produced collars were saved as replacements, to minimize dead times and data loss in case a collar stopped working. The monitored beef cattle belong to an extensive livestock farm located in the municipality of Carrascal de Barregas, in the province of Salamanca, Spain. During the study, collars were fitted to six cows (see Table 2). For eleven months (August 2020–July 2021) two experienced observers (one working in the morning shift and the other during the afternoon shift) annotated the actions of the animals for a total of 6 h every day. It is to be noted that although the observers were following the animals continuously, some situations introduced uncertainty in the data, for example, when cows stampede from one location to another. To mitigate the data deterioration, observers left the annotations blank when detecting these situations. To record the different states a laptop running the application previously presented was used while maintaining a clear line of sight with the animals. Every week, the data stored in the microSD card of the collar was downloaded, and the batteries were replaced. A detailed overview of the cow-wise distribution of the data is presented in Table 2. As datapoints were dumped in the SD card on an hourly basis, it is straightforward to know the number of hours we got data from. The total quantities shown in Table 2accumulate the number of hours the collar was recording data (raw data) and the number of hours the observer labelled behaviours for every studied cow. Table 2. Cow-wise distribution of the data. Cow Raw Data Collection Period(dd/mm/yyyy) Calving Date and Hour Hours of Raw Data Hours of Labelled Data 01 24/08/2020–17/02/2021 01/12/2020 13 h:3001.634 212 02 24/08/2020–25/05/2021 24/02/2021 08 h:3003.417 510 03 01/03/2021–15/06/2021 05/05/2021 16 h:4501.720 279 04 05/10/2020–12/07/2021 25/05/2021 13 h:3502.957 470 05 08/02/2021–30/07/2021 11/07/2021 20 h:1501.130 159 06 24/08/2020–27/01/2021 - 1.887 147 Total 24/08/2020–30/07/2021 - 12.745 1.777 The schema of data acquired with the collars is shown in Table 3. In total, more than 855 million (855,319,572) raw datapoints have been recorded, of which approximately 114 million (114,167,178) are labelled. Table 3. Schema of raw data collected by the collars. DB18B20 GNSS IMU Timestamp Temp Lon Lat Alt Speed ax ay az gx gy gz Temp Previous research work related to monitoring of pregnant cows around calving, (Jensen, 2012) and (Titler et al., 2015), has been focused on the period immediately around the time of calving (one and four days, respectively). This approach however is not practical for application where monitoring is less frequent, such as extensive livestock farms where large herds are held. Therefore, our research was focused on the long-term monitoring of the pregnant animals, which extended up to two months after calving. Using this approach, we could analyse the individual behavioural change during different stages of pregnancy, which previous research has proved can differ greatly between individuals [21]. Sensors 2021,21, 8060 8 of 15 The labels used for data annotation are presented in Tables 4and 5. These two sets of labels distinguish the two datasets introduced before, namely general behaviour and standing/lying behaviour. Considering both label sets, approximately 1.777 h of cow behaviour have been annotated by two experienced observers (working part-time in morning and afternoon shifts). Table 4. General behaviour annotations. ID Action A1 Grazing-Eating A2 Ruminating A3 Neutral A4 Walking Table 5. Standing/lying behaviour annotations. ID Action B1 Standing B2 Lying The distribution of the annotated actions is presented in Figure 4for general behaviour actions and in Figure 5for standing/lying behaviour. Sensors 2021, 21, x FOR PEER REVIEW 8 of 16 04 05/10/2020–12/07/2021 25/05/2021 13 h:35′ 2.957 470 05 08/02/2021–30/07/2021 11/07/2021 20 h:15′ 1.130 159 06 24/08/2020–27/01/2021 - 1.887 147 Total 24/08/2020–30/07/2021 - 12.745 1.777 The schema of data acquired with the collars is shown in Table 3. In total, more than 855 million (855,319,572) raw datapoints have been recorded, of which approximately 114 million (114,167,178) are labelled. Table 3. Schema of raw data collected by the collars. DB18B20 GNSS IMU Timestamp Temp Lon Lat Alt Speed ax ay az gx gy gz Temp Previous research work related to monitoring of pregnant cows around calving, (Jensen, 2012) and (Titler et al., 2015), has been focused on the period immediately around the time of calving (one and four days, respectively). This approach however is not practical for application where monitoring is less frequent, such as extensive livestock farms where large herds are held. Therefore, our research was focused on the long-term monitoring of the pregnant animals, which extended up to two months after calving. Using this approach, we could analyse the individual behavioural change during different stages of pregnancy, which previous research has proved can differ greatly between individuals [21]. The labels used for data annotation are presented in Tables 4 and 5. These two sets of labels distinguish the two datasets introduced before, namely general behaviour and standing/lying behaviour. Considering both label sets, approximately 1.777 h of cow behaviour have been annotated by two experienced observers (working part-time in morning and afternoon shifts) The distribution of the annotated actions is presented in Figure 4 for general behaviour actions and in Figure 5 for standing/lying behaviour. Table 4. General behaviour annotations. ID Action A1 Grazing-Eating A2 Ruminating A3 Neutral A4 Walking Table 5. Standing/lying behaviour annotations. ID Action B1 Standing B2 Lying Figure 4. Distribution of general behaviour annotations. Sensors 2021, 21, x FOR PEER REVIEW 9 of 16 Figure 4. Distribution of general behaviour annotations. Figure 5. Distribution of standing/lying behaviour annotations. General behaviour (Table 4) action prediction [33] has been explored during the study as additional input for calving prediction. Furthermore, data already gathered allows for further investigation in this field without the need of additional human-labelling. However, the general behaviour label set can lead to an imperfect classification of standing/lying behaviour (lying bouts) of the cow when transformed, as some actions can only be carried out in one posture (e.g., walking-standing), but other may occur both standing and lying. For this reason, a second set of labels (Table 5) only accounting for lying and standing postures has been used to ensure a correct classification of these two postures when needed due to the importance of lying bouts detection for calving detection (as discussed in the introduction). The annotated data from human observations show a small delay between the change of action of the cow and the annotation of this new action due to the response time of the observer and their need to operate the annotation software. For this reason, the datapoints recorded two minutes before an action change were ignored. This time period was chosen as a balance between loss of data and minimizing mislabelled data in the dataset. As cows did not change actions with high frequency in the recorded labels, two minutes resulted in enough certainty without losing a significative volume of data for each action. As discussed previously, the sensor readings from the collars and the annotations from the observers can be joined using the timestamp of each data point to form the final dataset. Three datasets were generated using the recorded data: • Non-annotated data from the devices. These data have been proved to be useful for unsupervised and semi supervised learning tasks. • General behaviour annotated data (Table 4). This dataset could be used to classify and predict the animal actions based on new reading from the devices. • Standing/lying behaviour annotated data (Table 5). This dataset, although smaller compared to b), serves for statistical learning tasks, as well as semi-supervised learning techniques. The experiments and algorithms have been developed using Python 3.7 [34] along several libraries, mainly: Pandas [35], Keras [36], Numpy [37], and SciPy [38]. Figures have been plotted using the Seaborn [39] library. 2.3.2. Parturition Prediction Algorithm Calving prediction is the main objective of this pilot study. To usefully notify parturition, it is necessary to detect it with enough anticipation using a low-memory algorithm suited for the microcontroller. The number of transitions of the animal between lying and standing has been empirically proved to be a good indicator of parturition in different studies [27,40,41]. This Figure 5. Distribution of standing/lying behaviour annotations. General behaviour (Table 4) action prediction [ 33 ] has been explored during the study as additional input for calving prediction. Furthermore, data already gathered allows for further investigation in this field without the need of additional human-labelling. However, the general behaviour label set can lead to an imperfect classification of standing/lying Sensors 2021,21, 8060 9 of 15 behaviour (lying bouts) of the cow when transformed, as some actions can only be carried out in one posture (e.g., walking-standing), but other may occur both standing and lying. For this reason, a second set of labels (Table 5) only accounting for lying and standing postures has been used to ensure a correct classification of these two postures when needed due to the importance of lying bouts detection for calving detection (as discussed in the introduction). The annotated data from human observations show a small delay between the change of action of the cow and the annotation of this new action due to the response time of the observer and their need to operate the annotation software. For this reason, the datapoints recorded two minutes before an action change were ignored. This time period was chosen as a balance between loss of data and minimizing mislabelled data in the dataset. As cows did not change actions with high frequency in the recorded labels, two minutes resulted in enough certainty without losing a significative volume of data for each action. As discussed previously, the sensor readings from the collars and the annotations from the observers can be joined using the timestamp of each data point to form the final dataset. Three datasets were generated using the recorded data: • Non-annotated data from the devices. These data have been proved to be useful for unsupervised and semi supervised learning tasks. • General behaviour annotated data (Table 4). This dataset could be used to classify and predict the animal actions based on new reading from the devices. • Standing/lying behaviour annotated data (Table 5). This dataset, although smaller compared to b, serves for statistical learning tasks, as well as semi-supervised learning techniques. The experiments and algorithms have been developed using Python 3.7 [ 34 ] along several libraries, mainly: Pandas [ 35 ], Keras [ 36 ], Numpy [ 37 ], and SciPy [ 38 ]. Figures have been plotted using the Seaborn [39] library. 2.3.2. Parturition Prediction Algorithm Calving prediction is the main objective of this pilot study. To usefully notify parturition, it is necessary to detect it with enough anticipation using a low-memory algorithm suited for the microcontroller. The number of transitions of the animal between lying and standing has been empirically proved to be a good indicator of parturition in different studies [ 27 , 40 , 41 ]. This measure serves as an indicator of the proximity of calving due to the relative increase of its value in the 8 to 2 h before the parturition event. Furthermore, a notable decrease in the number of lying bouts in the hours after calving is also observed. However, these works study intensive dairy cattle, while our work studies extensive beef cattle, with the according significantly less restricted environment since calving barns are not used. A new low-memory algorithm based on classification of two cow postures (lying and standing), from the collar sensors readings has been developed. To classify these behaviours, accelerometer readings, commonly used to distinguish between lying and standing, as well as GPS altitude readings were initially considered. However, GPS readings were discarded due to insufficient sensor resolution. In [ 21 ] is indicated that is more difficult to distinguish between a lying and standing position with a neck-based accelerometer since the two positions show similar accelerometer readings. Leg-based accelerometers show a distinct crossover of two axes and can easily be utilized to determine a standing or lying position. However, we have analyzed accelerometer signals read by our neck-mounted collars on the Y and Z axes (Figure 6). This figure shows that the analysis of the accelerometer signals provided by the cow’s collar, allow us to clearly distinguish cow lying position (green colour), and cow standing position (blue colour). These results are similar to those presented in [ 29 ] to classify the cow posture as standing or lying.