Feature Extraction and Random Forest to Identify Sheep Behavior from Accelerometer Data

Abstract

Sensor technologies play an essential part in the agricultural community and many other scientific and commercial communities. Accelerometer signals and Machine Learning techniques can be used to identify and observe behaviours of animals without the need for an exhaustive human observation which is labour intensive and time consuming. This study employed random forest algorithm to identify grazing, walking, scratching, and inactivity (standing, resting) of 8 Hebridean ewes located in Cheshire, Shotwick in the UK. We gathered accelerometer data from a sensor device which was fitted on the collar of the animals. The selection of the algorithm was based on previous research by which random forest achieved the best results among other benchmark techniques. Therefore, in this study, more focus was given to feature engineering to improve prediction performance. Seventeen features from time and frequency domain were calculated from the accelerometer measurements and the magnitude of the acceleration. Feature elimination was utilised in which highly correlated ones were removed, and only nine out of seventeen features were selected. The algorithm achieved an overall accuracy of 99.43% and a kappa value of 98.66%. The accuracy for grazing, walking, scratching, and inactive was 99.08%, 99.13%, 99.90%, and 99.85%, respectively. The overall results showed that there is a significant improvement over previous methods and studies for all mutually exclusive behaviours. Those results are promising, and the technique could be further tested for future real-time activity recognition.

Classifying the posture and activity of ewes and lambs using accelerometers and machine learning on a commercial flock

“…Random forest was chosen as it is relatively easy to inspect feature importance, is fairly robust to data imbalances, and has previously been used to classify behaviour in sheep with a high degree of success (e.g. Kleanthous et al, 2018Kleanthous et al, , 2019Kleanthous et al, , 2020Lush et al, 2018;Mansbridge et al, 2018;Walton et al, 2018;Kaler et al, 2020). When training the classifier, we specified an mtry (the number of variables tried at each split) equal to the square root of the number of variables used for classification (approximately 1.7) and ntree (the number of trees grown) of 500 (Barwick et al, 2018b;Fogarty et al, 2020a).…”

A survey of machine learning approaches in animal behaviour

“…Accelerometers measure the acceleration of motion and are a ubiquitous type of sensor in activity recognition problems because they are light weight, small in size, inexpensive, and offer low power consumption [18,34,[42][43][44][45][46]. Studies have reported that activities such as walking, grazing, scratching, lying, and standing can be easily recognized by using only accelerometers, yielding overall accuracies in excess of 98% [47,48]. Likewise, running activity was detected with an accuracy of 96.62% using information from only one acceleration axis of accelerometer data [49].…”

“…Eating sward at ground level with the head down. [2,5,34,37,47,48,51,52,55,56,57,[59][60][61][62][64][65][66] Infracting Eating from branches above a certain height.…”

“…[2, 5,6,34,37,47,50,51,53,[58][59][60][61][62][63][64]66] Moving This is an intended movement from one place to another. Naturally, the sheep is not looking for nutrition.…”

“…This step is critical since accelerometer signals provide valuable information about motion patterns in slots of data and not as a single variable measure. The windows are of the same size and are either disjoint [2,5,34,37,47,54,57,59,66] or overlapping, typically at 50% [77]. Overlapping windows are suggested because of their ability to capture the transitions of activities more precisely [77].…”