Application of deep learning algorithms to detect and count cattle from remotely sensed imagery
MetadataShow full item record
Effective livestock monitoring and control necessitate an accurate and up-to-date information on animal abundance. The paucity of accurate data on animal abundance has made the monitoring of greenhouse gas emission from livestock farming difficult. Manual counts use analysts to detect and identify animals in Remotely Sensed imagery but this can be time consuming, subjective and costly due to the labor costs for image pre-processing involved. Similarly, on ground count is also very expensive due to large number of human resources involved and limited by small spatial coverage per period. A good estimation of the number of animals is required for reliable reporting of GHG emissions over large areas. The main objective of this study was to integrate the use of convolutional neural networks (CNNs) algorithms to detect and count the number of cattle in high resolution images from Unmanned aerial vehicles (UAVs), and/or satellite-based platforms. The specific objectives were to develop and train an object detector model with aerial images of cows and to test the algorithms for performance accuracy. Remotely sensed images were acquired, preprocessed and annotated using custom python scripts. The You Only Look Once (YOLO) object detector model that utilizes convolutional neural network (CNN) architecture was adopted in building the proposed cattle detection algorithm. Tracking-by-detection approach was used for counting the cattle by applying SORT (Simple Online and Real-time Tracker) algorithms that tracks each detection by assigning unique identity (ID) to each bounding box, and as soon an object is lost due to occlusion, wrong identification, etc. the tracker assigns a new unique identity (ID) and start tracking the newfound object. The algorithm was trained on a total of 2,646 images with varying resolutions to increase its performance. Detection performance was further enhanced by optimizing the input resolution (to 768X768) during testing, and obtained high values of precision, recall and F1-score of up to 100% for high resolution images and fairly isolated animals. Also high Recall of 99.1%, precision of 96.9% and 96.5% F1-score were obtained for lower resolution images. In conclusion, with these results, the developed deep learning algorithm can be deemed scalable for real time cattle detection and counting and hence offering a solution to the paucity of accurate data on animal abundance and avoiding the drawbacks associated with the currently manual technique of animal census.