A systematic literature review on deep learning applications for precision cattle farming

In animal agriculture, deep learning-based approaches have been widely implemented as a decision support tool for precision farming. Several deep learning models have been applied to solve diverse problems related to cattle health and identification. However, an overview of the state-of-the-art of deep learning in precision cattle farming is needed, for which we performed a systematic literature review (SLR). This study aims to provide an overview of the recent progress in deep learning applications for precision cattle farming, in particular health and identification. In the initial search, we retrieved 678 studies from different electronic databases. Only 56 studies qualify the selection criteria, which were then analyzed to extract the data to answer the research questions. The two major applications of deep learning for cattle farming were identified: identification and health monitoring. About 58% of the selected studies are dedicated to cattle identification and the rest for health monitoring. We identified 20 deep learning models that were used to solve different problems, and Convolutional Neural Networks (CNNs) is the most adopted model than others, including Long Short-Term Memory (LSTM), Mask-Region Based Convolutional Neural Networks (Mask-RCNN), and Faster-RCNN. We identified 19 training networks and of which ResNet is by far the most used. From our selection, 12 model evaluation parameters were determined, of which seven were used more than five times. The challenges most encountered with image quality, data processing speed, dataset size, redundant information, and motion of the cattle during data acquisition. In closing, we consider that this SLR study will pave the way for future research towards developing automatic systems for cattle farming.