Smart Cattle: Cattle Live Weight Estimation Based on a Deep Learning Approach

Smart Cattle is a suite of technologies for boosting the quality and quantity of production in dairy farms. Measuring cattle’s live weight is essential as the captured data could be used to monitor cattle’s health and nutrition management. However, weighting cattle and consistently monitoring its weights of is time consuming and require intensive labour. Profitability of beef production system depends highly on the annual live weight gain, stocking rate, feed quality, feed quantity, and the balance between these factors. The lighter cattle causes farmers to lose profit. In feedlot, where the feed can be systematically rationed, this problem can be mitigated with feeding correction. The limitation of this approach is to monitor the live-weight and growth rate of the cattle, as the farmers need to weigh the animals at consistent interval such as once in a month during the production period. This task is laborious, time consuming and costly. Hence, the aim of this work is to develop a live weight estimation system that will be able to perform a physical weight estimation of a cattle using a deep learning approach. The objectives of this work are to design a deep learning model, specifically a Convolutional Neural Network (CNN) for cattle weighting estimation system based on features extracted from the images, to implement an optimized deep learning model for cattle physical weight estimation based on high resolution images and finally to evaluate the performance of the proposed cattle weight estimation system using a deep learning model. This approach can be used to replace traditional methods of weighing using electronic scale and direct observation and measurement can be made based on the cattle images. In this work, the impact of dropout rates on the performance of convolutional neural networks will also be studied for image classification. Based on the results obtained, the proposed CNN model with dropout regularization produced higher performance accuracy of 54.0%.