Application of AI in Smart Farming: A Deep Learning–Based Study for the Udder Segmentation and Milk Yield Prediction

Several methods for yield prediction have been proposed. Milk yield is predicted for either a lactation curve (305 days), six months, or three months. Because the milk yield varies along the lactation curve, predicting daily milk yield (DMY) is challenging. Prediction based on an average value will lead to inaccurate results. Interestingly, mammary gland (called udder) parameters are significantly related to DMY. Collecting these parameters is difficult because of the cows’ violent behavior. The proposed framework for yield prediction is based on udder size measured from images, previously recorded milking data, and deep learning techniques, including Mask Region–based Convolutional Neural Networks with Instance Segmentation (MRCNN‐IS), Segment Anything Model‐2 (SAM‐2), and Detection Transformers (DETR). There were 148 cows, and 3000 images of 10 lactating cows were collected in six months. The DETR, SAM‐2, and MRCNN‐IS models performed well, with the best attained average of 98% accuracy, 98% precision, and 97.5% recall on the dataset. The DETR algorithm was used on the validation set to measure accuracy for each object detection model dataset to generate a lactation curve. The MRCNN‐IS achieved the best accuracy of 98.7%, with MAE = 0.039 and RMSE = 0.015. The results showed that using deep learning models and computer vision, the proposed technique has achieved better results and ease than the available milk yield prediction techniques.