Udderly accurate: A deep learning based modelling for determination of dairyness of Sahiwal cow using computer vision

Abstract

Dairy farming is a crucial agricultural practice for food and nutritional security. Selection of the best milching animals has always been a challenging task for optimising production efficiency. The efficacy of milk yield production is contingent upon a number of factors including inherent linear traits. Linear traits refer to the quantifiable physical characteristics that are associated with the production and reproduction capabilities of dairy animals. This article presents an innovative approach for the classification of Sahiwal cows into high, medium and low yielder categories based on images featuring linear traits through emerging deep learning and computer vision techniques. A large dataset of 4110 images highlighting important linear traits such as udder size, shape, and texture of different categories of Sahiwal cows has been created for training, validation and testing of the model. Images were collected from the herd of Sahiwal cows maintained at the National Dairy Research Institute, Karnal. The dataset was pre-processed using image augmentation techniques to enhance the model’s robustness. Different architectures of CNN models namely InceptionV3, ResNet50 and GoogleNet were trained and optimised. The Inception V3 model demonstrated the best result with 85.64% testing accuracy among all these models in classifying the cow. The developed model can be used under field conditions to determine the dairyness of a cow in real time mode in place of human experts. Additionally, the model’s interpretability is evaluated through feature visualisation, showcasing the importance of different udder features in milk yield prediction.