Deep learning for lameness level detection in dairy cows

Abstract

Lameness detection using raw sensor data is a very challenging task, as the data are devoid of specific information regarding predictors such as gait distribution, weight among legs, etc. We have addressed this challenge using a deep learning technique, named LLP-Cow (Lameness level predictor for Cow), which is an application of artificial intelligence (AI). For objective comparison, LLP-Cow is validated using CowScreeningDB, an unbalanced public dataset composed of sensor data. This dataset is recorded during the normal life of dairy cows. Hence, LLP-Cow models the normal behaviour of cows and consists of feature extraction, application-specific deep network and a voting system. The technique presented is able to model the behaviour of a cow for both binary and multiclass classification. The precision and specificity reported by our technique stand at 0.94 and 0.98 for multiclass and 0.91 and 0.90 for binary protocols for the best case scenario. Moreover, F1 measure, Matthews correlation coefficient and Kappa are 0.94, 0.91, and 0.91, respectively. The technique introduced provides a margin for human intervention through the use of a voting system at the classification stage. The technique presented is therefore an implemented AI system for cow lameness detection that offers room for exploration in terms of real time implementation.