Developing an interpretable machine learning model for the detection of mimosa (Albizia julibrissin Durazz) grazing in goats

Recent advancements in machine learning for detecting animal behaviors, particularly goat activities, have faced challenges due to their complexity and lack of explainability in practical applications. This article presents an interpretable machine-learning framework using sensor-based data to differentiate mimosa grazing from other goat activities like grazing herb, resting and walking. BORUTA, an algorithm for selecting the most relevant features, and SHAP, a technique for interpreting the decision of a machine learning model are two fundamental components of the methodology used for developing the model. The resulting model, a gradient boost algorithm with 15 selected features has shown robust performance with accuracy, sensitivity, and precision between 82% and 86%. SHAP analysis further elucidates the model’s decision-making, highlighting the impact of features like ’Standing’ and ’%HeadDown,’ along with distance-related features on discriminating grazing mimosa from grazing herb. The simplicity of the model advocates for its potential in real-time systems and underscores the importance of explainability in improving and deploying these models in real-world scenarios.