XAI-Powered Early Warning System for Honey Bee Colonies Based on Multimodal Sensor Data

Abstract

Winter losses in honeybee colonies pose a critical problem for the sustainability of bee-keeping activities and ecosystem services. This study aims to predict the risk of winter loss in colonies using multi-sensor data collected during the active season and to interpret these predictions using explainable artificial intelligence methods. In this context, a mul-timodal feature set consisting of environmental, acoustic, and phenotypic indicators was created using the Multisensor Bee Phenotypic Dataset. An XGBoost-based supervised classification model was developed for winter loss prediction, and the model's ROC-AUC value was obtained as 0.94. Global and local explainability analyses were performed us-ing the SHAP method to ensure the interpretability of the model outputs. Global analyses showed that environmental variables played a dominant role in the model's deci-sion-making mechanism, while acoustic and phenotypic indicators provided comple-mentary contributions. Local explainability analyses revealed that predictions in individ-ual colonies were formed by the combined effect of multiple variables. Furthermore, the quantitative impact of hypothetical improvements in specific variables on winter loss risk was evaluated through counterfactual analyses. The findings demonstrate that explaina-ble and intervention-oriented AI approaches offer strong potential for developing early warning and decision support systems in precision beekeeping applications.