Adaptive Measurement Noise for Robust Kalman Filtering in Smart Beehive Telemetry

Abstract

Honeybee colony monitoring generates multimodal, non-stationary telemetry streams that require reliable recursive state estimation with well-calibrated uncertainty for digital apiculture. Although Kalman filtering is widely used in environmental monitoring, adaptive measurement-noise modeling has not been systematically evaluated for smart-hive telemetry under leak-free chronological protocols. The approach performs online measurement-noise covariance adaptation using innovation statistics identified as the primary source of calibration improvement, while an innovation-based normalized innovation squared (NIS) gate serves as a secondary robustness safeguard against rare extreme observations. The framework is benchmarked against fixed-noise linear Kalman filtering, the Extended Kalman Filter (EKF), the Ensemble Kalman Filter (EnKF), and ARIMA/SARIMAX baselines, with EKF and EnKF included primarily as stability reference baselines. Experiments are conducted using a strictly chronological train–validation–test replay protocol with one step-ahead forecasting on longitudinal data from instrumented honeybee colonies. Model performance is assessed using predictive accuracy, probabilistic calibration and innovation-based consistency diagnostics. Validation-based sensitivity and ablation analyses are included to evaluate robustness and to distinguish the contribution of adaptive measurement-noise estimation from the auxiliary robustness role of the normalized innovation squared (NIS) gating mechanism.