Let Me Sleep! A Machine Learning Approach for IoT Wake-Up Mechanism for Insect Detection

Agricultural pest monitoring IoT systems face a critical dilemma: continuous surveillance drains batteries within weeks while generating massive amounts of irrelevant data that overwhelm edge processing capabilities. While high-frequency sampling ensures improved detection accuracy, it drastically shortens operational lifetime in battery powered deployments and generates tons of non-relevant data hardly to be separated by relevant ones. This work introduces a smart wake-up mechanism that leverages machine learning and entomological insights to optimize energy efficiency and data volumes without compromising monitoring effectiveness. We formulate image acquisition (i.e., monitoring) as an anomaly detection problem, where a lightweight classifier activates image capture only under meteorologically and biologically relevant conditions. Using a dataset of six months of Brown Marmorated Stink Bug (BMSB) activity records in a pear orchard, we demonstrate that our method reduces image captures by 74% compared to fixed-interval sampling, while capturing up to 78.1% relevant images. The system integrates environmental data (i.e., temperature, humidity) and insect phenology metrics (degree-day accumulations) to predict optimal sensing times.