IoT Edge Acoustic Sensing with a Tiny CNN for Real-Time Detection of Illegal Chainsaw Activity in Forests

Abstract

This study focuses on the capabilities of edge IoT devices for long-term monitoring in forest environments. IoT devices vary in capability, which must be evaluated for acoustic sensing. The main aim of this research is to develop a tiny deep-learning model to detect sound events to identify unauthorized chainsaw activity under dense forest conditions. The system runs entirely on a low-power Raspberry Pi node, USB microphone, and a compact CNN model that operates on log-Mel spectrograms. The audio is pre-processed (noise reduction and normalization), converted to time-frequency features, and classified on the device. Using a small, unbalanced dataset and noisy backgrounds, the model produces reliable segment- and event-level detections in real time. During the event-level evaluation, 15 test cases were assessed, and in each case the system detected multiple chainsaw events. These results indicate that tiny models running on inexpensive edge hardware offer a practical solution for real-time ecoacoustic surveillance and can support scalable, cost-effective forest-protection workflows.