ForestGuard: A TinyML-Based Acoustic Surveillance System for Intelligent Forest Monitoring

Forests are increasingly threatened by illegal logging, wildlife poaching, forest fires, and unauthorized human intrusion. Conventional surveillance approaches such as manual patrolling, camera traps, and satellite monitoring are limited by high operational costs, restricted coverage, delayed response times, and dependency on visibility conditions. Acoustic monitoring presents a promising alternative, as many harmful forest activities produce distinctive sound signatures. This paper presents ForestGuard, a complete end-to-end TinyML-based acoustic surveillance system for intelligent forest monitoring. The proposed system employs an ESP32 microcontroller integrated with a digital I2S microphone to perform real-time on-device sound classification using a quantized deep learning model. The classifier detects chainsaw activity, gunshots, elephant vocalizations, lion/tiger vocalizations, fire crackling sounds, and human screams, along with an additional “unknown” class to handle environmental background noise. The deployed system achieved 85.2% validation accuracy and 70.53% real-world testing accuracy, with an AUC of 0.95, while maintaining a compact model size of approximately 69 KB suitable for embedded deployment. The architecture integrates a Spring Boot backend server and a React Native mobile application for real-time alert visualization and scalable multi-device monitoring. The results demonstrate the feasibility of scalable, low-cost, edge-based acoustic surveillance for forest protection and wildlife conservation