ForestGuard: An IP66 Edge-AI Raspberry Pi Node for Illegal Logging and Early Fire/Smoke Detection

Illegal logging and wildfires cause rapid biodiversity loss and economic damage, yet remote forests lack affordable, robust monitoring. This paper presents ForestGuard, a tiny, IP66-rated edge-AI node built on Raspberry Pi 4B that fuses acoustic, inertial, and gas/optical sensing to detect illegal tree cutting and early fire/smoke events in real time. A USB microphone streams audio to a TensorFlow model exported from Teachable Machine; the classifier was trained on 100 samples per class (chainsaw, drilling, axe strikes) and achieves 94% overall accuracy on held-out audio. An MPU6050 inertial unit monitors free-fall/impact as a proxy signal for branch severance or device tampering, while an MQ135 gas sensor and a digital flame detector—both read through an ADS1115 over I²C—provide smoke and ignition cues. Multimodal decisions are fused with rule-based logic to minimize false alarms; when any activity exceeds calibrated thresholds, the node immediately emails the control room with timestamp and event type. All sensing, inference, and messaging execute on-device to avoid latency and to remain functional in low-connectivity settings. Hardware is packaged in an enclosure meeting IP66, enabling unobtrusive deployment on trees or poles. Key details include duty cycling, denounced inertial interrupts, and I²C coordination between ADS1115 and MPU6050. Bench and outdoor trials show reliable recognition of target acoustic events and timely fire/smoke indication, with rapid notifications during combined scenarios. The contributions are: (i) an IP66 edge device unifying audio AI with inertial and gas/optical sensing; (ii) a lightweight training-to-deployment pipeline using Teachable Machine and TensorFlow on Raspberry Pi; and (iii) a simple alerting workflow suitable for ranger operations. We also detail threshold calibration for MQ135 using controlled smoke sources and validate flame sensing with open-flame tests, improving sensitivity without spurious triggers. Future work will expand the sound taxonomy, add continual learning, and integrate long-range radios and camera snapshots for evidence capture.