A lightweight deep learning framework for real-time smouldering forest fire detection on edge devices

Smouldering forest fires are characterized by low combustion temperature, flameless burning, weak thermal radiation signature, and prolonged latency period, presenting significant detection and monitoring challenges. These characteristics make early identification difficult, and without timely intervention, such fires readily escalate into large-scale wildfires that cause severe ecological degradation. Existing monitoring methods have critical limitations: satellite sensing suffers from occlusion and low resolution, ground patrols have limited coverage and high costs, and fixed sensors lack field reliability. This paper proposes YOLOv10-WT, a lightweight detection framework for UAV-based forest surveillance. The framework achieves balance between weak feature extraction and lightweight deployment through four optimized modules working in concert. The wavelet basis feature enhancement module amplifies high-frequency thermal radiation characteristics of smouldering fires through Haar wavelet decomposition and reconstruction. The GhostConv module reduces parameter redundancy. A dual-branch ECA attention mechanism strengthens fire-point channel localization and spatial precision. The GatedSConv module suppresses background noise interference. Experimental validation on 5,128 thermal infrared images acquired by UAVs demonstrates that the proposed framework achieves 93.97% mAP@0.5 for smouldering forest fire detection, representing a 2.44% point improvement over YOLOv10n. With only 2.19 million parameters, the model achieves real-time inference at 6 FPS on Jetson Nano edge devices, satisfying operational requirements for autonomous UAV-based forest patrols. This work provides an effective early-warning solution for resource-constrained forest monitoring, addressing core smouldering fire detection challenges.