A Lightweight Asymmetric RGB–Thermal Multi-Scale Detection Network for UAV Forest Fire Monitoring

Abstract

Forest fires pose severe threats to ecological systems and human safety, yet existing UAV-based detectors struggle to balance accuracy and real-time performance due to severe target scale variations and limited edge computing resources. This paper proposes YOLO-ATF, a lightweight asymmetric RGB–thermal multi-scale detection network built upon RepViT. The model adopts a full-resolution RGB stream paired with a half-capacity infrared stream to leverage complementary visible texture and thermal signals. Three key components are introduced: a P2 high-resolution detection head (stride 4) for tiny fires, a cross-modal feature aggregation network (FAN) with channel-gated residual refinement, and a thermal-guided low-frequency attention (TLA) module, which provide the major gains. A cross-modal alignment loss (CWA) and a parameter-free SimAM module serve as auxiliary components without increasing inference overhead. On the RGBT-3M single-class fire protocol, YOLO-ATF achieves 97.03% mAP@0.5 and 61.03% mAP@0.5:0.95 on the validation set, with only 4.02M parameters and 38.72 GFLOPs under 640 × 640 input, comprehensively outperforming mainstream lightweight detectors and other state-of-the-art fire detectors. The proposed method provides a low-cost, high-accuracy, and deployable solution for early forest fire detection on computationally constrained UAV platforms.