Key points are not available for this paper at this time.
Detecting forest fires at the incipient stage enables timely suppression before escalation. Unmanned aerial vehicles (UAVs) equipped with visible-light and thermal-infrared sensors can capture complementary information for this task, but the high computational cost of existing multimodal detectors limits real-time deployment on UAV platforms. This paper proposes RTA-FireNet, a lightweight RGB-thermal detector built on YOLO11 for early forest fire detection. RTA-FireNet adopts an asymmetric architecture in which the visible branch serves as the primary feature extractor and the thermal branch provides auxiliary cues at selected stages, thereby avoiding the overhead of symmetric dual-stream processing. The detector integrates an Asymmetric Cross-Modal Backbone (ACBackbone), a Bidirectional Fusion Neck (BFNeck), and a Semi-Coupled Detection Head (SCHead). Experiments on the RGBT-3M dataset show that RTA-FireNet achieves an AP of 66.0%, an AP50 of 95.2%, and an F1-score of 94.0%, with only 2.2 M parameters and 9.9 GFLOPs. Compared with other RGB-thermal detectors evaluated in this study, RTA-FireNet achieves the highest AP and F1-score with the fewest parameters and the lowest computational cost, making it suitable for real-time early fire detection on resource-constrained UAV platforms.
Yang et al. (Mon,) studied this question.