SBDM-YOLO: an enhanced object detection network with adaptive boundary selection for straw burning monitoring via UAV RGB-TIR fusion

Although straw burning plays a positive role in some aspects of agricultural production, it is also a major source of seasonal air pollution and thus must be strictly controlled. Current detection methods mainly rely on satellite remote sensing but suffer from obvious deficiencies in accuracy and timeliness. This paper presents an unmanned aerial vehicle (UAV)-based method for straw burning detection, termed SBDM-YOLO, which leverages deep-learning image-level fusion of visible-light (RGB) and thermal infrared (TIR) data. The model adopts HGNetv2 as its backbone and incorporates GhostConv for efficient feature extraction, alongside a refined feature enhancement module (C2PSA-Mona) and a boundary-selective SE-attention PAN (BSE-PAN). These components are designed to strengthen the model's ability to distinguish subtle targets while suppressing interference from complex agricultural backgrounds. A hybrid loss function is also introduced to improve localization accuracy. Experimental results indicate that SBDM-YOLO achieves a mean Average Precision (mAP) of 0.929 at IoU (Intersection over Union) = 0.5 and a recall of 0.879, demonstrating superior performance compared to other methods. Significant progress has been made in detecting small-target fire spots, while the false alarm rate has been effectively reduced. This paper proposes an effective technical solution for intelligent monitoring in the agricultural environment, which is of great significance for promoting the sustainable development of agriculture.