The SD-YOLO Model for Complex Scenarios: A New Approach to Enhance Fire and Smoke Detection Performance

Fire detection is critical for safeguarding human life and property. However, traditional detection methods face challenges of high feature redundancy and insufficient multi-scale feature fusion in complex environments. To address these limitations, this study proposes SD-YOLO, a fire detection model based on the YOLO framework, which integrates the SCConv spatialchannel module and SimAM attention mechanism to improve detection performance. Specifically, in the backbone network, SCConv replaces the original C3K2 module. Through spatial redundancy filtering and channel adaptation, it reduces redundant computations and model parameters while boosting detection accuracy for small-scale smoke targets. In the neck network, a novel feature fusion architecture (DFPN) is introduced to optimize multi-scale feature integration. With the SimAM attention mechanism, the model effectively suppresses interference from clouds, reflections, and other irrelevant elements. Experimental results show that SD-YOLO achieves a recall rate of 76.5% ($\mathbf{1. 0 \%}$ higher than the baseline YOLO11) and a precision rate of $\mathbf{8 5. 2 \%} \boldsymbol{(} \mathbf{0. 9 \%}$ higher than YOLO11). This model is well-suited for forest patrol drones and industrial monitoring systems, providing a reliable solution for early fire warning and enhancing object detection in safety-critical domains.