A lightweight YOLOv11n model with enhanced multi scale feature representation for wildfire detection

The increasing frequency and severity of wildfires pose serious threats to ecosystems, human safety, and property security. Accurate and timely detection of fire and smoke from satellite imagery is therefore essential for early warning and emergency response. However, wildfire detection in satellite images remains challenging due to complex background interference, weak and irregular smoke boundaries, and the limited representation of small-scale fire regions. To address these challenges, this study proposes DME-YOLO, a lightweight wildfire detection model based on YOLOv11n. The proposed model introduces three architectural improvements to enhance feature representation and detection efficiency. First, a C3k2-DWR-DRB module is designed by integrating multi-dilation-rate convolution and re-parameterized convolution into the original C3k2 structure, thereby improving multi-scale contextual feature extraction. Second, a C2PSA-Mona module is developed by embedding a lightweight multi-scale convolutional adapter into the partial self-attention mechanism, which enhances local spatial refinement while maintaining global feature modeling. Third, a Detect-Efficient head based on group convolution is introduced to simplify the prediction branch and reduce computational complexity while preserving multi-scale detection capability. Experimental results on the Wildfire Detection Dataset demonstrate that the proposed DME-YOLO achieves better detection accuracy and computational efficiency than the baseline YOLOv11n. Specifically, Precision, Recall, $$mAP_{50}$$ , and $$mAP_{50:95}$$ are improved from 0.834, 0.737, 0.821, and 0.565 to 0.853, 0.769, 0.844, and 0.601, respectively. Meanwhile, the computational cost is reduced from 6.313 GFLOPs to 5.247 GFLOPs. These results demonstrate that DME-YOLO achieves an improved accuracy–efficiency trade-off, providing accurate wildfire detection with reduced computational overhead. This characteristic makes the proposed model particularly suitable for real-time and resource-constrained wildfire monitoring applications and other resource-constrained remote sensing platforms, where both detection accuracy and computational efficiency are critical for timely early warning and emergency response.