DMPNet: A lightweight remote sensing forest wildfire detection network based on multi-scale heterogeneous attention mechanism and dynamic scaling fusion strategy

Unmanned Aerial Vehicle (UAV)-based remote sensing technology has emerged as a critical tool for forest fire detection. However, existing methods face significant challenges in simultaneously achieving high detection accuracy and real-time performance, particularly in scenarios characterized by multi-scale variations, non-stationary behaviors, and complex occlusions. To address these issues, we propose DMPNet, a lightweight neural network specifically designed for forest fire detection. Unlike conventional approaches that sacrifice accuracy for parameter reduction, DMPNet employs an optimized feature extraction architecture (PCSPNet), which not only preserves robust feature representation but also reduces the number of parameters by 16.67% (approximately 500K) compared to YOLOv8n. Furthermore, the dynamic fusion scaling strategy (DSFNet) is integrated into the network's neck to dynamically adjust the size and weight of feature maps, overcoming the limitations of static fusion. Additionally, DMPNet incorporates a multi-scale heterogeneous attention mechanism (MSHA), which effectively addresses occlusion issues through multi-scale contextual reasoning and cross-scale feature interactions. Experimental results demonstrate that DMPNet achieves an inference speed of 1.2 ms, with mAP50 reaching 86.8% and mAP@50-95 at 59%. When compared to YOLOv10n under similar parameter constraints, DMPNet improves mAP50 by 4.2% and surpasses MobileNet, GhostNet, and EfficientNet by over 20% in accuracy. By balancing model compactness, precision, and efficiency, DMPNet provides a robust solution for real-time forest fire monitoring.