Forest Fire Detection Based on Enhanced Feature Information Capture and Long-Range Dependency

Accurate and timely forest fire detection is critical for early warnings and effective wildfire containment, safeguarding both ecological systems and human lives. However, in the early stages of a fire, the flames are typically small, and the smoke often bears a close resemblance to mist or fog, which makes prompt and accurate detection difficult and frequently causing the loss of the optimal response window. To tackle this challenge, we present an enhanced single-stage object detector, that excels at identifying fires, especially in their nascent stages. First, we incorporate a contrast enhancement technique that leverages an atmospheric scattering model and entropy optimization to clarify the blurred smoke edges. Next, we integrate deformable convolutions into the original feature extraction module, enabling more effective capture of long-range dependencies and spatial relationships, thus improving multi-scale fire detection accuracy. Additionally, we introduce a novel Hybrid Spatial Pyramid Pooling (HSPP) structure, which uses parallel average and max pooling operations to enrich feature extraction, facilitating better recognition of small-scale early fires. Finally, we propose a Dynamic Channel Attention (DCA) module to adaptively optimize cross-channel interactions across multiple feature extraction layers, bolstering cross-scale feature representations. Extensive experimental results show that on the M4SFWD and FLAME public datasets, our method achieves a 1% improvement in both precision and recall compared to the SOTA methods, while reducing the number of parameters by about 40%, making it more suitable for practical application.