To address the challenges of feature attenuation, background interference, and real-time bottlenecks in forest fire detection, this study proposes a novel algorithm based on long-range dependency modeling and dynamic feature fusion. First, a long-range-local dual-mode feature enhancement mechanism is constructed by integrating the multi-scale perception capabilities of the strip pooling and pyramid pooling, which strengthens the global morphological representation and local detail extraction of target features. Second, a spatially enhanced attention head (Detect-SEAM) is designed to suppress background noise through a channel-spatial dual-attention coordination mechanism, thereby enhancing the spatial feature response of occluded targets. Finally, the dynamic upsampling operator (DySample) is introduced into the neck network to adaptively adjust the sampling strategy based on input features, reducing feature information loss while balancing detection accuracy with real-time performance. Experimental results on the Wildfire dataset show that the improved model reaches 86.5% mAP with a 3.7% increasement, 85.1% precision with a 2.2% increasement, 78.2% recall with a 3.4% increasement, and an inference speed of 68.86 f/s. The proposed model realizes a synergistic optimization of detection accuracy and inference efficiency, offering an effective solution for forest fire detection.