DGIM-SAM: dual-kernel gated interaction modulation and spectral adaptive modulation for real-time wildfire detection

Wildfire detection in unconstrained natural scenes is challenging due to weak fire responses, ambiguous smoke boundaries, and interference from clouds, haze, vegetation, and illumination changes. To address these challenges, we propose DGIM-SAM , a real-time wildfire detection framework that combines dual-kernel gated interaction modulation with spectral adaptive modulation. DGIM-SAM consists of three key components. First, the Dual-Kernel Enhanced Cross Stage Partial block enlarges the effective receptive field through dual 3 × 3 convolutions, improving the representation of irregular flames and diffuse smoke. Second, the Spectral Adaptive Feed Forward Network adaptively modulates frequency responses to enhance discriminative high-frequency fire cues while suppressing low-frequency environmental noise. Third, the Channel Linearized Alignment Gating Fusion module aligns multi-scale features by jointly modeling grid-level spatial dependencies and channel-wise semantic importance, reducing feature mismatch during fusion. Experiments on D-Fire, FASDD, FLAME and a Self-Constructed Wildfire Dataset with 18 000 images demonstrate that DGIM-SAM achieves strong accuracy and efficiency. DGIM-SAM achieves an average mAP 50 : 95 of 55.6% and 145 FPS on a desktop GPU, while reaching 38.7 FPS with 25.8 ms latency on the Jetson Orin Nano, demonstrating strong accuracy, efficiency, and edge-deployment feasibility. These results show that DGIM-SAM provides an effective and scalable solution for real-time forest surveillance on resource-constrained edge devices.