FF-fusion: A knowledge-distilled lightweight visible–infrared image fusion framework for Forest fire monitoring

Abstract

Visible–infrared image fusion offers an effective strategy for forest fire monitoring by integrating complementary information from visible-light and thermal-infrared modalities, particularly under smoke interference, low illumination, and complex forest backgrounds. However, forest-fire-specific paired datasets remain scarce, and lightweight fusion methods have seldom been evaluated under real-world monitoring scenarios and edge-deployment constraints. To address these limitations, this study constructs TF-1770, a spatially aligned visible–infrared forest fire dataset collected from ground-based and UAV-based firefighting scenes. The dataset comprises paired visible-light and thermal-infrared images captured under diverse viewpoints, smoke densities, illumination conditions, and target scales. Based on TF-1770, we propose FF-Fusion, a lightweight visible–infrared image fusion framework based on knowledge distillation. The teacher network learns robust multimodal representations through wavelet-based fusion and cross-modal feature modeling, while the student network incorporates re-parameterizable convolution blocks to enable efficient edge inference. Experimental results on TF-1770 show that FF-Fusion achieves entropy, standard deviation, average gradient, visual information fidelity, and edge information transfer metric values of 7.10, 43.66, 10.83, 0.649, and 0.584, respectively. The deployable student model maintains a compact size of only 1.34 MB. Downstream detection experiments demonstrate that the fused images enhance the perception of flames, smoke, and firefighters. Moreover, edge-deployment tests on heterogeneous platforms verify the feasibility of FF-Fusion for resource-constrained devices. These results indicate that lightweight visible–infrared image fusion can provide effective visual support for practical forest fire monitoring. The proposed workflow, spanning dataset construction, model design, downstream validation, and edge deployment, may also serve as a reference for future deployable ecological monitoring systems.