Wildfire-Diff: A Controllable Diffusion Framework for Synchronized Generation of Remote Sensing Wildfire Images and Masks

Burned area mapping based on remote sensing images is a key task in wildfire monitoring and holds significant importance for disaster assessment. In recent years, deep learning (DL) methods have made significant progress in this task and are gradually becoming the mainstream approach. However, the issue of existing training sample scarcity severely limits the accuracy and generalization of DL-based burned area mapping methods. To address this issue, we propose a controllable diffusion model, namely Wildfire-Diff, for synchronized generation of paired wildfire images and pixel-level masks. Different from current conditional diffusion models, we strengthen the realness of generated samples by constraining the spatial distribution and background of wildfires. Specifically, we design a vegetation-constrained wildfire mask generation network that can restrict burned areas to vegetation-covered regions, thereby reducing unrealistic fire scenes (e.g. wildfire on the river). Furthermore, we develop a background-aware wildfire image generation network to reduce the randomness of generated image background and mitigate the distortion issue. Benefiting from these two frameworks, our Wildfire-Diff can generate high-quality wildfire change detection samples to promote burned area mapping. Experimental results demonstrate the superiority of our Wildfire-Diff (30.42 FID) compared to other advanced diffusion models (e.g. ControlNet). Meanwhile, in downstream burned area mapping tasks, applying wildfire samples generated by our Wildfire-Diff can significantly improve the accuracy (IoU) of four change detection models (i.e., BIT, CGNet, EfficientCD and BAMCD) by at least 20%, which further reveals its great potential to address training sample scarcity and promotes downstream applications.