SAR-To-Optical Translation Using Conditional Diffusion Models for Wildfire-Burned Area Segmentation

This study presents a Conditional Denoising Diffusion Probabilistic Model (CDDPM) for Synthetic Aperture Radar (SAR) to Optical image translation, specifically focusing on areas affected by wildfires. Given the escalating wildfire incidences due to climate change, satellite monitoring is an essential tool for firefighting efforts. Although Sentinel-2 imagery provides accurate data for burned area mapping, its effectiveness is strongly influenced by cloud cover. While this obstacle is mitigated by Sentinel-1, which can conduct surface measurements even during sub-optimal visibility conditions, its inherent noise instead makes segmentation difficult. Our pro-posed model utilizes a multi-image, multi-temporal approach that integrates prior Sentinel-1/2 data as well as current Sentinel-1 data for context-aware output. After translation, a U-Net is trained to determine the burned area. For enhancing the efficacy of the burned area prediction, our diffusion model simultaneously learns mask generation and translation tasks. The model, trained on a dataset including 541 wildfire events in Canada from 2017 to 2021, significantly outperforms existing methods in segmentation metrics.