Skip to content
Open access

SAR-to-Optical Translation Using Swin-U-DiT for Burned Area Mapping

2026 · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · Vol 19, pp. 36369-36381 · 0 citations · 48 references

Abstract

Accurate and timely burned area mapping is critical for wildfire management and ecological assessment. Optical satellite sensors, such as Sentinel-2 MultiSpectral Instrument (MSI), provide well-suited data for this task but are often obstructed by clouds and smoke during and immediately after fire events. Synthetic aperture radar (SAR) sensors offer all-weather imaging capabilities, but SAR data are challenging to interpret due to speckle noise and complex scattering mechanisms. This article introduces Swin-U-DiT, a conditional diffusion model designed to bridge this gap by translating SAR data into optical-like imagery. Our approach combines a U-Net backbone with efficient Transformer-based processing to generate high-fidelity optical equivalents from multitemporal Sentinel-1 SAR and prefire Sentinel-2 MSI data. It is more efficient than denoising diffusion probabilistic model architectures and performs better than GAN-based methods in the downstream task. We train and evaluate the model on a dataset comprising 2053 wildfire events across Canada during 2017–2022. The generated images enable downstream burned area mapping that achieves an F1-score of 0.8037. Given the rapidly increasing availability of SAR data, including RADARSAT constellation mission, NISAR, ICEYE, and other commercial constellations, this work offers strong promise for near-real-time burned area mapping in cloud-prone regions worldwide.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.