Generative Modeling of Post-Fire Disturbance Vegetation Recovery in Boreal Peatlands

Canadian boreal peatlands hold a globally significant store of soil orcarbon, and wildfire activity across the region has surged over the past several years as the climate warms. Given satellite imagery of a burn scar over three consecutive years, we predict its visual recovery with a generative diffusion model, a Denoising Diffusion Probabilistic Model (DDPM) with a U-Net backbone. Beyond assembling a multimodal dataset of environmental covariates and remote-sensing imagery, we find that the model is most sensitive to the early-recovery frame and the near-infrared channel, whose removal drops SSIM sharply. This framework and dataset aim to take a step toward scenario-based recovery forecasting.