Generative Modeling of Post-Fire Recovery In Boreal Peatlands

Generative Modeling of Post-Fire Recovery In Boreal Peatlands Peatlands cover roughly 3% of global land area but store about 30% of global soil carbon (Blodau, 2002). Fire increasingly releases this carbon back to the atmosphere. The 2023 Canadian fire season burned more than seven times the four-decade average annual area. Recovery rate after fire varies with soil moisture, precipitation, pre-burn biomass, burn severity and local microclimate. Understanding it is important for decision making and resource allocation, in particular for conservation triaging, hunting and trapping, planning for Indigenous groups and watershed runoff. Physical process models address this, ie. CARDAMOM, data fusion framework for terrestrial carbon dynamics (Bloom et al., 2016). A generative model producing a spatial prediction together with its own uncertainty quantification would complement such process-based modelling. Research question: Given an image of a burn scar before the fire, during the year it burned, and one year after, together with soil, climate and geographic covariates can a conditional diffusion model learn to predict the two-year post-fire vegetation state? The data includesSentinel-2 surface reflectance at 10 m (RGB and NIR), ERA5 monthly meteorology over the April–October growing season, SoilGrids static soil properties, and a digital elevation model. A DDPM with a denoising backbone is a three-scale U-Net encoder–decoder (Ronneberger et al., 2015) with 128 base channels, trained with an ε-prediction objective (Ho et al., 2020) under a cosine noise schedule, p2 loss weighting (γ = 0.5) and EMA averaging (decay 0.999). Conditioning imagery is concatenated channel-wise with the noisy target at the network input, while the timestep embedding and the scalar climate and soil covariates are injected into every residual block via FiLM (Perez et al., 2018). On the Sentinel-2 sprint dataset the model reaches an evaluation MSE of 0.748, comparable to a memorization copy-early persistence baseline of 1.286. In the visualization below, the first three panels are the conditioning inputs and the rightmost panel is the model prediction; structural features of the scene are recovered in the generated output. References Blodau, C. (2002). Carbon cycling in peatlands: A review of processes and controls. Environmental Reviews, 10(2), 111-134. Bloom, A. A., Exbrayat, J.-F., van der Velde, I. R., Feng, L., & Williams, M. (2016). The decadal state of the terrestrial carbon cycle. PNAS, 113(5), 1285-1290. Byrne, B., et al. (2024). Carbon emissions from the 2023 Canadian wildfires. Nature, 633, 835-839. Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. NeurIPS. Perez, E., Strub, F., de Vries, H., Dumoulin, V., & Courville, A. (2018). FiLM: Visual reasoning with a general conditioning layer. AAAI. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. MICCAI.