An AI-driven approach to enhancing wildfire representation and climate feedbacks in the UVic-ESCM v2.10

Wildfire trends vary by region and are influenced by climate, vegetation, and human activity. Regional trends over the past 20 years have varied, though overall have driven a 60% increase in global wildfire carbon emissions, primarily from carbon-dense boreal forests. In addition to releasing carbon, wildfires alter surface albedo, aerosols, and vegetation dynamics, producing complex climate feedbacks. Representing patterns and quantities of burned areas across the globe is thus crucial to accurately predict future climate, but is difficult due to the nonlinear and spatially heterogeneous nature of wildfire drivers. In this work, we develop an artificial intelligence (AI)-based model to predict patterns and quantities of burned areas across the globe,  with the goal of integrating it within the University of Victoria Earth System Climate Model (UVic-ESCM v2.10). Our model consists of a deep neural network trained with a new custom, spectral-based loss function (DNN-FFTLoss). We compare it with deep neural networks trained with a mean-square error loss function (DNN-MSE) and random forests (RF), using a consistent set of climate and vegetation predictors from the UVic-ESCM v2.10.Training is performed using climate and vegetation predictors from CMIP6 simulations (1850–2100, including multiple Shared Socioeconomic Pathway (SSP) scenarios) alongside satellite-based Global Fire Emissions Database (GFED) 4 burned area observations (2001–2015). Transfer learning is then performed using the GFED4 dataset to impose observational constraints, reduce biases, and improve burned area predictions and the representation of fire-climate interactions. A comparison with the independent test year (2014) reveals that the DNN-FFTloss more accurately reproduces the spatial and seasonal variability of global burned area than the DNN-MSE and RF. However, the DNN-FFTloss still exhibits regional biases, overestimating burned area in Northern and Southern Africa and Australia and underestimating it in Europe. Nevertheless, these discrepancies are reduced relative to the other architectures. Additionally, the global cumulative density function of burned area is best captured by the DNN-FFTloss, indicating improved representation of both high- and low-burn regions. All model configurations show reduced skill temporally during the spring transition (e.g., March-April), when global Pearson correlations drop to 0.3 for the DNN-MSE model and 0.6 for the DNN-FFTloss model. Overall, the DNN-FFTloss better represents the global behaviour of wildfire burned area and will provide new insights into how climate change alters wildfire regimes and their impact on terrestrial carbon storage.