Improving extreme Fire Radiative Power forecasting with Gaussian loss-based transformer and LSTM models

Fire Radiative Power (FRP) forecasting is essential for biomass burning intensity, smoke emissions, air quality, and climate applications, yet remains difficult because satellite FRP is highly skewed and heavy tailed. Conventional deep learning often underestimates high‑intensity fires when abundant low FRP samples dominate training. We evaluate a Gaussian weighted Huber objective with Transformer and Long Short Term Memory (LSTM) models for global FRP forecasting using MODIS informed products and environmental predictors from 2017 to 2020 across California, the Congo Basin, East African savanna, and Mato Grosso, Brazil. For the Transformer, Gaussian loss yields lower point-estimate RMSE and reduces underestimation in selected FRP regimes, whereas LSTM responses are weaker and more region-dependent. The same correction can introduce localized overestimation, revealing a trade off between bias correction and prediction stability. Overall, a Gaussian weighted training objective is a practical way to reduce skewness-induced underprediction in satellite‑based global FRP forecasting, with the clearest gains when paired with a Transformer.