Forecasting Subseasonal Fire Radiative Power with Novel TabNet-GPR and GNN-LSTM

In recent years, wildfires have become an escalating threat, with 8.9 million acres burned in the U.S. in 2024 alone. To combat this, two pipelines were developed and compared for predicting subseasonal Fire Radiative Power (FRP): a spatio-temporal Graph Neural Network (GNN) + Long Short-Term Memory (LSTM) model and a data-driven TabNet Neural Network + Gaussian Process Regressor (GPR). FRP quantifies fire heat intensity and serves as a reliable indicator for estimating wildfire strength. The data was assembled from 2-day lags of satellite, weather, and historical fire data, including features such as wind speed, temperature, and pixel neighbor FRP values. Bayesian Optimization was applied to improve both pipelines by tuning hyperparameters. GNN-LSTM modeled spatio-temporal patterns via graph embeddings and recursive rollout, where nodes represented locations with Latitude and Longitude. TabNet-GPR used residual learning to refine predictions. Models were evaluated with MAE, RMSE, and R² against three baselines: Persistence, Climatology, and Linear Regression. In same-day predictions, GNN-LSTM reduced MAE by 9.2% compared to the best baseline and achieved the highest R² (0.3169), above TabNet-GPR and Linear Regression, and MAE was below all baselines. During forecasting, TabNet-GPR showed spikes and negative values, showing the pipeline’s instability for small datasets with high variance due to lack of memory. GNN-LSTM had few spikes in forecasting due to spatio-temporal awareness, as shown by higher R2. For subseasonal forecasting, GNN‑LSTM outperformed TabNet and TabNet‑GPR. Temporal modeling with LSTM captured sequential patterns, while GNN leveraged spatial information to improve predictions. TabNet alone achieved high error and low R2, though GPR degraded performance significantly. Though GPR improves short-term predictions, its lack of memory leads to amplified noise in recursive rollout, leading to error drift and spikes. GNN-LSTM’s errors stemmed from node sparsity and GraphSAGE’s static graph construction, while TabNet-GPR’s high error rates were from micro-sampling necessitated by GPR. Results were limited by computational constraints. Next steps include dataset expansion to improve R², quantile regression for uncertainty, and replacing GPR with LightGBM to boost efficiency.