PyroAdapt: Adapting Wildfire Prediction under Spatial Heterogeneity and Temporal Shift
PyroAdapt adapts wildfire prediction to changing environments by combining dynamic manifold retrieval with preference-based optimization. It creates a localized, relevant dataset from historical data and then fine-tunes the model using a hybrid loss function that balances standard classification with ranking rare extreme events. This results in more robust predictions under spatial and temporal shifts.