Forecasting wildfire hazard across northwestern south America

Fire hazard is often mapped as a static conditional probability of fire characteristics’ occurrence. We developed a dynamic product for operational risk management to forecast the probability of occurrence of fire radiative power in the locally possible near-maximum fire intensity range. We applied standard machine learning techniques to remotely sensed data. We used a block maxima approach to sample the most extreme fire radiative power (FRP) MODIS retrievals in free-burning fuels for each fire season between 2001 and 2020 and associated weather, fuel, and topography features in northwestern south America. We used the random forest algorithm for both classification and regression, implementing the backward stepwise repression procedure. We solved the classification problem predicting the probability of occurrence of near-maximum wildfire intensity with 75% recall out-of-sample in ten annual test sets running time series cross validation, and 77% recall and 85% ROC-AUC out-of-sample in a twenty-fold cross-validation to gauge a realistic expectation of model performance in production. We solved the regression problem predicting FRP with 86% r2 in-sample, but out-of-sample performance was unsatisfactory. Our model predicts well fatal and near-fatal incidents reported in Peru and Colombia out-of-sample in mountainous areas and unimodal fire regimes, the signal decays in bimodal fire regimes.