Short‐Term Forecasting of the Human‐Caused Wildland Fire Occurrence Using a Measurement Error Corrected Random Forest Approach

A key component for wildfire risk management is predicting when and where fires might happen. We build short‐term (1–4 days ahead) fire occurrence prediction models by adapting the random forest algorithm using an errors in variables approach to account for the inherent forecasting error in short‐term forecasts of fuel moisture. Error distributions are estimated using local‐likelihood since there were found to depend on the forecasted value due to fuel moisture having a finite range. Our proposed method simulates data from the conditional measurement error distribution and subtract it from the observed variable to create the new error‐free variable that is used as the predictor in the random forest model. By doing this repeatedly and aggregating the results from each of these models, we obtain error‐free predictions. To assess the model performances, we conduct leave‐one‐out cross validation using several model performance metrics on our historical data set, and compared across different models. Our modified random forest model, using a errors in variable approach, outperforms the naive models which ignore measurement errors and a simple regression calibration approach. Furthermore, simulations reveal that our modified random forest approach more accurately estimates the underlying cutoff (i.e., split location) whereas the naive approach exhibits bias.