Search Results

This study investigates the impact of location uncertainty on the predictive performance of Bayesian Logistic Regression (BLR) for forest fire ignition prediction in Austria. Historical forest fire ignitions are used to create a dataset for training models with the capability to assess the general forest fire ignition susceptibility. Each recorded fire ignition contains a timestamp, the estimated location of the ignition and a radius defining the area within which the unknown true location of the ignition point is located. As the values of the predictive features are calculated based on the assumed location, and not the unknown true location, the training data is biased due to input uncertainties. This study is set to assess the impact of input data uncertainty on the predictive performance of the model. For this we use a data binning approach that splits the input data into groups based on their location uncertainty and use them later for training multiple BLR models. The predictive performance of the models is then compared based on their accuracy, area under the receiver operating characteristic curve (AUC) scores and brier scores. The study revealed that higher location uncertainty leads to decreased accuracy and AUC score, accompanied by an increase in the brier score, while demonstrating that the BLR model trained on a smaller high-quality dataset outperforms the model trained on the full dataset, despite its smaller size. The study’s contribution is to provide insights into the practical implications of location uncertainty on the quality of forest fire susceptibility predictions, with potential implications for forest risk management and forest fire documentation.