Daily Spread Prediction of European Wildfires based on Historical Burned Area Time Series from Earth Observation Data using a Spatio-temporal Graph Neural Network

Kurzfassung

This study aims to overcome the existing accuracy and transferability limitations in operational wildfire spread models by developing a data-driven Spatio-Temporal Graph Neural Network for predicting the next day´s wildfire spread. The model was trained and tested using historic burned area time series and additional environmental variables in a regional setting in Portugal and a continental setting in the Mediterranean. Prediction accuracies were low with weighted macro-mean IoU values of 0.37 (Portugal) and 0.36 (Mediterranean), likely related to inadequate reference data quality. Spatial and temporal dimensions of wildfires significantly influenced the model performance. Nevertheless, an increased capacity in terms of model transferability was observed for the Mediterranean model.