Performance of Transformer-Based Time Series Models in Predicting Wildfire Ignition Points in Mediterranean Maquis 72 Hours in Advance: A Comparative Analysis with LSTM

Mediterranean maquis, despite its high biodiversity value, faces severe threats from increasing wildfire frequency driven by climate change. The effectiveness of early warning systems depends on the accurate prediction of fire ignition points from meteorological data. This study comparatively examines the accuracy of Transformer-based time series models (Temporal Fusion Transformer, Informer, and PatchTST) and Long Short-Term Memory (LSTM) networks in predicting wildfire ignition points in Mediterranean maquis 72 hours in advance using meteorological data. The research utilizes meteorological variables (temperature, humidity, wind speed, precipitation, Keetch-Byram Drought Index) collected from Turkey, Greece, and Italy between 2010 and 2024, along with satellite-based fire ignition records. Experimental findings demonstrate that the Temporal Fusion Transformer model achieves the highest performance with an F1-score of 84.7%, while the single-layer LSTM model remains at 71.3%. The multi-variate attention mechanisms were found to provide a distinct advantage in capturing long-range meteorological dependencies. Results indicate that Transformer architectures offer a statistically significant superiority over LSTM in Mediterranean fire early warning systems.