A Multimodal Ensemble Deep Learning Model for Wildfire Prediction in Greece Using Satellite Imagery and Multi-Source Remote Sensing Data

Wildfire events pose significant threats to global ecosystems, with Greece experiencing substantial economic losses exceeding EUR 1.7 billion in 2023 alone, generating immediate financial burdens while contributing to atmospheric carbon dioxide emissions and accelerating climate change effects. This study presents a group of classification models for Greece wildfires utilizing historical datasets spanning 2017 to 2021, incorporating satellite-derived remote sensing data, topographical characteristics, and meteorological observations through a multimodal methodology that integrates satellite imagery processing with traditional numerical data analysis techniques. The framework encompasses multiple deep learning architectures, specifically implementing four standalone models comprising two convolutional neural networks optimized for spatial image processing and long short-term memory networks designed for temporal pattern recognition, extending classification approaches by incorporating visual satellite data alongside established numerical datasets to enable the system to leverage both spatial visual patterns and temporal numerical trends. The implementation employs an ensemble methodology that combines individual model classifications through systematic voting mechanisms, harnessing the complementary strengths of each architectural approach to deliver enhanced predictive capabilities and demonstrate the substantial benefits achieved through multimodal data integration for comprehensive wildfire risk assessment applications.