Artificial Intelligence for Early Wildfire Detection in Northern Morocco: A Robust and Interpretable CatBoost–Transformer Framework Using Multisource Environmental Data

Abstract

Forest fires pose a major threat to Mediterranean ecosystems and northern Morocco is particularly vulnerable due to climate pressure, rugged terrain and limited firefighting resources. This study aims to build an effective and interpretable prediction model adapted to these local conditions. Using daily meteorological data, fire danger indices and satellite observations from 2019 to 2024 a range of machine learning, deep learning and hybrid models were evaluated. The hybrid CatBoost transformer achieved the best performance with 92.5% accuracy and a ROC AUC of 97.2% .The SHAP analysis identified geographic location, surface pressure and drought indices as the most influential features . These results highlight the value of hybrid and explainable AI models for strengthening the early warning and prevention of wildfires in the Mediterranean region of Morocco.