Exploring forest fire susceptibility and management strategies in Western Himalaya: Integrating ensemble machine learning and explainable AI for accurate prediction and comprehensive analysis

Forest fires pose a significant threat to ecosystems and socio-economic activities, necessitating the development of accurate predictive models for effective management and mitigation. In this study, we present a novel machine learning approach combined with Explainable Artificial Intelligence (XAI) techniques to predict forest fire susceptibility in Nainital district. Our innovative methodology integrates several robust models - AdaBoost, Gradient Boosting Machine (GBM), XGBoost and Random Forest - with a Deep Neural Network (DNN) as a meta-model in a stacking framework. This approach not only utilises the individual strengths of these models, but also improves the overall prediction performance and reliability. By using XAI techniques, in particular SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), we improve the interpretability of the models and provide insights into the decisionmaking processes. Our results show the effectiveness of the ensemble model in categorising forest fire susceptibility into different zones: very low, low, moderate, high and very high. In particular, the stacking ensemble and XGBoost models identified extensive areas of high and very high susceptibility, with precision, recall and F1 values underpinning their effectiveness. These models achieved ROC AUC values above 0.90, with XGBoost performing exceptionally well with an AUC of 0.94. The precision, recall and F1 values for these models are remarkably high. The inclusion of confidence intervals for the most important metrics in all models emphasises their robustness and reliability and supports their practical use in forest fire management. Through SHAP summary plots, we analyze the global variable importance, revealing annual rainfall and Evapotranspiration (ET) as key factors influencing susceptibility. Local analysis using LIME consistently highlights the importance of annual rainfall, ET, and distance from roads across all models. This study fills an important research gap by providing a comprehensive and interpretable modelling approach that improves our ability to predict and effectively manage forest fire risk and is consistent with environmental protection and sustainable development goals.