Recent advances in explainable Machine Learning models for wildfire prediction

• Published in: Applied computing and geosciences Vol. 27; p. 100266
• Main Authors: Sengupta, Abira, Woodford, Brendon J.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2025; Elsevier
• Subjects: Explainable artificial intelligence; Forest fires; Hyper-parameter optimisation; Machine learning; SHAP values; Explainable artificial intelligence; Hyper-parameter optimisation; SHAP values; Forest fires; Machine learning
• ISSN: 2590-1974; 2590-1974
• DOI: 10.1016/j.acags.2025.100266

Climate change has caused increasingly frequent occurrences of forest fires around the world. Machine Learning (ML) and Artificial Intelligence models have emerged to predict both the onset of wildfires and evaluate the extent of damage a wildfire would cause. However, understanding what factors lead to generating models that exhibit optimal performance and providing insight into the importance of features on model outcomes is the subject of ongoing research. To help answer these questions, we propose a framework which adopts recent advances in methods for obtaining optimal models along with the application of SHAP (SHapley Additive exPlanations) values to obtain the most important features which affect the performance of wildfire prediction models. We use this framework as a classification task to predict the likelihood of wildfire occurrence based on environmental conditions, using a data set which represents instances of forest fires in Algerian, and as a regression task to predict the burned area once a wildfire has begun, using a data set from Portugal that recorded the area burned after a fire event. Insights provided by this framework allow us to assess the efficacy of specific ML models for wildfire prediction, ultimately making recommendations as to which ML models are more suited towards these challenging tasks. [Display omitted] •Machine learning model framework for wildfire forecasting and burned area estimation.•Shapley values show interactions between features that influence model predictions.•Adopting a new method to show contributions of hyper-parameters on model learning.