Causal Graph Neural Networks for robust wildfire forecasting across geographic shifts

Machine learning has become a powerful tool for modeling the relationships between environmental factors and fire events. However, beyond the predictive performance, we argue that critical decision-making requires an understanding of fire mechanisms to improve reliability. Causality offers a promising framework for explicitly analyzing the interdependencies among factors; however, its integration into deep learning and further application in disaster management remain largely underexplored. To map the relationship between historical inputs and resulting burned areas, we proposed a causally informed deep learning approach using graph models. The graph representation is constructed through a learnable approach supervised by causal knowledge. A graph pooling layer, informed by backdoor adjustment criteria, mitigates the potential confounding effects of hidden variables on the target variable. Our experiments demonstrate that our model is more robust, reducing the AUROC standard deviation at longer forecasting horizons by 64% and improving performance under geographical distribution shifts by 2 points compared to the baseline. Compared with fully connected and correlation-based graphs, the causally informed graph proved to be more resilient to input perturbations. Additionally, our model revealed the lagged effect of Oceanic Climate Index variables on local fire events and the critical role of short-term local precipitation—indicating that Mediterranean fires are mostly drought-driven. • Need for models that are right for the right reasons in wildfire forecasting. • Propose a graph network that integrates causality to predict wildfires up to eight weeks ahead. • The new method is more robust under geographic shifts and perturbations. • Explainable AI reveals the lagged effect of teleconnections and the dominant role of local precipitation in driving Mediterranean wildfires.