Toward Spatio‐Temporally Consistent Multi‐Site Fire Danger Downscaling With Explainable Deep Learning

This study introduces a novel Convolutional Long Short‐Term Memory neural networks (ConvLSTM)‐based multi‐site downscaling approach for fire danger prediction, that leverages the properties of Long‐Short Term Memory (LSTM) Recursive Neural Networks and Convolutional Neural Networks (CNNs) by learning daily Multivariate‐Gaussian distributions conditioned on large‐scale atmospheric predictors. The ConvLSTM‐Multivariate‐Gaussian (MG) model enhances the predictive accuracy, spatial coherence, and temporal alignment of the downscaled Fire Weather Index (FWI). We compared its performance with Generalized Linear Models and a CNN‐based benchmark across multiple locations in Spain, focusing on extreme FWI events. Our findings show that ConvLSTM‐MG outperforms in predictive accuracy and distributional consistency, effectively capturing spatial and temporal variability. It reduces correlation length bias by over 50% and mutual information error in 90th percentile of FWI by over 80%, demonstrating robustness in representing spatial correlations under extreme conditions. The model's temporal performance aligns closely with observed data as measured by the autocorrelation function, making it a promising tool for multi‐site downscaling. Additionally, the use of eXplainable Artificial Intelligence techniques enhances model interpretability, providing insights into influential variables. Unlike other deep learning models, ConvLSTM‐MG prioritizes simplicity and ease of training, making it accessible and practical for regional weather station networks. This approach offers significant improvements in fire danger prediction, crucial for climate impact assessment and fire prevention.