Daily Data-Driven Emulation of the Fire Weather Index: Deep Learning Solutions for Wildfire Risk Prediction

Wildfires are an intensifying global challenge, driven by climate change, which increases their frequency, severity, and spatial extent. Accurate wildfire risk assessment and forecasting are essential for effective mitigation, resource allocation, and long-term planning. The Canadian Fire Weather Index (FWI) is a widely used fire danger rating system that integrates four primary daily meteorological variables—24-hour accumulated precipitation, wind speed, relative humidity, and temperature—into six components representing fuel moisture, ignition probability, and fire spread potential. Its temporal "memory" feature, which tracks moisture changes over time, makes it particularly valuable for capturing wildfire dynamics.However, the FWI reliance on specific daily input data at noon poses challenges for its application in regions or scenarios lacking such precise temporal measurements. To address this limitation, FWI proxies computed using daily mean data offer a practical alternative. Yet, these proxies often lack the fidelity required to fully replicate the FWI values.This study focuses on enhancing the emulation of the original FWI using daily mean data and other proxy variables by leveraging advanced deep learning techniques. We explore a spectrum of architectures, ranging from conventional machine learning models to state-of-the-art approaches like convolutional neural networks (CNNs) and Convolutional Long Short-Term Memory (ConvLSTM) networks. These models are tailored to capture the spatial and temporal complexities of wildfire behavior while maintaining robustness in the face of variable data availability.Our research centers on the Iberian Peninsula, a Mediterranean region highly vulnerable to extreme wildfire events. By utilizing high-resolution, geo-referenced datasets, we validate the ability of these models to emulate the original FWI with high accuracy. To enhance model interpretability, we integrate eXplainable Artificial Intelligence (XAI) techniques, providing actionable insights into the decision-making processes and addressing concerns about the "black box" nature of deep learning.This work demonstrates how daily data, combined with cutting-edge deep learning methods, can effectively emulate the FWI, offering a scalable and reliable solution for wildfire risk prediction in regions where traditional inputs are unavailable. The proposed models bridge the gap between limited data availability and the growing need for precise fire danger indices, enabling improved assessment and planning for wildfire-prone regions.By advancing the science of wildfire modeling through daily data-driven approaches, this study contributes to a deeper understanding of spatial and temporal wildfire dynamics. It highlights the potential of integrating geoscience, climatology, and artificial intelligence to develop practical tools for wildfire risk mitigation, resilience, and decision-making in a rapidly changing climate. Acknowledgments: This research work is part of R+D+i project CORDyS (PID2020-116595RB-I00) with funding from the Spanish Ministry of Science, Innovation and Universities MCIN/AEI/10.13039/501100011033. O.M. has received the research grant PRE2021-100292 funded by MCIN/AEI /10.13039/501100011033.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.