Wildfire spread forecasting with Deep Learning

TL;DR

This work tackles the problem of forecasting the final burned area from wildfire ignition using a deep learning framework on a large spatio-temporal Mediterranean dataset. It systematically evaluates temporal context by comparing ignition-day baselines with models trained on extended windows up to four days pre- and five days post-ignition, finding that incorporating post-ignition data yields notable improvements in $F1$/$IoU$ scores. The authors provide a publicly available dataset of ~9,568 fire events and implement multiple architectures, including 2D/3D U-Nets and a Vision Transformer, identifying the 10-day $3$D U-Net as the best performer with a Dice score of $53.6\%$ and IoU of $36.6\%$ on the test set, while discussing limitations and future directions. The study demonstrates the practical potential of data-driven wildfire modeling for risk management and response, and paves the way for further improvements via higher-resolution data, additional environmental variables, and advanced temporal models. $F1$ and $IoU$ gains, together with a public release of data and code, enhance reproducibility and cross-domain applicability of data-driven wildfire spread forecasting.

Abstract

Accurate prediction of wildfire spread is crucial for effective risk management, emergency response, and strategic resource allocation. In this study, we present a deep learning (DL)-based framework for forecasting the final extent of burned areas, using data available at the time of ignition. We leverage a spatio-temporal dataset that covers the Mediterranean region from 2006 to 2022, incorporating remote sensing data, meteorological observations, vegetation maps, land cover classifications, anthropogenic factors, topography data, and thermal anomalies. To evaluate the influence of temporal context, we conduct an ablation study examining how the inclusion of pre- and post-ignition data affects model performance, benchmarking the temporal-aware DL models against a baseline trained exclusively on ignition-day inputs. Our results indicate that multi-day observational data substantially improve predictive accuracy. Particularly, the best-performing model, incorporating a temporal window of four days before to five days after ignition, improves both the F1 score and the Intersection over Union by almost 5% in comparison to the baseline on the test dataset. We publicly release our dataset and models to enhance research into data-driven approaches for wildfire modeling and response.