Real-time Fire Evolution Forecasting

Abstract What an active wildfire does in the hours after detection determines suppression demands, risks to communities and emissions, yet wide-scale operational forecasting at this timescale remains limited. Mechanistic spread models need detailed inputs and substantial computing resources, while fire danger systems work at daily or longer timescales and do not describe individual fires. We present Sparky-Evolve, a probabilistic machine-learning model that forecasts the persistence and intensity of active fires within ~ 1 km grid cells at 15-minute resolution up to 6 h ahead. It is trained on 1.25 million fire events observed by the Flexible Combined Imager aboard Meteosat Third Generation (MTG), using fire radiative power (FRP) with weather, vegetation and human predictors. An XGBoost classifier estimates whether FRP remains detectable, a Temporal Fusion Transformer estimates FRP quantiles conditional on persistence, and the two are coupled in an autoregressive ensemble. On held-out events, the ensemble reduced conditional FRP error by 30% relative to persistence, with well-calibrated uncertainty up to 2 h, and estimated when detectable activity would end with a mean absolute error of 39 min. The most intense fires were underestimated. For three 2026 fires in Angola, Spain and Ukraine, the ensemble captured the post-peak decline in FRP that persistence misses, reducing errors by 68–87%, and kept survival probabilities well above climatology. Forecasts can be issued about 1 h after first detection and are inexpensive enough to run for every active fire in the MTG domain. Sparky-Evolve can therefore help identify which newly detected fires are likely to persist, supporting the prioritisation of suppression resources and near-real-time air-quality forecasting.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11553556/v1
Primary Topic
Fire effects on ecosystems
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Real-time Fire Evolution Forecasting

Joe McNorton, Francesca Giuseppe
2 links
Research Square
Fire effects on ecosystems
preprint

Real-time Fire Evolution Forecasting

Joe McNorton, Francesca Giuseppe
preprint en

Abstract

Abstract What an active wildfire does in the hours after detection determines suppression demands, risks to communities and emissions, yet wide-scale operational forecasting at this timescale remains limited. Mechanistic spread models need detailed inputs and substantial computing resources, while fire danger systems work at daily or longer timescales and do not describe individual fires. We present Sparky-Evolve, a probabilistic machine-learning model that forecasts the persistence and intensity of active fires within ~ 1 km grid cells at 15-minute resolution up to 6 h ahead. It is trained on 1.25 million fire events observed by the Flexible Combined Imager aboard Meteosat Third Generation (MTG), using fire radiative power (FRP) with weather, vegetation and human predictors. An XGBoost classifier estimates whether FRP remains detectable, a Temporal Fusion Transformer estimates FRP quantiles conditional on persistence, and the two are coupled in an autoregressive ensemble. On held-out events, the ensemble reduced conditional FRP error by 30% relative to persistence, with well-calibrated uncertainty up to 2 h, and estimated when detectable activity would end with a mean absolute error of 39 min. The most intense fires were underestimated. For three 2026 fires in Angola, Spain and Ukraine, the ensemble captured the post-peak decline in FRP that persistence misses, reducing errors by 68–87%, and kept survival probabilities well above climatology. Forecasts can be issued about 1 h after first detection and are inexpensive enough to run for every active fire in the MTG domain. Sparky-Evolve can therefore help identify which newly detected fires are likely to persist, supporting the prioritisation of suppression resources and near-real-time air-quality forecasting.

Research Square
European Centre for Medium-Range Weather Forecasts (GB)
European Commission, Joint Research Centre
Climate action
Fire effects on ecosystems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon