Where there’s smoke, there’s data: SeasFire’s mission to predict wildfires in Europe

SeasFire, an ESA Φ-lab-supported initiative, uses cutting-edge deep learning algorithms to explore the spatio-temporal connections between Earth system variables and fire regimes, gaining valuable insights into predicting potential wildfires. This project is aligned with ESA’s mission to develop innovative applications of Earth Observation (EO) data that address important societal and environmental challenges. 

Planet Earth undergoes several complex physical processes that occur at variable spatial and temporal scales. Wildfires are notable examples of such processes, since they do not behave the same way in different areas and years. As major hazards, wildfires are deeply influenced by a combination of multiple human and natural factors, such as temperature, soil moisture, relative humidity, wind speed, vegetation – commonly referred to as ‘fire drivers’.

Wildfires disrupt natural ecosystems and cause the loss of lives, properties and infrastructure. Due to climate change, an increase in the number of fires in Europe and around the world is expected, with major wildfire events extending to evergreen forests and boreal regions. Therefore, it is important to improve our capabilities to anticipate fire danger and understand its driving mechanisms at a global scale.

SeasFire – Earth System Deep Learning for Seasonal Fire Forecasting in Europe – emerges as an innovative solution for fire forecasting. Supported by ESA Φ-lab and implemented by the National Observatory of Athens, the National Technical University of Athens, the Harokopio University of Athens and the Max Planck Institute for Biogeochemistry, SeasFire proposes to explore and capture the potential spatio-temporal asynchronous links between pre-occurring and non-overlapping fire-driving forces in the Earth system and European fire regimes to predict seasonal burned area extent in Europe.

How has this been accomplished? SeasFire has made use of two major advancements of our time, namely the availability of a huge amount of satellite data with good spatio-temporal resolution, and Deep Learning (DL) techniques that have proven capable of capturing the spatio-temporal interactions of Earth system variables, treating Earth as an interconnected system.

Some of the DL models developed in this project include an encoder-decoder architecture that takes as input snapshots of fire drivers and is trained to predict burned area patterns in the future; FireCastNet is a Graph Neural Network that can leverage local, mid-range and long-range spatial connections; and TeleViT is a transformer-based architecture, which combines information from local fire drivers and teleconnections to improve long-term forecasting.

Current CO₂ estimates rely on factors such as burned areas, vegetation carbon stocks, and combustion completeness. While Machine Learning algorithms have been developed to forecast burned area patterns, prior estimates from vegetation carbon stocks and combustion completeness are model-based and assume fixed emission factors that may not accurately capture larger changes in carbon stocks over longer periods.

SeasFire overcomes these limitations by integrating Earth Observation (EO) data with climate datasets in a hybrid modelling framework. This approach combines process-oriented modelling with observation-based learning, enabling more accurate model parameterisation and reducing biases from model initialisation. By comparing existing carbon models with data from the Copernicus Atmosphere Monitoring Service (CAMS), SeasFire enhances the reliability and precision of carbon cycle predictions.    

The SeasFire DataCube, a public global analysis-ready and cloud-friendly dataset for seasonal fire forecasting, between the years 2001-2021 and at a spatio-temporal resolution of 0.25° x 0.25° x 0.25° x 8 days, includes a combination of variables describing seasonal fire drivers, namely climate, vegetation, oceanic indices, human factors, land cover and the burned areas. In the future, this datacube can also be exploited as a template for modelling different natural hazards like floods, heatwaves and droughts.

From this initiative resulted also an interactive toolkit that allows the visualisation of EO data and model outputs stored in a Zarr file, accessed via the SeasFire GitHub organisation. In this repository, it is possible to find all trained models regarding seasonal wildfire forecasting and modelling of CO₂ emissions.

“This project explores a key gap by attempting to capture the interannual variability of seasonal, high-impact wildfire events — an area where traditional numerical models often fall short. By leveraging Deep Learning and teleconnections which dictate climate dynamics, we explore novel machine learning methods that treat the Earth as a system to enhance long term forecast capabilities. Moving forward, we aim to refine models for more precise regional predictions, establish clear benchmarks for comparing our models with existing and future approaches, and strengthen collaborations to validate and operationalise our methodology”, says Ilektra Karasante, SeasFire Project Manager.

“Φ-lab’s support has been essential from the project’s conception to its implementation. Beyond the initial trust in funding this high-risk/high-gain research, Φ-lab has further accelerated our developments by providing access to computational resources through the Network of Resources (NoR), fostered collaborations with leading scientists in the field, and enhanced the project’s visibility through ESA and ECMWF workshops, strengthening our research network and helping integrate our work into the broader scientific community. Through all the above Φ-lab-supported activities, SeasFire outcomes reached a much broader audience.”

Φ-lab hosted a workshop about “Innovations in Data-Driven Seasonal Fire Forecasting: From Models to Visuals” on 7 February 2025. Patrick Ebel, Internal Research Fellow at Φ-lab and SeasFire Technical Officer comments: “It was my pleasure being ESA’s Technical Officer for this activity and its enthusiastic consortium. SeasFire shows how recent advances in Deep Learning and weather forecasting can be harnessed to model wildfires and their impact to tackle one of the greatest challenges Europe will be facing in the coming decades under a changing climate. I appreciated the public’s scientific interest in SeasFire and look forward to future advances building on the achievements of the activity.”

To know more: ESA Φ-lab, SeasFire

Photo courtesy of Unsplash/Mike Newbry