SeasFire: Earth System Deep learning for Seasonal Fire Forecasting in Europe 

The SeasFire project, funded by the ESA, is taking an innovative approach to predicting seasonal wildfire patterns in Europe. SeasFire uses cutting-edge technology, such as modern deep learning models and Earth Observation data, to explore the spatio-temporal connections between atmospheric conditions and fire regimes to gain valuable insights into predicting potential wildfires. The project aligns with the ESA’s mission to develop innovative applications of Earth Observation data to address important societal and environmental challenges.

Achievements

A. SeasFire Cube: a public, global, analysis-ready, cloud-friendly dataset of burned areas, fire drivers and ocean-climate indices. We propose it as a test bed for wildfire-related models at sub-seasonal to seasonal scales. SeasFire Cube, includes:

• 30 global variables related to fire drivers, burned areas and fire emissions.
• Global extent from 2001 to 2021 at 8-day x 0.25° x 0.25° resolution.

B. Data Analytics and Causality: Explored seasonal and temporal patterns in fire occurrence data by analyzing the historical trends and patterns in the data on different temporal scales (monthly, annual) and areas (biomes,  GFED areas). Conducted also causal analysis in timeseries of weather variables, ocean climate indices and burnt areas in Europe’s biomes, to identify causal links between those variables at different time lags, through lagged independence tests.

C. Burned Area Pattern Forecasting based on Machine Learning shows good predictive skill several weeks and even months in advance. The machine learning models we developed are:

(a) A segmentation-based U-Net takes as input snapshots of the fire drivers and is trained to predict the burned area patterns in the future. It demonstrates a higher predictive skill than the statistical baseline of the average seasonal cycle and motivates the use of Deep Learning for the problem.

(b) Graph image-based models, can handle long-range interactions and provide explanations through the learned edge connections. Those connections could help identify similar fire regimes.

(c) A Graph Neural Network with temporal attention, can leverage local, mid-range and long range spatial connections profiting from the versatility of graph-based modeling.

(d) Transformer-based architecture, is the first model that combines information from local fire drivers and oceanic indices that are relevant in larger temporal scales. Importantly, it is shown that such models can enhance predictive capabilities, as they achieve the best performance.

D. A. prototype application has been developed based on the U-Net model and ERA5 data, producing real-time predictions of the burned area patterns for the next 6 weeks in Europe. Though the application cannot serve yet as an operational tool, it demonstrates the road from research to production and guides future steps.

Additional contributions