Detection of wildfire burn scars in the UK using geospatial foundation models

Wildfires are posing ecological and social challenges in the United Kingdom [1]. Therefore, it is important to better understand this phenomenon and locate the wildfires to identify their contributing factors. However, the analysis of UK wildfires is mainly based on the European Forest Fire Information System (EFFIS) [2], which uses MODIS and VIIRS imagery [1]. Due to the moderate pixel resolution of these satellites, only fires greater than 30 hectares are reliably recorded. This is a limitation for the study of UK wildfires as 99% are smaller than 30 hectares [1]. Thanks to higher resolution sensors such as Sentinel-2 MSI and Landsat 8 OLI, it has now become possible to map smaller wildfires [3]. However, we have found no evidence that these sensors were used to locate smalla wildfires in the UK.  

Recently, geospatial foundation models have made significant improvements in the processing of satellite imagery. More specifically, Prithvi-EO-2.0 [4] and TerraMind [5] have outperformed typical machine learning models in many applications.  

With this work, we studied how the Prithvi-EO-2.0 and TerraMind geospatial foundation models generalized and performed for the detection of wildfires in the UK. First, we created a dataset made of Harmonized Landsat and Sentinel images matched to UK EFFIS wildfire polygons (1409 large wildfire polygons covering the UK), as well as wildfire polygons from the UK Dorset region (typically smaller wildfire polygons obtained from 1147 wildfire intervention records of the Dorset Fire Intervention service). Then, we compared the performance of Prithvi-EO-2.0 and TerraMind over this dataset, using different fine-tuning configurations to analyze their performance and generalization capabilities. These models were also compared with typical ML and rule-based wildfire detection methods in order to confirm the relevance of our models. 

We demonstrated that the use of geospatial foundation models, once fine-tuned over UK wildfire data, allowed us to increase the detection of the wildfire from 0.58 MIoU (rule-based baseline models) and 0.73 MIoU (ML based baseline models) to 0.78 (Prithvi-EO-2.0) and 0.81 (TerraMind) MIoU. We have found that this increase in performance is especially important for the detection of smaller wildfires relevant to our study. 

This work thus provides a novel approach to detect smaller wildfires in the UK and the rest of the world using geospatial foundation models, but also highlights the necessity to train the geospatial foundation models with diverse data to improve its generalizability. 

[1] Belcher, C. M et al.: UK wildfires and their climate challenges. Expert Led report prepared for the third climate change risk assessment (2021). 
[2] San-Miguel-Ayanz, J. et al.: Towards a coherent forest fire information system in Europe: the European Forest Fire Information System (EFFIS) (2002). 
[3] Filipponi, F.: Exploitation of sentinel-2 time series to map burned areas at the national level: A case study on the 2017 italy wildfires. Remote Sensing, 11(6), 622 (2019). 
[4] Jakubik, J. et al.: Foundation Models for Generalist Geospatial Artificial Intelligence. Preprint Available on arxiv:2310.18660 (2023). 
[5] Jakubik, J. et al.: TerraMind: Large-Scale Generative Multimodality for Earth Observation. IEEE/CVF International Conference on Computer Vision (ICCV) (2025).