Object-based models map burned forests in Mediterranean transfer tests

Object-based machine-learning models identified burned forest areas from uni-temporal Sentinel-2 imagery and retained overall accuracies above 83% when transferred to selected, ecologically similar Mediterranean sites in Greece and Spain. In the original study of four forest areas in Türkiye, Random Forest models achieved overall accuracies of 93.5–97.7%, while Gradient Tree Boosting models achieved 92.9–96.8%. The study, published in Natural Hazards in 2026, evaluated an approach designed to map post-fire burned forests using satellite imagery acquired at one point in time.

Forest fires pose significant threats to ecosystems and can cause irreversible damage beyond the burned areas. The workflow combined Object-Based Image Analysis with Random Forest and Gradient Tree Boosting algorithms on the Google Earth Engine platform. The researchers segmented the Sentinel-2 imagery with the Simple Non-Iterative Clustering algorithm before applying the classifiers.

Four remotely sensed indices were calculated and included as model inputs: the Mid-Infrared Burn Index, Normalized Burn Ratio 2, Burn Area Index and Normalized Difference Vegetation Index. The study’s scope included four Mediterranean forest areas in Türkiye, followed by transferability tests at selected sites in Greece and Spain. Both machine-learning approaches produced high reported overall accuracies in the initial analysis, with Random Forest reaching 93.5–97.7% and Gradient Tree Boosting reaching 92.9–96.8%.

When the models were transferred to the comparison sites in Greece and Spain, both retained overall accuracies above 83%. These results support the potential use of object-based Random Forest and Gradient Tree Boosting models for burned-forest mapping within the tested Mediterranean environments. They do not establish that the same performance would extend to locations with different climates, vegetation types, fire regimes or satellite-image conditions.

The accessible source for this report provides the study abstract rather than the complete methods and results. As a result, the number of test samples, uncertainty estimates, site-specific error patterns and any comparisons with other classification methods cannot be assessed here. The reported percentages describe overall accuracy for the evaluated models, not perfect identification of burned forest in every landscape.

Even so, the study demonstrates that an object-based Sentinel-2 workflow can combine image segmentation, four burn- and vegetation-related indices and tree-based machine learning to map burned forests in Türkiye and maintain useful reported accuracy at selected Mediterranean transfer sites.

Subject of Research: Transferability of object-based machine-learning models for burned-forest mapping with Sentinel-2 imagery in Mediterranean regions

Article Title: Post-fire burned forest area mapping with Sentinel-2 imagery: evaluating object-based machine learning models and transferability

Article References: Bulut, F. S., Turker, M., & Ateşoğlu, A. (2026). Post-fire burned forest area mapping with Sentinel-2 imagery: evaluating object-based machine learning models and transferability. Natural Hazards, 122(22), Article 679. https://doi.org/10.1007/s11069-026-08470-6

Image Credits: AI Generated

DOI: 10.1007/s11069-026-08470-6

Keywords: Sentinel-2, burned forest mapping, machine learning, Random Forest, Gradient Tree Boosting, Object-Based Image Analysis, Google Earth Engine, Mediterranean forests, remote sensing, post-fire assessment, Post-fire, burned