Evaluating AlphaEarth and TESSERA Geospatial Embeddings for Machine Learning-Based Burned Area Mapping in Portugal

Abstract

Accurate burned area mapping is essential for wildfire damage assessment, post-fire management and fire regime analysis in fire-prone Mediterranean landscapes. Classical satellite-based approaches typically rely on spectral indices derived from paired pre- and post-fire optical imagery. Recently released geospatial embedding datasets, such as AlphaEarth and TESSERA, provide compact, ready-to-use representations of Earth observation data that may support burned area classification with lower event-specific processing requirements. However, their performance relative to established spectral workflows remains insufficiently assessed.
This study compares AlphaEarth and TESSERA against a Sentinel-2 spectral baseline for machine learning-based burned area classification across twenty wildfire events from the official 2024 ICNF burned area catalogue for Portugal. Burned and unburned samples were generated from official fire perimeters using an adaptive, area-proportional sampling strategy. The Sentinel-2 baseline was constructed from pre- and post-fire temporal windows combining original bands, spectral indices and difference features. Random Forest, Support Vector Machine and Gradient Boosting classifiers were evaluated using five-fold GroupKFold cross-validation grouped by fire event, ensuring that no wildfire contributed samples to both training and test sets.
TESSERA consistently outperformed AlphaEarth among the embedding-based representations. With Support Vector Machine, TESSERA achieved a mean F1-score of 0.860±0.046, compared with 0.812±0.083 for AlphaEarth. However, neither embedding matched the Sentinel-2 spectral baseline, which achieved the strongest overall performance with a mean F1-score of 0.895±0.027. Both embeddings showed reduced performance when features were extracted from the year following the fire rather than from the fire year itself, with F1-scores decreasing from 0.812 to 0.722 for AlphaEarth and from 0.860 to 0.796 for TESSERA.
Spatial prediction maps generated with the best-performing TESSERA/SVM model reproduced coherent burned area patterns for two large fire events, although the quantitative evaluation remained sample-based rather than fully cartographic.
These results indicate that geospatial embeddings are promising scalable complements to spectral workflows, rather than direct replacements. TESSERA, in particular, approaches spectral-baseline performance while avoiding event-specific image differencing, making it useful when a single annual feature layer must support the classification of multiple fire events. Nevertheless, when accurate fire dates and suitable Sentinel-2 imagery are available, a purpose-built spectral baseline remains the stronger option.