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

Technical Summary: Evaluating AlphaEarth and TESSERA Geospatial Embeddings for Burned Area Mapping in Portugal

Problem Statement
Accurate burned area mapping is critical for wildfire damage assessment, post-fire management, and fire regime analysis, particularly in fire-prone Mediterranean landscapes like Portugal. Traditional satellite-based approaches rely on spectral indices (e.g., dNBR) derived from paired pre- and post-fire optical imagery. While effective, these methods are operationally demanding, requiring event-specific image selection, cloud masking, temporal compositing, and differencing for every individual fire. Recent geospatial embedding datasets, such as AlphaEarth and TESSERA, offer compact, ready-to-use representations of Earth observation data that could potentially support classification with lower processing requirements. However, their performance relative to established spectral workflows and their ability to generalize across unseen fire events remain insufficiently assessed.

Methodology
The study evaluated AlphaEarth and TESSERA embeddings against a Sentinel-2 spectral baseline for machine learning-based burned area classification. The experimental framework utilized 20 wildfire events selected from the official 2024 ICNF burned area catalogue for Portugal, stratified by fire size to ensure a balanced dataset.

• Data Sampling: A binary point-based sampling design was employed. Burned samples (class 1) were generated within official fire perimeters, while unburned samples (class 0) were drawn from an adaptive external buffer scaled to the fire's perimeter area to maintain comparable spatial support. Sampling density was proportional to fire size, constrained between 100 and 5,000 points per class.
• Feature Stacks: Three feature representations were extracted for the sampled points:
  1. AlphaEarth: 64-dimensional annual geospatial embeddings (2024 and 2025 variants).
  2. TESSERA: 128-dimensional spectral-temporal embeddings derived from satellite time series (2024 and 2025 variants).
  3. Sentinel-2 Spectral Baseline: A 45-predictor stack comprising 10 original bands, 5 spectral indices (NDVI, NBR, NBR2, NDMI, BSI), and their temporal differences (pre- vs. post-fire windows) for the 2024 fire year.
• Machine Learning: Three supervised classifiers—Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting (GB)—were evaluated.
• Validation Strategy: A rigorous five-fold GroupKFold cross-validation was used, grouping samples by fire_id. This ensured that no samples from a specific wildfire event appeared in both the training and test sets, providing a conservative measure of generalization to unseen fire events.
• Analysis: Performance was assessed using F1-score, balanced accuracy, precision, recall, ROC-AUC, and a cell-based Intersection over Union (IoU). Feature importance was analyzed using RF and LIME (for the spectral baseline), and spatial prediction maps were generated for two large fire events using the best-performing model.

Key Results

• Performance Hierarchy: The Sentinel-2 spectral baseline achieved the highest overall performance, with a mean F1-score of 0.895 ± 0.027. Among the embedding-based representations, TESSERA consistently outperformed AlphaEarth. TESSERA (SVM) achieved a mean F1-score of 0.860 ± 0.046, while AlphaEarth (SVM) scored 0.812 ± 0.083.
• Classifier Performance: SVM generally yielded the best results across all feature stacks, though the gap between classifiers was smaller than the gap between feature representations.
• Temporal Stability: Fire-year (2024) embeddings significantly outperformed following-year (2025) embeddings. For TESSERA, the F1-score dropped from 0.860 to 0.796 when using 2025 data; for AlphaEarth, it dropped from 0.812 to 0.722.
• Feature Importance:
  • TESSERA: Showed low temporal stability in feature importance, with only one dimension shared between the top-15 rankings of 2024 and 2025, suggesting high sensitivity to year-specific spectral-temporal changes.
  • AlphaEarth: Demonstrated higher temporal consistency, with nine dimensions shared in the top-15 rankings.
  • Sentinel-2 Baseline: LIME analysis identified the differenced Normalized Burn Ratio (dNBR) as the most predictive variable, followed by post-fire NBR2 and near-infrared bands, confirming the physical grounding of the spectral approach.
• Spatial Mapping: The TESSERA/SVM model produced spatially coherent burned area patterns for two large fire events (8,310 ha and 299 ha). While the model reproduced broad spatial extents, some predicted areas extended beyond official perimeters, occasionally coinciding with nearby fire occurrences rather than representing pure commission errors.

Significance and Claims
The paper concludes that geospatial embeddings are promising scalable complements to spectral workflows rather than direct replacements.

• Operational Trade-off: While the purpose-built Sentinel-2 spectral baseline remains the superior option when accurate fire dates and suitable imagery are available, TESSERA approaches this performance without requiring event-specific image differencing. This makes TESSERA particularly useful for scenarios where a single annual feature layer must support the classification of multiple fire events.
• Limitations: The study notes that embeddings do not yet match the accuracy of spectral baselines. Furthermore, the evaluation was limited to 20 events from a single year, and the validation strategy did not test temporal transferability across different fire seasons. The reported IoU is a cell-based approximation based on sample points, not a full raster-to-polygon comparison.
• Future Directions: The authors suggest future work should focus on full raster-to-perimeter validation, assessing temporal transferability, and exploring hybrid feature stacks that combine embeddings with interpretable spectral features to balance performance, scalability, and interpretability.