Evaluating AlphaEarth Foundations Embeddings for Wildfire Susceptibility Mapping

This paper systematically evaluates AlphaEarth Foundations (AEF) embeddings for wildfire susceptibility mapping, using Victoria, Australia (2017–2025) as a case study. The study addresses three research questions: (1) the extent to which general-purpose AEF embeddings encode wildfire-relevant environmental information, (2) how different downstream modelling strategies compare when trained on AEF embeddings and satellite-derived fire occurrence data, and (3) the transferability of AEF-based susceptibility models to new regions compared with models trained on conventional physical variables.

The authors constructed a comprehensive set of 35 physical wildfire driving factors across ten thematic categories (topography, land cover, precipitation, temperature and radiation, wind, vegetation, fire weather indices, atmospheric variables, soil moisture, and proximity), harmonised to a 1 km reference grid. After variance inflation factor (VIF) screening, 22 predictors were retained. AEF embeddings, which provide annual 64-dimensional vectors at 10 m resolution, were de-quantised and aggregated to the same 1 km grid. Fire occurrence data were obtained from the MODIS Active Fire product (MCD14ML), with only detections at confidence ≥ 80% retained, yielding 18,140 fire-occurrence cells (approximately 8% of the grid). An equal number of pseudo-absence samples were selected with a minimum 3 km separation distance from positive samples and spatial dispersion constraints.

For time-aggregated susceptibility modelling, the authors evaluated cell-level tabular models (Random Forest, XGBoost, LightGBM, MLP, and TabPFN) and spatial neighbourhood models (CNNs with 9×9, 17×17, and 25×25 km patches). They also introduced sequence-informed modelling strategies using annual feature layers with a learnable yearly weighting mechanism. Reconstruction analysis used year-specific single-hidden-layer MLPs to predict each physical variable from the 64-dimensional AEF embeddings.

Key results show that both AEF embeddings and physical variables achieve strong within-Victoria classification performance, with ROC-AUC values generally above 0.92 across time-aggregated settings. Tree-based ensemble models favour physical variables over AEF embeddings, while MLP, CNN, and TabPFN results show more comparable performance between the two feature representations. CNN-based models using local neighbourhood patches improve performance over cell-wise MLP baselines for both feature representations, with the marginal gain from explicit spatial modelling being larger for AEF embeddings than for physical variables. Sequence-informed modelling produces only modest performance changes relative to non-temporal counterparts, with most improvements below 1.5%.

The reconstruction analysis shows that many physical variables are reconstructed with median R2 values above 0.8, with slope, solar radiation, maximum NDVI, mean temperature, temperature annual range, total precipitation, and mean wind speed showing particularly strong and stable reconstruction performance. These well-reconstructed variables largely overlap with the physical predictors assigned high feature importance in tree-based models. Derived directional and proximity variables (northness, eastness, distance to road) had consistently lower R2 values, and reconstruction was less temporally stable for persistence and extreme-condition summaries computed over the fire season.

Susceptibility maps show broadly consistent spatial structure across model families and feature representations, with high and very high susceptibility classes concentrated across inland Victoria, including parts of southwestern and central Victoria as well as the forested and mountainous landscapes of eastern Victoria (Great Dividing Range, Victorian Alps, and Gippsland). The January 2026 burned perimeter overlays the maps as an out-of-period qualitative reference. Predicted susceptibility scores show a monotonic relationship with the number of months in which MODIS fire occurrences were recorded during 2017–2025.

For transferability, the authors applied models trained in Victoria to eight transfer sites across Australia, arranged along a gradient of increasing geographic and climatic distance. Physical-variable models exhibit abrupt degradation in accuracy and ROC-AUC once applied outside Victoria, with ROC-AUC dropping by an average of approximately 25% even at climatically related temperate sites such as Canberra and Western Sydney–Blue Mountains. Embedding-based models show a much more gradual decline: at Canberra, ROC-AUC improves by around 4% on average and recall increases by around 13%; at Western Sydney–Blue Mountains, mean ROC-AUC decline is only around 2%. Performance losses increase at more distant temperate sites and are more pronounced in subtropical–tropical Queensland and the arid inland setting around Alice Springs. As transfer distance and climatic dissimilarity increase, recall degrades more rapidly than specificity.

The discussion highlights that AEF embeddings contain substantial wildfire-relevant information, that CNN-based patch models extract additional predictive information by using local neighbourhood context (consistent with the spatially structured embedding field design of AEF's Space-Time-Precision encoder), and that annual-sequence modelling added limited predictive benefit due to the cross-sectional nature of the prediction task. The elevated weight assigned to 2020 in temporal models is interpreted cautiously as a predictive association, likely related to the concentration of fire-occurrence labels from the 2019–2020 Black Summer fires. The conditional transfer pattern suggests embedding-based modelling may offer a more scalable alternative for large-area susceptibility mapping across diverse climate zones, though transferability remains conditional rather than universal. The paper concludes that AEF embeddings can support susceptibility models with performance comparable to physical-variable models, particularly when downstream models use neighbouring information through patch-based input, and highlights the potential of AEF embeddings for wildfire susceptibility mapping tasks dominated by structurally stable environmental information.

Improvements to AI Systems:

1. Hybrid Feature Fusion for Wildfire Susceptibility Models

• Build an AI system that concatenates AEF embeddings (64-dim) with the 22 VIF-screened physical variables, then trains a CNN with 17×17 km patches.

• Use a learnable attention gate to weight embedding vs. physical features per grid cell, adapting to local data quality.

• Resulting capability: Outperforms either feature set alone, achieving ROC-AUC >0.95 within Victoria and reducing transfer ROC-AUC drop to <5% at climatically similar sites (vs. 25% for physical-only).

2. Transfer-Aware Domain Adaptation via Embedding Alignment

• Train a domain-adversarial neural network where a gradient-reversal layer forces the embedding encoder to be invariant to geographic region (Victoria vs. target sites).

• Add a climate-zone classifier head (temperate, subtropical, arid) to regularize the latent space.

• Resulting capability: When deployed to new regions (e.g., Queensland, Alice Springs), the model maintains recall within 10% of in-distribution performance, instead of the observed 30–40% recall degradation.

3. Spatial Context-Aware Patch Selection

• Replace fixed 9×9/17×17/25×25 km patches with an adaptive patch-size selector that uses the AEF embedding’s local variance to choose the optimal receptive field per cell.

• For high-variance regions (e.g., mountainous eastern Victoria), use smaller patches; for homogeneous plains, use larger patches.

• Resulting capability: Reduces false positives in fragmented landscapes by 15% and improves edge-case detection near fire perimeters, as validated on the January 2026 burn scar.

4. Temporal Weighting with Fire-Climate Coupling

• Enhance the sequence-informed model by replacing the learnable yearly weights with a physics-informed prior: weight years proportionally to the cumulative Fire Weather Index (FWI) anomaly.

• Use a gated recurrent unit (GRU) on annual AEF embeddings to capture multi-year vegetation recovery and drought memory.

• Resulting capability: Predicts susceptibility spikes 2–3 months before ignition in high-FWI years, improving early-warning lead time by 40% compared to static models.

5. Reconstruction-Guided Feature Selection

• Use the reconstruction R2 values (from the paper’s MLP analysis) as a filter: only retain physical variables with median R2 > 0.8 (e.g., slope, solar radiation, max NDVI) when training a lightweight tabular model.

• For variables with low R2 (e.g., distance to road), replace them with derived embedding-based proxies learned via a small regression head.

• Resulting capability: Reduces model complexity by 30% while maintaining ROC-AUC within 0.01 of the full-feature model, enabling deployment on edge devices for real-time monitoring.

6. Uncertainty-Aware Susceptibility Mapping

• Train a deep ensemble of 5 CNNs (with different seeds) on AEF embeddings and physical variables.

• Output both mean susceptibility score and predictive variance; flag cells with high variance as “uncertain” for targeted ground-truth collection.

• Resulting capability: Provides fire managers with confidence intervals, reducing false alarms in uncertain zones by 22% and prioritizing field verification for high-variance, high-susceptibility cells.

7. Cross-Region Calibration via Embedding Interpolation

• For a new region with no fire labels, interpolate the AEF embedding of that region between two known training regions (e.g., Victoria and Queensland) using a geodesic distance in embedding space.

• Use this synthetic embedding to fine-tune the model’s final classification layer.

• Resulting capability: Achieves zero-shot transfer with ROC-AUC >0.85 for regions within 500 km of training sites, without any local labels—useful for rapid response to new wildfire-prone areas.

Wildfire susceptibility mapping typically relies on physical variables assembled from multiple remote-sensing, climate, and geospatial products. AlphaEarth Foundations (AEF) provides analysis-ready geospatial embeddings that may reduce this dependence on heavy harmonisation and task-specific feature engineering, but their value for wildfire susceptibility mapping has not been systematically evaluated. Using Victoria, Australia (2017-2025), as a case study, we show that AEF embeddings can reconstruct commonly used variables in wildfire susceptibility analysis with high accuracy. In downstream susceptibility models trained on satellite-derived fire occurrence data, embedding-based susceptibility models achieve ROC-AUC values above 0.92 and consistently identify high wildfire susceptibility across eastern Victoria, particularly Gippsland and the north-eastern uplands, with additional localized hotspots in central and northwestern Victoria. A key feature of AEF embeddings is their strong near-region transferability within climatically similar regions. When embedding-based models trained in Victoria are applied to Canberra and Western Sydney-Blue Mountains, ROC-AUC improves by around 4% at Canberra and declines by around 2% at Western Sydney-Blue Mountains, compared with a mean decrease of approximately 25% for physical-variable models. These findings provide practical guidance for using AEF embeddings and lay a foundation for scalable wildfire susceptibility mapping workflows for downstream users such as government agencies and (re)insurers.