📊 statistics

Evaluating AlphaEarth Foundations Embeddings for Wildfire Susceptibility Mapping

This study demonstrates that AlphaEarth Foundations embeddings serve as a highly effective, transferable alternative to traditional physical variables for wildfire susceptibility mapping, achieving superior model performance and cross-regional applicability in Victoria, Australia, and surrounding areas.

Original authors: Yuan Zhuang, Sanaa Hobeichi, Peng Shi, Fei Huang

Published 2026-08-14

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuan Zhuang, Sanaa Hobeichi, Peng Shi, Fei Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict where a wildfire might strike next. For decades, scientists have treated this like a giant, complicated recipe. To get the right answer, they had to gather dozens of specific ingredients from different jars: maps of the hills, satellite photos of the grass, weather reports on wind speed, and soil moisture readings. The problem? These ingredients often come in different units, from different years, and sometimes the jars are empty or the labels are missing. It's like trying to bake a cake when the flour is in a bucket, the sugar is in a bag, and the oven manual is in a different language. You have to spend hours just measuring and mixing everything together before you can even turn the oven on.

Recently, a new kind of "super-ingredient" has appeared in the world of Earth science called a "foundation model." Think of this as a pre-mixed, all-in-one cake batter. Instead of measuring flour and sugar separately, you just grab a scoop of this batter, which has already learned what the world looks like by studying billions of photos and data points. The big question for scientists is: Can we just use this pre-mixed batter to predict wildfires, or do we still need to measure every single ingredient ourselves? This is the story of a new study that tested if this "super-batter" works for keeping our communities safe from fire.

The Great Fire Prediction Experiment

In this study, researchers decided to put the "super-batter" to the test using the state of Victoria, Australia, as their giant kitchen. Victoria is a perfect place for this experiment because it's a land of extremes: it has busy cities, dry deserts, and lush, fire-prone forests. The team wanted to see if they could predict wildfire risks using the new AlphaEarth Foundations (AEF) embeddings (the fancy name for the "super-batter") compared to the old-school method of mixing physical variables like temperature and wind manually.

The Ingredients: Old vs. New
The traditional method is like building a puzzle piece by piece. Scientists gather data on elevation, how much rain fell, how dry the soil is, and what kind of plants are growing. They have to clean this data, make sure the pieces fit together, and then feed it into a computer model. It's accurate, but it's a lot of work.

The new method uses AEF embeddings. Imagine a robot that has looked at the entire Earth for years, taking pictures and reading weather data. Instead of giving you a list of numbers, it gives you a "summary code" for every patch of land. This code is a 64-number vector that captures everything about that spot—the trees, the hills, the weather history—all wrapped up in a neat package. The researchers asked: Does this summary code know enough about fire to be useful?

The Taste Test: How Well Did It Work?
The researchers trained computer models to predict where fires happened between 2017 and 2025. They used two types of models:

The "Cell-by-Cell" Model: This looked at each square of land individually, like checking one cookie at a time.
The "Neighborhood" Model: This looked at a patch of land (like a 17x17 grid of squares) to see how the surroundings influenced the fire risk, similar to checking if the whole cookie tray is baking evenly.

The results were surprisingly tasty. When using the AEF embeddings, the models achieved a score (called ROC-AUC) above 0.92. This is a very high score, meaning the model was excellent at distinguishing between places that burned and places that didn't. In fact, for the "neighborhood" models, the AEF embeddings performed just as well as, and sometimes even better than, the carefully hand-crafted physical variables.

The maps they produced looked very realistic. They correctly identified the high-risk areas in eastern Victoria (like the Gippsland region) and the mountainous uplands, which matches what experts already know. They also spotted some smaller, isolated hotspots in the northwest. When they overlaid these maps with the actual burned areas from a massive fire in January 2026 (which happened after their data ended), the maps lined up perfectly with where the fire actually went. This suggests the model learned the real rules of fire, not just memorized the past.

The Secret Sauce: What's Inside the Code?
You might wonder, "How does a code know about fire?" The researchers did a "reconstruction" test. They tried to reverse-engineer the code to see if they could pull out the original ingredients, like temperature or wind speed.

The Good News: They could pull out the basics with high accuracy. Things like slope, solar radiation, total rainfall, and wind speed were clearly encoded in the AEF summary.
The Bad News: They struggled to pull out very specific, extreme details, like "the number of days the fire danger index was above 50." The code is great at the general picture but a bit fuzzy on the extreme, short-term spikes.

The Real Magic: Traveling to New Places
Here is where the story gets really exciting. Usually, if you train a fire model in one place (like Victoria) and try to use it in another place (like Canberra or Sydney), it falls apart. It's like a chef who knows how to cook Australian BBQ but tries to cook a Japanese sushi dinner using the same recipe; it usually doesn't work.

When the researchers tested their models on new regions:

The Old Method (Physical Variables): When they took the traditional models trained in Victoria and applied them to nearby Canberra, the accuracy dropped by about 25%. It was a disaster.
The New Method (AEF Embeddings): When they used the AEF models, the accuracy at Canberra actually improved by about 4%. Even in Western Sydney, the drop in performance was tiny (only about 2%).

This suggests that the "super-batter" learned universal rules about the Earth that apply across different regions, whereas the old method was too stuck on the specific details of Victoria. However, the researchers noted that if they went too far away to very different climates (like the tropical north or the arid interior), the performance did start to drop, just not as sharply as the old method.

What Didn't Work?
The researchers also tried to feed the models a "time sequence," giving them data year-by-year from 2017 to 2025 to see if the model could learn how fire risk changes over time. Surprisingly, this didn't help much. The models didn't get significantly better at predicting fire risk just by knowing the history of each year. This suggests that for wildfire susceptibility (how likely a place is to burn), the permanent features of the land (hills, vegetation, climate) matter more than the year-to-year changes.

The Takeaway
This study suggests that we might not need to spend hours manually gathering and cleaning dozens of different data files to predict wildfires anymore. The AlphaEarth Foundations embeddings act as a powerful, ready-to-use tool that captures the essential information about the landscape. While it might not be perfect for predicting the exact moment a fire starts tomorrow, it is incredibly good at mapping out which areas are naturally prone to fire.

For governments and insurance companies, this is a game-changer. It means they can create large-scale fire risk maps much faster and with less effort, and these maps can be trusted to work even when they move from one state to another. It's like having a master chef who can cook a delicious meal for a whole country using a single, pre-mixed ingredient, rather than needing a different recipe for every single town.

Technical Summary: Evaluating AlphaEarth Foundations Embeddings for Wildfire Susceptibility Mapping

Problem Statement
Wildfire susceptibility mapping is critical for emergency planning and risk assessment but faces significant practical bottlenecks. Traditional data-driven approaches rely on assembling and harmonizing diverse physical variables (e.g., climate, topography, vegetation) from multiple remote sensing and geospatial sources. This process requires substantial feature engineering, scientific judgment for variable selection, and complex data harmonization to ensure consistent spatial resolutions and formats. Furthermore, models trained on these physical variables often fail to transfer reliably to new regions, necessitating the redevelopment of data pipelines and retraining of models for each new study area. This study investigates whether AlphaEarth Foundations (AEF) embeddings—a publicly available, analysis-ready geospatial product derived from a foundation model trained on billions of Earth observation records—can serve as a robust, reusable alternative to hand-engineered physical variables for wildfire susceptibility mapping.

Methodology
The study utilizes Victoria, Australia (2017–2025) as the primary source domain, with eight transfer sites across Australia (ranging from climatically similar temperate zones to distinct subtropical and arid environments) to evaluate spatial generalization.

Data Preparation:
- Physical Variables: A comprehensive set of 35 predictors (topography, land cover, climate, fire weather, etc.) was constructed and harmonized to a 1 km grid. Variance Inflation Factor (VIF) screening reduced this to 22 retained predictors.
- AEF Embeddings: Annual 64-dimensional embedding vectors (10 m native resolution) were de-quantized, aggregated to the 1 km grid, and L2-normalized.
- Labels: Fire occurrence data (MODIS hotspots, 2017–2025) were rasterized to create a binary susceptibility label (fire vs. no-fire) for each grid cell. Pseudo-absence samples were selected with spatial constraints to avoid environmental similarity with fire cells.
Modelling Frameworks:
The study compared models trained on physical variables versus AEF embeddings across three settings:
- Time-Aggregated (Cross-Sectional): Using aggregated feature vectors for the entire study period. Models included tree-based ensembles (Random Forest, XGBoost, LightGBM), Multi-Layer Perceptrons (MLP), TabPFN (a tabular foundation model), and Convolutional Neural Networks (CNNs) using local spatial patches (9×9, 17×17, 25×25 km).
- Sequence-Informed: Models incorporating annual feature sequences (2017–2025) with a learnable weighting mechanism to capture interannual variations, predicting the same cross-sectional fire occurrence label.
- Reconstruction Analysis: A regression task was performed to predict individual physical variables directly from AEF embeddings to quantify how much wildfire-relevant environmental information is encoded within the embeddings.
Evaluation:
- Within-Victoria: Performance was measured using Accuracy, Recall, Specificity, F1-score, and ROC-AUC on a held-out test set.
- Transferability: Models trained in Victoria were applied without modification to the eight transfer sites. Performance degradation was quantified relative to the source region.

Key Results

Predictive Performance: AEF embeddings achieved high predictive accuracy, with ROC-AUC values consistently above 0.92 in Victoria, comparable to models using hand-engineered physical variables.
- Model Architecture: While tree-based ensembles slightly favored physical variables, CNN-based models utilizing local spatial patches significantly outperformed tabular baselines (including TabPFN) when using AEF embeddings. The marginal gain from spatial context was larger for embeddings than for physical variables, suggesting embeddings encode rich spatial structure.
- Temporal Sequencing: Incorporating annual sequences provided only modest performance gains (<1.5%), suggesting that for susceptibility (a persistent trait), time-aggregated representations are sufficient.
Reconstruction Analysis: AEF embeddings successfully reconstructed many key physical drivers (e.g., slope, solar radiation, temperature, precipitation, wind) with median $R^2$ values above 0.8. However, variables related to fire-season persistence and extreme conditions (e.g., consecutive dry days) showed lower and less stable reconstruction accuracy.
Transferability:
- Physical Variables: Models trained on physical variables suffered a sharp performance drop (approx. 25% decrease in ROC-AUC) when applied to nearby climatically similar regions like Canberra and Western Sydney.
- AEF Embeddings: Embedding-based models demonstrated superior transferability. In Canberra, performance improved by ~4%, and in Western Sydney, the decline was minimal (~2%). Performance degradation was more gradual and only became significant in climatically distinct environments (e.g., tropical Queensland, arid Northern Territory).
Spatial Patterns: Both embedding-based and physical-variable models successfully identified high-susceptibility zones in eastern and southwestern Victoria, aligning with historical fire patterns and the January 2026 bushfire perimeter. CNN-based maps produced smoother, more contiguous susceptibility zones compared to the finer-grained, pixelated outputs of tree-based models.

Significance and Claims
The paper claims to provide the first systematic evaluation of Earth foundation model embeddings for wildfire susceptibility mapping. Its primary contributions are:

Validation of Embeddings: Demonstrating that AEF embeddings encode sufficient wildfire-relevant environmental information to support high-accuracy susceptibility models without the need for complex feature engineering.
Scalability and Transferability: Establishing that embedding-based models offer a more scalable workflow for large-area mapping. Unlike physical-variable models that require region-specific redevelopment, AEF models show strong "near-region" transferability, potentially reducing the data and technical barriers for government agencies and insurers operating across diverse jurisdictions.
Architectural Insight: Highlighting that foundation model embeddings should be treated as spatially structured fields rather than simple tabular covariates, as CNNs leveraging local neighborhood context extract additional predictive signal.
Practical Baseline: Positioning AEF embeddings as a meaningful performance baseline for researchers; if a physical-variable model cannot exceed the performance of an embedding-based baseline, it suggests a need to revisit variable selection or data processing.

The authors conclude that while AEF embeddings are not a universal solution for all climatic regimes, they represent a powerful, reusable geospatial data asset for creating scalable wildfire risk mapping workflows, particularly in regions with climatic similarity to the training domain.

The Great Fire Prediction Experiment

More like this