𝗗𝗲𝗲𝗽-𝗱𝗶𝘃𝗲: 𝘜𝘴𝘪𝘯𝘨 𝘎𝘰𝘰𝘨𝘭𝘦 𝘚𝘢𝘵𝘦𝘭𝘭𝘪𝘵𝘦 𝘌𝘮𝘣𝘦𝘥𝘥𝘪𝘯𝘨𝘴 𝘵𝘰 𝘗𝘳𝘦𝘥𝘪𝘤𝘵 𝘑𝘙𝘊 𝘍𝘰𝘳𝘦𝘴𝘵 𝘈𝘳𝘦𝘢𝘴 🌳🛰️
Not so long ago, Google DeepMind’s AlphaEarth Satellite Embeddings were released (𝘑𝘶𝘭𝘺, 2025), and the JRC Global Forest Cover Data dropped recently as well (𝘈𝘶𝘨𝘶𝘴𝘵, 2025). This presented the perfect circumstance to bring both datasets together in Google Earth Engine to test a key hypothesis 🧪.
𝗧𝗵𝗲 𝗴𝗼𝗮𝗹: 𝘊𝘢𝘯 𝘸𝘦 𝘤𝘳𝘦𝘢𝘵𝘦 𝘢 𝘧𝘢𝘴𝘵, 𝘩𝘪𝘨𝘩𝘭𝘺 𝘢𝘤𝘤𝘶𝘳𝘢𝘵𝘦, 𝘢𝘯𝘥 𝘳𝘦𝘭𝘪𝘢𝘣𝘭𝘦 𝘵𝘰𝘰𝘭 𝘧𝘰𝘳 𝘌𝘜𝘋𝘙 𝘤𝘰𝘮𝘱𝘭𝘪𝘢𝘯𝘤𝘦 𝘢𝘯𝘥 𝘳𝘪𝘴𝘬 𝘢𝘴𝘴𝘦𝘴𝘴𝘮𝘦𝘯𝘵? 🎯
The results were impressive: After tuning a Random Forest model, a high AUC (⬆️) and a low Brier Score (⬇️) give us confidence that AI-derived features are a game-changer for monitoring forest health and degradation. The Google DeepMind Foundation Model simplifies processing and eliminates the need for complex, manual feature engineering.
Check out the full analysis, which includes the methodology, detailed model performance comparisons (kNN, Random Forest vs. XGBoost), and the automated setup for feature selection.
Read the full article here 👇
The embeddings have great potential! I'm looking forward to any feedback, comments, or collaboration opportunities to build this out further. Let's connect 🤝 or follow me for more updates! https://www.linkedin.com/in/gvddool/
Notebook: Using Google Satellite Embeddings to Predict JRC Forest Areas [online]. Available at: https://gist.github.com/GvdDool/1e4d893319a487c49584cc70bc0d90d0
Recently, the Joint Research Centre (JRC) of the European Commission released the JRC Global Forest Cover 2020 dataset [1]. This dataset provides a global map of forest cover, offering a spatially explicit depiction of forest areas with standing trees for the year 2020 at a 10-metre spatial resolution in geographic coordinates (latitude and longitude). It was developed to support the implementation of the EU Deforestation Regulation (EUDR). The Global Forest Cover 2020 (GFC 2020) dataset, updated in 2024 (version 2), incorporates user feedback and revised spatial inputs, with its methodology, improvements, and accuracy assessments documented in a series of JRC technical reports. Although the GFC 2020 maps have no legal status, they are valuable tools for assessing deforestation and forest degradation risks in support of due diligence under the EUDR.
At the time of writing, the status of the EUDR remains uncertain; however, deforestation and land degradation continue worldwide regardless of its implementation. This ongoing environmental decline highlights the urgent need to maximise the value of emerging geospatial data products such as the JRC GFC 2020 dataset. In this context, I explored how this dataset could be integrated with AlphaEarth Foundations’ annual embeddings, developed by the Google DeepMind team [2]. These embeddings offer advanced, machine-learned representations of the Earth’s surface, which can enhance the monitoring, analysis, and understanding of global forest dynamics.
The Google Satellite Embedding dataset offers a powerful new way to represent the Earth’s surface using artificial intelligence. Instead of relying on traditional satellite imagery composed of visible or infrared spectral bands, this dataset combines information from multiple satellite sensors, including optical, radar, and LiDAR, to create compact “embeddings,” or numerical summaries of surface conditions. Each 10-metre pixel on the Earth’s surface is represented by a 64-value vector that captures how that location changes throughout the year. These embeddings summarise complex environmental patterns such as vegetation growth, water dynamics, and land use into a format that can be directly analysed by computer models.
One of the key advantages of this dataset is its near-global coverage, encompassing most terrestrial and shallow-water areas, making it suitable for environmental monitoring, land-cover mapping, and change detection studies. Because the embeddings are learned from multiple data sources, they are far less affected by clouds, data gaps, or other sensor artefacts than traditional optical imagery. The embedding system is also consistent across years, allowing comparisons over time to identify environmental change or degradation.
Although the embeddings themselves are not human-readable, their underlying structure makes them exceptionally powerful for automated analysis. By offering a ready-to-use, high-resolution, and AI-enhanced representation of the planet’s surface, the Google Satellite Embeddings dataset enables scientists, policymakers, and developers to study environmental change and support initiatives like the EUDR with greater accuracy and efficiency.
This raises a key question: can Google Satellite Embeddings be used to predict forest cover as defined by the JRC Global Forest Cover 2020 dataset? In other words, do these embeddings contain sufficient ecological information to distinguish forested from non-forested areas? Addressing this hypothesis could demonstrate the potential of embeddings as a flexible, high-resolution tool for environmental monitoring, complementing conventional Earth observation products and supporting initiatives such as the EUDR.
To test whether Google Satellite Embeddings contain information about forest cover, we treat each pixel as a 64-value vector from the embeddings dataset, and the corresponding forest label from the JRC Global Forest Cover 2020 dataset as a simple “forest” or “non-forest” category. The null hypothesis is that the embeddings contain no predictive information, meaning that a classifier trained on them would perform no better than a random classifier. The alternative hypothesis is that the embeddings contain meaningful ecological information, enabling a classifier to distinguish reliably between forested and non-forested pixels.
We can test this by training different models; a suggestion of such an analysis is detailed in the accompanying notebook [3]. The notebook uses three different models to analyse forest cover: k-Nearest Neighbors (kNN) as a baseline (as suggested by the Google Earth Engine (GEE) development team [4]), Random Forest, and XGBoost. The 64-dimensional embeddings serve as input, while forest/non-forest labels from the JRC dataset are used as the target.
Model performance is evaluated using key classification metrics, including the Area Under the Curve (AUC), Precision-Recall AUC, and Brier score. To ensure robust results, spatial cross-validation was performed grouped by 5×5 km blocks during training. To demonstrate that the embeddings contain useful information and reject the null hypothesis, any of the tested models should achieve an AUC score higher than 0.8 and a Brier score below 0.15.*
The analysis targeted a pilot area near Ebolowa, Cameroon (Figure 1), selected as a representative site within the tropical climate zone with diverse land use. A 100 km × 100 km study window was defined and divided into 5 km × 5 km blocks, creating ~400 blocks in total. Since forest is the predominant land cover type, the models find it easier to learn the characteristics of “forest.” Therefore, it was essential to create a balanced dataset that distinguishes between “forest,” as defined by the JRC, and other land cover types. Pixels were sampled from each block to ensure balanced representation of forested and non-forested areas, with a minimum distance of 100 meters to minimise spatial autocorrelation. To ensure the training data was viable, only blocks with a sufficient number of non-forest pixels were included. From the available blocks, 46 met this criterion, resulting in 46 × 150 forest points and 46 × 150 non-forest points. Six blocks had slightly fewer than 150 non-forest points, resulting in a final dataset of 13,523 sampled pixels.
Figure 1: Study Area and Sample Blocks
In this study, making the model understand the definition of a “forest” is the most important task because the JRC data product aims to align with the definitions established by the EU Deforestation Regulation (EUDR) and the Food and Agriculture Organization (FAO, 2018), and not to pixel-count the “green” pixels in an image.
Article 2(4) in the EUDR defines forest as:
- “‘forest’ means land spanning more than 0.5 hectares with trees higher than 5 metres and a canopy cover of more than 10 %, or trees able to reach those thresholds in situ, excluding land that is predominantly under agricultural or urban land use;”
Agricultural use and agricultural plantation are defined in Article 2 (5) and 2 (6), respectively, as follows:
- “‘Agricultural use’ means the use of land for the purpose of agriculture, including for agricultural plantations and set-aside agricultural areas, and for rearing livestock.”
- “‘Agricultural plantation’ means land with tree stands in agricultural production systems, such as fruit tree plantations, oil palm plantations, olive orchards and agroforestry systems where crops are grown under tree cover; it includes all plantations of relevant commodities other than wood; agricultural plantations are excluded from the definition of ‘forest’. “
In other words, all plantations of relevant commodities other than wood, such as cattle, cocoa, coffee, oil palm, rubber, and soya, are excluded from the definition of forest. As a result, the JRC forest layer represents only a subset of the broader land cover used as input data. For example, the map accurately identifies cacao and rubber plantations in Côte d’Ivoire and Liberia as non-forest, even when these plantations were established relatively recently within older and denser forest areas. However, some misclassifications occur in more complex settings, such as cacao plantations within degraded forests or locations where cacao and palm oil are intermixed. Additional confusion arises from the frequent omission of Eucalyptus plantations and, less commonly, from commission errors involving orchards and banana fields [5].
Figure 2: Globel Map of Forest Cover 2020 - version 2
The analysis used the 64-dimensional Google Satellite Embeddings for the year 2020, extracted at each sampled pixel. These data, available annually from 2017 to 2024, provide compact numerical representations of surface conditions derived from multiple satellite sensors and time periods. For this study, the 2020 embeddings were chosen to align with the JRC Global Forest Cover dataset. The embeddings serve as ready-to-use features for predictive modelling, allowing forest patterns to be inferred directly from their learned representations without the need for manual spectral index calculations. This concept is illustrated in the figure below, which shows how each location on Earth is represented as a 64-dimensional embedding vector within the AlphaEarth Foundations model, capturing temporal and multi-sensor information in a compact form.
Figure 3: Conceptual diagram of the AlphaEarth Foundation embedding field
Each location on Earth, as represented in Figure 3, is encoded by a 64-dimensional embedding vector that encompasses temporal and multisensor surface information. Adapted from DeepMind (2024), “AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data.” [6]
Random Forest (RF) classifiers were applied to the embeddings to predict the presence of forests. RF was selected for its strong performance with high-dimensional inputs, robustness to overfitting, and interpretability through feature importance. The k-Nearest Neighbors (kNN) baseline was also evaluated but ultimately rejected due to lower performance. For completeness, XGBoost was included in the comparative experiments; however, it was not chosen as the primary model because its results were very similar to those of RF, while being more complex and less interpretable.
Model performance (Embeddings only):
- Random Forest: AUC = 0.948, PR-AUC = 0.955, Brier = 0.090
- XGBoost: AUC = 0.951, PR-AUC = 0.958, Brier = 0.089
- k-Nearest Neighbors (kNN): AUC = 0.926, PR-AUC = 0.915, Brier = 0.103
To evaluate consistency and reliability, each model was tested using 5-fold spatial cross-validation. Figure 2 shows the ROC curves for Random Forest (RF), XGBoost, and k-Nearest Neighbors (kNN) side by side, with each curve representing one fold. The results demonstrate that Google Satellite Embeddings are highly effective for predicting forest presence. For practical applications, a Random Forest model using embeddings alone provides the best combination of predictive performance, calibration, and interpretability, while adding conventional spectral indices or switching to XGBoost offers minimal additional benefit.
Figure 4: RoC curves for Rf, XGBoost, and kNN
Figure 4 shows the. ROC curves for RF, XGBoost, and kNN across five spatial folds (Embeddings only). Each fold demonstrates model performance stability, with AUC, PR-AUC, and Brier scores indicated. RF and XGBoost consistently outperform the kNN baseline, confirming that Google Satellite Embeddings contain strong ecological information for predicting forest presence.
Before verifying the predictive capacity of the embeddings through geospatial analysis, a regional Random Forest (RF) model was developed for the project area. Initial validation used average configurations to provide a general overview of performance across various data folds (Figure 4); however, this approach did not yield the most optimised setup. To create a better-fitting model, the hyperparameters were refined using three-fold stratified cross-validation, and a final model was constructed using the complete dataset. The optimised RF achieved strong performance, with a mean AUC of 0.951 (matching previous XGBoost results), a mean PR-AUC of 0.959, a Brier score of 0.087, and a mean accuracy of 0.880. Spatial cross-validation confirmed consistency across folds, with ROC and PR curves demonstrating stable predictive behaviour.
To assess the predictive power of the embeddings in a real-world setting, a representative 5×5 km block near Ebolowa, Cameroon, was selected. This block was chosen because it contained a median number of non-forest pixels relative to the dataset, providing a balanced test case for validation. The final Random Forest model, trained on the full sampled dataset, was used to predict forest probability for each pixel in the block.
Predicted probabilities were mapped back to a 2D grid and compared against the JRC forest mask using two complementary thresholds: the F1-optimal threshold and Youden’s J statistic*. The model closely matched the JRC ground truth for this block, achieving an accuracy of 0.879 (slightly lower than the accuracy on the full dataset, but still well above the acceptance criteria of the hypothesis), an ROC AUC of 0.969, a PR AUC of 0.997, and a Brier score of 0.068. Figure 5 illustrates the predicted forest probabilities alongside the JRC classification, showing strong spatial agreement and demonstrating that the embeddings alone capture ecologically meaningful information sufficient for accurate forest mapping, including areas aligned with the FAO forest definition.
Figure 5: Predictive Forest Probability vs JRC Ground Truth
Figure 5 - Left: Random Forest predictions using Google Satellite Embeddings, with probability thresholds indicated by F1-optimal and Youden’s J statistic. Right: Binary forest/non-forest classification from the JRC Global Forest Cover 2020 dataset. The predicted probabilities closely match the ground truth, demonstrating the embeddings’ ability to capture forest patterns.
The analysis demonstrates that Google Satellite Embeddings are highly effective for predicting forest cover. A Random Forest model trained solely on embeddings yields the best combination of predictive performance, calibration, and interpretability, with minimal benefit from adding conventional spectral indices or switching to more complex models, such as XGBoost. The model shows consistent results across cross-validation folds and when geospatially validated on a representative 5×5 km block, indicating that the embeddings encode robust and ecologically meaningful information. These results suggest that embeddings can complement or even partially replace traditional remote-sensing inputs in operational forest monitoring, offering a more efficient, well-calibrated, and globally consistent approach.
Key findings:
High predictive power with minimal input – Unlike many traditional forest mapping approaches that rely on multiple spectral indices or complex pre-processing, this analysis shows that the 64-dimensional Google Satellite Embeddings alone are sufficient to produce highly accurate predictions. Consistency across folds and regions – The model performs consistently across cross-validation folds and when geospatially validated on a representative 5×5 km block, suggesting robustness of embeddings compared to other datasets that might be more sensitive to clouds, sensor gaps, or local anomalies. Calibration advantage – The Brier score is low, indicating that predicted probabilities are well-calibrated. Together with the AUC score, this provides confidence that the model is suitable for use in risk assessments or deforestation monitoring. Practical interpretability – Random Forest’s feature importance allows some interpretability, whereas more complex models like XGBoost or deep networks may give slightly higher accuracy but at the cost of transparency. Alignment with official datasets – Achieving these results using a pilot area validated against JRC GFC 2020 shows that embeddings can reproduce or complement authoritative datasets, potentially reducing the need for expensive, large-scale field validation.
Taken together, these results demonstrate not only the strong predictive power of Google Satellite Embeddings for a single year but also their potential as a foundation for longitudinal analyses, operational monitoring, and collaborative research in forest conservation. Building on these results, the following outlook highlights potential avenues for further leveraging embeddings in forest monitoring and environmental research:
Completeness of the embedding field – During exploratory experiments, adding conventional spectral indices or other derived datasets did not significantly improve predictive performance. This suggests that the Google Satellite Embeddings provide a highly complete representation of surface conditions, with additional data likely yielding benefits only in very specific contexts. A more detailed analysis of this topic will be presented in a forthcoming write-up.
Extending to temporal monitoring – The strong alignment with the 2020 JRC forest dataset opens the possibility to extend this approach to subsequent years. By using embeddings from 2021 to 2024 with the JRC data as labelled input, it becomes feasible to detect land-cover changes and monitor forest dynamics over time. This represents an opportunity for collaboration with environmental scientists, policy stakeholders, and conservation organisations to leverage embeddings for timely, global-scale forest monitoring initiatives.
This work was carried out independently and without external funding, as a non-commercial exploratory research activity. Many thanks to European Commission EU Science, Research and Innovation (#JRC) for the Global Forest Cover dataset, Google/ Google DeepMind for the AlphaEarth Foundations embeddings, and Google Earth Engine GEE for the computational framework; making these resources openly available enabled this analysis and exploration in forest monitoring.
a: Thresholds (Area under the Curve: AUC > 0.8, Brier < 0.15) were selected as practical benchmarks to indicate strong predictive performance in ecological mapping; they are not strict statistical cutoffs.
The Area Under the Curve (AUC) measures a model’s ability to distinguish between classes. It summarises the trade-off between the true positive rate and false positive rate across all possible classification thresholds. An AUC of 0.5 or below corresponds to random predictions, while higher values indicate better discriminative performance. The practical threshold of 0.8 was selected because predicting forest cover is challenging, and even strong models often struggle to achieve values above approximately 0.85.
The Brier score is a metric for evaluating probabilistic predictions in classification tasks. It calculates the mean squared difference between the predicted probability of an event (e.g., a pixel being forest) and the actual outcome (0 for non-forest, 1 for forest). Lower Brier scores indicate that the model’s predicted probabilities are closer to the true labels, reflecting both accuracy and calibration of the predictions.
b: Random Forest predictions using Google Satellite Embeddings, with probability thresholds indicated by F1-optimal and Youden’s J statistic.
Youden’s J statistic – Measures the effectiveness of a binary classifier by finding the threshold that maximises the difference between the true positive rate (sensitivity) and the false positive rate (1 – specificity). A higher value indicates a threshold that balances sensitivity and specificity, often used to select an optimal cut-off in diagnostic or classification tasks.
F1 score – The harmonic mean of precision and recall, providing a single metric that balances false positives and false negatives. It is particularly useful for imbalanced datasets, where accuracy alone may be misleading. Thresholds can be chosen to maximise F1, emphasising overall classification quality rather than only one aspect of performance.
[1] Bourgoin, C., Verhegghen, A., Degreve, L., Ameztoy, I., Carboni, S., Colditz, R. & Achard, F., 2024. Global map of forest cover 2020 – version 2 [dataset]. European Commission, Joint Research Centre (JRC). Available at: http://data.europa.eu/89h/e554d6fb-6340-45d5-9309-332337e5bc26. Also available through Google Earth Engine as JRC/GFC2020/V2 [Accessed 8 October 2025].
2] DeepMind, 2024. AlphaEarth Foundations helps map our planet in unprecedented detail [online]. Available at: https://deepmind.google/discover/blog/alphaearth-foundations-helps-map-our-planet-in-unprecedented-detail/. Also available through Google Earth Engine as GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL [Accessed 8 October 2025].
[3] Dool, 2025. Notebook: Using Google Satellite Embeddings to Predict JRC Forest Areas [online]. Available at: https://gist.github.com/GvdDool/1e4d893319a487c49584cc70bc0d90d0 [Accessed 9 October 2025].
[4] Google Earth Engine, 2025. Supervised classification using satellite embeddings [online]. Available at: https://developers.google.com/earth-engine/tutorials/community/satellite-embedding-03-supervised-classification [Accessed 9 October 2025].
[5] European Commission, Joint Research Centre, Bourgoin, C. et al., 2024. Mapping global forest cover of the year 2020 to support the EU Regulation on deforestation-free supply chains. Luxembourg: Publications Office of the European Union. Available at: https://data.europa.eu/doi/10.2760/262532 [Accessed 9 October 2025].
[6] Brown, C.F. et al., 2025. AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv preprint arXiv:2507.22291. Available at: https://arxiv.org/pdf/2507.22291 [Accessed 9 October 2025].