🤖 machine learning

Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings

The Planetary Prediction Engine (PPE) is an autonomous AI system that streamlines geospatial modeling by automatically retrieving and fusing multimodal data with foundation model embeddings to optimize task-specific architectures, consistently outperforming manual expert baselines across diverse global challenges such as health monitoring, disaster risk assessment, and food security.

Original authors: Evelyn Ma, Rama Kumar Pasumarthi, Kishwar Shafin, Mandar Sharma, Mimi Sun, Hamed Sadeghi, Dav M. Ebengo, Mbulayi Onesime, Rouslan Solomakhin, John Wamburu, William Ogallo, Aisha Walcott-Bryant, Sanxin

Published 2026-08-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Evelyn Ma, Rama Kumar Pasumarthi, Kishwar Shafin, Mandar Sharma, Mimi Sun, Hamed Sadeghi, Dav M. Ebengo, Mbulayi Onesime, Rouslan Solomakhin, John Wamburu, William Ogallo, Aisha Walcott-Bryant, Sanxing Chen, Arbaaz Muslim, Yael Mayer, Ronald Ho, Roy Lee, Ruth Alcantara, Abdoulaye Diack, Monica Bharel, Lambert Rosique, Jeremy Amez-Droz, Christopher Haire, James Manyika, Yossi Matias, Niv Efron, Gautam Prasad, Shravya Shetty

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 a world where the most critical decisions about our planet's health—where to send food aid during a drought, which neighborhoods need vaccines first, or how to prepare for a spreading disease—depend on maps that are either too blurry to be useful or take months to draw. For decades, creating these high-fidelity maps has been a slow, manual process. Experts have had to hunt through scattered digital archives, clean up messy data, and painstakingly stitch together satellite images with census records and weather reports. This work requires deep specialization and significant time, often leaving the most vulnerable regions without timely insights when crises strike. The challenge has not been a lack of data, but rather the difficulty of connecting the right pieces of information at the right speed to build a clear picture of the future.

A new system called the Planetary Prediction Engine, developed by researchers at Google, aims to change this dynamic by automating the entire process of building these predictive maps. Instead of requiring a team of specialists to spend weeks gathering data and tuning models, this system allows a user to simply ask a question in plain language, such as "predict the spread of an outbreak" or "estimate food insecurity in a specific region." The system then acts as an autonomous researcher, scouring the web and massive satellite databases to find the most relevant information, cleaning and combining it, and finally training a computer model to make predictions. It does this by using a type of advanced artificial intelligence that understands the deep patterns of human populations and land use, allowing it to see connections that a human might miss. The result is a tool that can generate expert-level forecasts in a fraction of the time it used to take, making high-quality geospatial analysis accessible to organizations that need it most.

The researchers tested this engine on three very different types of challenges to see if it could truly replace the slow, manual methods. First, they looked at how well it could predict the spread of disease, specifically tracking a real-world Ebola outbreak in the Democratic Republic of the Congo in 2026. In this scenario, the system had to identify which health zones were likely to be hit next. While previous expert models could identify about 73 percent of the new outbreak zones, the autonomous engine successfully predicted 83 percent of them. It achieved this by combining standard disease data with unique signals about how people move and where they live, which the system discovered and selected on its own. This improvement means that health workers could have been alerted to dangerous areas sooner, potentially saving lives by directing resources to the right places before the virus arrived.

Next, the team tested the system's ability to zoom in on data that was originally too coarse to be helpful, a process known as downscaling. In Nigeria, food security data is often only available at the state level, which is too broad to help local communities. The researchers asked the engine to predict food insecurity at the much smaller level of local government areas. A traditional approach using basic statistics and simple interpolation managed to explain only about 31 percent of the variation in food security at this local level. The Planetary Prediction Engine, however, doubled that accuracy, reaching 66 percent. It accomplished this by automatically finding and using specific local signals, such as changes in market prices, weather patterns, and even the brightness of city lights at night, to fill in the gaps where direct surveys were missing. This capability suggests that the system can provide detailed, actionable maps for regions where data is scarce, turning rough regional estimates into precise local guides.

Finally, the researchers evaluated the system on standard statistical tasks across the United States, where data is abundant and expert models already exist. They asked the engine to predict health outcomes like obesity and diabetes rates, as well as environmental risks like flood exposure. In these areas, the system consistently outperformed the best manually crafted models created by human experts. For instance, when predicting health indicators across the country, the engine achieved an accuracy score of 76.8 percent, compared to 60 percent for the expert baseline. Similarly, for predicting social vulnerability, it improved the accuracy from roughly 58 percent to 66 percent. These results demonstrate that the system does not just work in data-poor environments; it can also find better ways to combine information in data-rich ones, uncovering patterns that human experts had overlooked.

The success of the Planetary Prediction Engine relies on how it handles the complexity of the data it gathers. When the system starts a task, it does not just grab the first available numbers; it acts like a rigorous scientist, checking every piece of data to ensure it does not accidentally use information that would not be available at the time of prediction. It filters out any data that is a direct result of the thing being predicted, ensuring the model learns to forecast the future rather than just memorizing the past. It also balances different types of information, mixing hard numbers like census counts with the subtle, learned patterns found in satellite imagery and population models. This careful balancing act allows the system to build a robust understanding of a place, combining the concrete facts of demographics with the hidden textures of the landscape.

What makes this work particularly significant is that it shifts the role of the human researcher from a data engineer to a director. Instead of spending weeks cleaning spreadsheets and debugging code, a user can now focus on the question they want to answer. The system handles the heavy lifting of finding the data, fusing different sources, and testing thousands of model variations to find the best one. While the system is not perfect and still faces challenges, such as knowing exactly how to handle the noise that comes with very high-resolution data, it represents a major step forward. It proves that autonomous systems can not only match but often exceed the performance of human experts in complex geospatial tasks, offering a new way to understand and protect our planet. By lowering the technical barriers to entry, this technology opens the door for more organizations to use advanced modeling to solve urgent global problems, from famine to disease, with a speed and precision that was previously out of reach.

Technical Summary: Planetary Prediction Engine (PPE)

1. Problem Statement

Addressing critical global challenges—such as food insecurity, disaster risk mitigation, and disease outbreak tracking—requires high-fidelity, real-time geospatial modeling. However, the current workflow for building planetary-scale predictive models is severely bottlenecked by a fragmented data ecosystem. Traditional approaches rely on labor-intensive, manual processes involving:

Data Discovery: Manually navigating decentralized public and private repositories to find relevant proxy signals.
Curation and Fusion: Retrieving, cleaning, and fusing multimodal data (e.g., satellite imagery, demographic statistics, open-web reports) with domain-specific geospatial foundation models (e.g., Population Dynamics Foundation Model [PDFM], AlphaEarth).
Model Engineering: Designing architectures that handle spatial dependencies while enforcing strict validation guardrails to prevent target leakage.

These workflows often involve hundreds of steps (e.g., >700 steps for epidemiological nowcasting), requiring specialized teams and weeks of development time. Existing automated machine learning (AutoML) frameworks are ill-suited for this domain as they typically operate on pre-curated tabular data and lack the capability to autonomously discover, align, and fuse heterogeneous geospatial embeddings or mechanistic epidemiological models.

2. Methodology: The Planetary Prediction Engine (PPE)

The Planetary Prediction Engine is an autonomous AI system designed to execute end-to-end geospatial prediction workflows directly from natural-language queries. The system is built on a modular architecture orchestrated by Large Language Models (LLMs) and consists of three distinct stages:

Stage 1: Intelligent Data Selection

This stage transforms a natural-language query into a prioritized, join-ready multimodal DataFrame.

Task Inference: The system identifies the predictive paradigm (e.g., spatial regression, super-resolution, epidemiological nowcasting) from the user prompt to constrain downstream data retrieval.
Signal Discovery: It performs "Grounded Signal Discovery," distinguishing between direct signals and proxy signals validated against literature.
Dynamic Retrieval: The system queries established repositories (Data Commons, Google Earth Engine, Google Maps Platform) and performs live open-web discovery for datasets (CSV, GeoJSON, Parquet) from government and academic sources.
Prioritization: Datasets are ranked using a five-dimension scoring rubric, heavily weighting provenance and license openness.
Anti-Leakage: A "Feature Gate" enforces strict criteria to prevent target leakage, ensuring covariates are not mathematical components of the target, do not share survey sources, are causally upstream, and are temporally valid.

Stage 2: Multimodal Dataset Curation

This stage fuses the selected statistical covariates with pre-trained geospatial foundation model embeddings.

Embedding Integration: The system integrates tabular covariates with latent representations from the Population Dynamics Foundation Model (PDFM) (330/512-dim socio-demographic embeddings) and AlphaEarth (64-dim satellite-derived land-use semantics).
Preprocessing: It applies automated target leakage mitigation, including "Split-Isolated Imputation" (computing imputation statistics only on the training set) and standard feature engineering (skewness correction, outlier clipping, scaling).
Data Handling: Data artifacts are passed between stages via opaque handles to avoid LLM context-window limitations.

Stage 3: Automated Model Building and Prediction

This stage performs iterative model selection and optimization.

Model Search: The system searches across diverse model families: Regularized Linear Models, Gradient Boosted Trees (HistGradientBoosting, XGBoost), and Multi-Layer Perceptrons (MLP).
Validation Strategies: It employs Random Split, Spatial Group Split (to prevent spatial autocorrelation leakage), and K-Fold Cross-Validation.
Overfitting Guards: A multi-layered protocol includes pre-training risk assessment (based on sample size and feature ratios) and a self-correction loop that discards models showing catastrophic generalization failure and restarts with conservative configurations.

3. Key Contributions

The paper outlines four primary methodological contributions:

Autonomous Modular Agent Architecture: A fully automated end-to-end system that performs research-grounded signal discovery, feature engineering, and model training without manual intervention.
Task Auto-Identification & Objective Alignment: The ability to infer specific geospatial tasks (e.g., super-resolution, nowcasting) from high-level human specifications and align the AutoML optimization accordingly.
Intelligent Data Selection & Multimodal Curation: A protocol that combines heterogeneous geospatial embeddings with dynamically selected statistical covariates, systematically exploring feature combinations that human practitioners might overlook while enforcing strict leakage mitigation.
Automated Model Building & Prediction: An iterative search protocol incorporating overfitting guards and self-correction loops to select optimal models for final predictions.

4. Experimental Results

The system was benchmarked across a multidimensional matrix covering three paradigms: Epidemiological Nowcasting, Super-Resolution Downscaling, and Spatial Regression.

Epidemiological Transmission (DRC Ebola Outbreak)

Task: Nowcasting the spatial invasion of the 2026 Bundibugyo Ebola outbreak in the Democratic Republic of Congo (DRC).
Performance: PPE achieved a Recall@10 of 83.3%, correctly identifying 15 of 18 newly invaded health zones across five weekly forecasts.
Comparison: This represents a +10.3 percentage point absolute improvement over the public state-of-the-art Bayesian modeling baseline (~73%).

Super-Resolution Downscaling

Nigeria Food Security: Downscaling Food Consumption Group (FCG) indicators from State (ADM1) to Local Government Area (ADM2) levels.
- Result: PPE achieved an $R^2$ of 66.1%, doubling the baseline accuracy of 31.5%.
US Social Vulnerability Index (SVI): Downscaling from County to ZIP Code Tabulation Areas (ZCTA).
- Result: PPE achieved an $R^2$ of 37.6%, a 26.6 percentage point improvement over the baseline of 11.0%. The fusion of covariates with PDFM embeddings was identified as the key driver.

Spatial Regression (US Data-Rich Settings)

CDC Health Indicators: Predicting 21 health variables (e.g., obesity, diabetes) at the Census Tract level.
- Result: PPE achieved a mean $R^2$ of 76.8%, significantly outperforming the manually tuned expert baseline of 60.0%.
FEMA Environmental Risk: Predicting 21 risk indices.
- Result: PPE achieved a mean $R^2$ of 64.9%, outperforming the expert baseline of 60.0% (and 59.9% for the full suite).
SVI Spatial Regression: Predicting 5 vulnerability scores at the County level.
- Result: PPE achieved a mean $R^2$ of 66.2%, a 6.8 percentage point improvement over the statistical covariate baseline (58.6%).

5. Significance and Claims

The paper claims that the Planetary Prediction Engine addresses a fundamental bottleneck in planetary analytics: the reliance on specialized human engineering teams for data discovery and model construction. By combining autonomous multimodal planetary data discovery with targeted model optimization, PPE:

Lowers Technical Barriers: Enables non-experts to instantly build predictive models for specific objectives via natural language.
Achieves Expert-Level Performance: Consistently outperforms state-of-the-art or manually tuned expert baselines across diverse tasks, geographies (Global North and South), and domains (public health, environmental risk, epidemiology).
Enables Rapid Deployment: Reduces development time from weeks to minutes, facilitating immediate humanitarian and policy responses in crisis-prone regions.

The authors emphasize that the system's success relies on the co-optimization of data and models, noting that neither intelligent data curation nor automated model selection alone achieves peak performance. The work is presented as a step toward democratizing geospatial prediction, shifting the researcher's role from manual data engineering to high-level hypothesis direction.