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

Abstract: Addressing critical global challenges, from food security and disaster risk to disease outbreaks and socio-economic vulnerability, demands high-fidelity geospatial modeling. However, building predictive planetary models remains bottlenecked by a fragmented data ecosystem, requiring manual data retrieval, multimodal data curation and fusion along with iterative model selection. We present the Planetary Prediction Engine (PPE), an autonomous AI system that executes this end-to-end workflow directly from natural-language queries. PPE synthesizes multimodal datasets on the fly, retrieving spatiotemporally relevant covariates across open-web and Earth observation platforms (Data Commons, Google Earth Engine) and fusing them with geospatial foundation model embeddings (PDFM, AlphaEarth). Simultaneously, it searches over task-tailored model architecture families with automated overfitting guards. Across diverse tasks, geographies, and scientific domains, PPE consistently outperforms state-of-the-art or manually tuned expert baselines. For US spatial regression, PPE improves mean $R<sup>2$ across 21 CDC health indicators (76.8% vs. 60.0%), FEMA national risk indices (64.9% vs. 60.0%), and the Social Vulnerability Index (66.2% vs. 58.6%). For spatial downscaling in data-scarce settings, PPE integrates localized proxies to double baseline accuracy in Nigerian food security indicators ($R<sup>2$ of 66.1% vs. 31.5%). For epidemiological nowcasting of the 2026 DRC Bundibugyo Ebola outbreak, PPE achieves a Recall@10 of 83.3% (identifying 15 of 18 newly invaded health zones across five weekly forecasts), a +10.3 percentage-point improvement over the public state-of-the-art modeling (~73%). By combining autonomous multimodal planetary data discovery with targeted model optimization, PPE lowers the technical barrier to planetary-scale analytics, enabling rapid, customized, expert-level deployment.

• The paper presents the Planetary Prediction Engine (PPE), an autonomous geospatial prediction system that optimizes data selection and model configuration; it combined structured covariates with foundation model embeddings, achieving up to 76.8% mean R2 for 21 CDC indicators and 83.3% Recall@10 for DRC outbreak hotspot prediction.
• PPE maps natural-language queries to geospatial prediction workflows with stages of Intelligent Data Selection, Multimodal Dataset Curation, and Automated Model Building for task-specific prediction. For Nigerian food-security downscaling, the system reached an R2 of 66.1%, significantly improving upon a baseline of 31.5%, demonstrating the system's adaptability across different data and tasks.
• For US CDC health indicators, PPE achieved superior performance with a 95% confidence of 76.1% to 77.6% mean R2, showing a 36.8% improvement over a model of censored embeddings and regulators

The Planetary Prediction Engine (PPE) is presented as an autonomous compound system for constructing geospatial predictive models from natural-language specifications. Its central premise is that planetary-scale analytics is constrained less by the absence of predictive algorithms than by the manual labor required to identify relevant signals, retrieve heterogeneous data, harmonize spatial and temporal supports, select representations, and enforce leakage-resistant validation. PPE addresses this bottleneck by integrating LLM-based orchestration, dynamic data discovery, geospatial foundation-model embeddings, automated feature engineering, and task-specific model search into a modular pipeline (2608.26088).

Problem formulation and system architecture

PPE maps a natural-language query—such as forecasting obesity rates, downscaling food insecurity, or predicting disease-transmission hotspots—into an executed geospatial prediction workflow. The architecture separates the process into three stages: Intelligent Data Selection, Multimodal Dataset Curation, and Automated Model Building and Prediction. Each stage has a fixed tool boundary and communicates through opaque artifact handles rather than embedding large datasets in LLM prompts. This design limits context-window dependence and prevents the prediction agent from retrieving or modifying data after curation. LLM calls use temperature zero, although this improves nominal reproducibility without eliminating nondeterminism originating from external repositories, evolving APIs, or model-serving infrastructure.

The first stage infers spatial granularity, join keys, temporal scope, and predictive task from the query and supplied labeled data. It then produces a research-grounded Signal Guide containing direct and proxy variables. Candidate data are retrieved from Data Commons, Google Earth Engine, Google Maps Platform Insights, OpenStreetMap, government portals, NGO repositories, and academic data stores. A provenance-first rubric prioritizes licensing and institutional provenance, followed by spatiotemporal fitness, signal alignment, data quality, and redundancy. The weighting scheme assigns provenance and license a fivefold weight, signal alignment a threefold weight, spatiotemporal fitness a twofold weight, and format quality and redundancy unit weights. This is a meaningful systems choice: the retrieval policy is not optimized solely for predictive utility, but also for operational legitimacy and reproducibility.

The second stage standardizes and fuses structured covariates with pretrained geospatial representations. The principal embeddings are the 330- or 512-dimensional Population Dynamics Foundation Model (PDFM), which encodes socioeconomic and demographic structure, and the 64-dimensional AlphaEarth Foundation (AEF) representation, which captures satellite-derived land-use, terrain, vegetation, and ecological semantics. Tabular covariates are log-transformed when highly skewed, clipped at the first and ninety-ninth percentiles, decorrelated, and standardized. Embeddings are instead L2-normalized without dimension-wise statistical transformations, preserving angular relationships in the learned representation space.

PPE also incorporates explicit anti-leakage controls. The Feature Gate excludes mathematical components of the target, variables derived from the same survey or imputation system, downstream effects, and observations occurring after the prediction window. Split-Isolated Imputation computes missing-value statistics solely on the training partition. These safeguards are particularly important for composite targets such as the Social Vulnerability Index (SVI), where apparently independent socioeconomic variables may in fact be ingredients of the target definition. The paper’s causal-direction filter is conceptually appropriate but ultimately heuristic; it cannot formally establish that every retained variable is upstream of the target.

The final stage performs model search over regularized linear models, histogram-based gradient boosting, XGBoost, and MLPs. Validation may use random, spatial-group, or three-fold cross-validation depending on the task. An Overfitting Guard uses sample size, feature-to-sample ratio, and spatial structure to constrain model complexity. If validation reveals catastrophic degradation or a large train–validation gap, PPE restarts with stronger regularization, although the self-correction loop is limited to one iteration. The system therefore automates a bounded search rather than unrestricted program synthesis.

Evaluation design

The evaluation spans three predictive paradigms, two broad geographic contexts, and several application domains. The benchmarks include US spatial regression for CDC health indicators, FEMA National Risk Index variables, and SVI; US county-to-ZCTA SVI downscaling; Nigerian state-to-LGA food-security downscaling; and epidemiological nowcasting of the 2026 Bundibugyo ebolavirus outbreak in the Democratic Republic of the Congo.

The tasks differ substantially in statistical structure. Spatial regression predicts missing values at the same geographic granularity. Super-resolution downscaling trains on coarse administrative observations and applies the model to fine-grained units with no corresponding labels. Epidemiological nowcasting predicts one-week increases in case counts or identifies previously uninfected health zones that subsequently experience infection. These settings make the reported results informative about adaptability, but they also make aggregate comparisons difficult: the baselines, sample sizes, spatial supports, and validation schemes are not uniform across tasks.

Epidemiological nowcasting and transmission prediction

For the DRC outbreak, PPE models weekly invasion risk across 519 health zones using surveillance signals, nowcasted incidence, mobility information, infrastructure, demographic variables, environmental covariates, and PDFM embeddings. The target is the increase in confirmed caseload over a seven-day horizon. Mobility information combines Flowminder relocation estimates with gravity and radiation models, while epidemiological preprocessing adjusts reporting delays and convolves incidence with several generation-time distributions.

The complete system obtains Recall@10 of 83.3%, identifying 15 of 18 newly invaded health zones over five weekly forecasts. This exceeds the published Bayesian baseline of approximately 73% by 10.3 percentage points. Adding geospatial covariates without the full selection procedure yields 77.8%, indicating that auxiliary spatial signals are useful even before foundation-model fusion and intelligent selection. The strongest configuration combines epidemiological signals, selected geospatial covariates, and PDFM embeddings.

The result is operationally relevant because top-ranked hotspot detection and accurate case-volume forecasting are distinct objectives. The paper’s expanded comparison reports that spatial transmission regression attains test top-10 accuracy of 0.8, whereas the enhanced SEIR nowcasting configuration achieves a test RMSE of 0.9204. PPE therefore supports the paper’s claim that frontier identification and caseload trajectory estimation should not necessarily be treated as the same modeling problem. The conclusion is constrained by the evaluation’s dependence on a single outbreak, a small number of weekly folds, and surveillance data whose preprocessing includes manually transcribed situation-report information. The confidence interval for the complete Recall@10 estimate is broad, 60.8%–94.2%, reflecting the limited number of newly invaded zones.

Food-security downscaling in Nigeria

The Nigerian benchmark evaluates whether coarse state-level food-security observations can be projected to 581 local government areas over 40 months. The baseline uses only month-of-year Fourier terms, a linear time trend, and interpolation, thereby representing seasonal and macroeconomic variation without localized spatial information. PPE adds food-price indices, WFP food-security measures, precipitation, NDVI, monthly vegetation indices, vegetation anomalies, nighttime lights, PDFM embeddings, and AEF features.

The full system reaches an $R^2$ of 66.1%, compared with 31.5% for the macro-covariate baseline. The intermediate configurations clarify the contribution of localized proxies: nighttime lights alone raises $R^2$ to 49.8%, vegetation features to 60.1%, and the full selected covariate suite to 66.1%. The reported out-of-fold MAE against independently estimated ADM2 targets is 10.0%, compared with 13.6% for the baseline.

This is the paper’s clearest evidence that intelligent data discovery can matter as much as model selection. The selected predictors encode agricultural seasonality, drought and vegetation shocks, market conditions, and urban or commercial activity—signals that a purely temporal baseline cannot represent. However, the fine-grained ground truth is not directly surveyed at the LGA level; it is estimated using Multilevel Regression and Poststratification and withheld from training. Consequently, the benchmark measures agreement with an independently produced model-based estimate rather than direct observation. The result demonstrates cross-scale predictive utility, but it does not establish that PPE recovers true household-level food insecurity without MRP-related uncertainty or bias.

SVI super-resolution

The US SVI downscaling experiment projects county-level information to ZIP Code Tabulation Areas. The baseline achieves only 11.0% mean $R^2$. Geospatial covariates increase performance to 25.6%, PDFM alone to 36.9%, and the covariate-plus-PDFM system to 37.6%. This establishes that PDFM representations contain cross-scale socioeconomic information that remains useful after county-level aggregation.

The manuscript contains an internal inconsistency concerning this benchmark. The headline contribution reports an SVI downscaling result of $R^2=37.6\%$ versus 11.0%, whereas the discussion describes an alternative covariate–PDFM configuration with $R^2=52.0\%$ and a reduction to 40.1% after adding AEF features. The supplied results do not fully reconcile these figures. The qualitative conclusion is nevertheless consistent: high-resolution physical features do not automatically improve socioeconomic downscaling. AEF can introduce high-frequency variation that is predictive of land-use or terrain but weakly aligned with sub-county socioeconomic targets, thereby increasing spurious correlation and degrading cross-scale transfer.

Spatial regression for health, vulnerability, and environmental risk

The strongest US spatial-regression result concerns 21 CDC health indicators. The manual expert baseline has mean $R^2=60.0\%$, while PDFM alone reaches 59.7% and PDFM plus AEF reaches 61.8%. The full PPE configuration, combining embeddings with Data Commons covariates and intelligent selection, reaches 76.8% with a 95% confidence interval of 76.1%–77.6%. The substantial gap between embedding-only and full-stack performance indicates that foundation models do not replace structured covariates; their value arises from complementary fusion and task-specific selection.

For FEMA risk variables, the full system achieves a nationwide mean $R^2$ of 64.9%, compared with 59.9% for the hand-curated expert baseline. The gain is heterogeneous. Intelligent selection raises the socioeconomic and composite category to 69.4%, the atmospheric and climatological category to 68.3%, and the geophysical and hydrological category to 56.2%. Individual improvements are large for some targets: avalanche risk increases from 14.7% to 68.8%, landslide risk from 56.1% to 75.5%, and tornado risk from 75.4% to 91.7%. Conversely, intelligent selection reduces performance for hurricane risk from 82.7% to 77.3% and earthquake risk from 86.1% to 79.9%. These reversals are important: autonomous feature discovery is not uniformly beneficial, and the full-suite mean can conceal target-specific regressions.

For county-level SVI spatial regression, the paper reports a baseline foundation-model configuration of 60.3% mean $R^2$, geospatial covariates at 50.2%, and a covariate-plus-PDFM configuration at 66.2%. Elsewhere, the manuscript characterizes the comparison baseline as 58.6%, corresponding to PDFM alone, and describes the improvement as 6.8 percentage points. The difference between the 60.3% and 58.6% baseline values is not explained in the provided text. Regardless of which baseline is used, the ablation supports multimodal synergy: PDFM captures latent socioeconomic structure, while explicit covariates contribute interpretable and target-aligned information.

What the results establish

Across the reported experiments, PPE’s principal empirical claim is not that a particular learner dominates all alternatives. The model families are conventional supervised estimators, and gradient boosting is selected in the Nigerian experiment. The performance advantage instead arises from co-optimizing the feature space and hypothesis space. Passive ingestion of all available variables is inferior to selecting a smaller task-specific set, while foundation embeddings are most effective when combined with structured covariates.

The results also show that the value of a modality is target-dependent. PDFM is consistently useful for socioeconomic and demographic outcomes, whereas AEF is more naturally aligned with ecological, physical, and land-use variables. Yet even this correspondence is imperfect, as illustrated by the SVI downscaling degradation and several FEMA regressions. PPE’s automation therefore functions as an empirical selection mechanism rather than a guarantee that semantically plausible features will transfer across spatial resolutions.

The system’s data-discovery component is also a substantive methodological contribution. It operationalizes signal discovery, repository retrieval, provenance ranking, schema normalization, and audit metadata as part of the predictive pipeline. In data-scarce settings, this can substitute for substantial manual search effort. The paper reports that an Ebola workflow executed 793 steps across approximately 55 minutes of sessions, demonstrating that the system is capable of conducting nontrivial end-to-end searches, although runtime, API costs, retrieval failures, and sensitivity to the underlying LLM are not comprehensively evaluated.

Limitations and open questions

PPE uses PDFM and AEF as frozen feature extractors. The experiments therefore do not test downstream fine-tuning, and it remains unclear whether task-specific adaptation would improve performance or amplify overfitting in small-sample and cross-scale regimes. The validation protocols are also not fully standardized across benchmarks: the US census-tract experiments use random 80:20 partitions, while some downscaling and epidemiological tasks use leave-one-state-out, spatial clustering, or expanding temporal windows. Random spatial partitions can preserve strong spatial autocorrelation and may overstate out-of-region generalization relative to spatial-group validation.

The anti-leakage protocol is more rigorous than ordinary automated feature engineering, but the causal-direction criterion remains dependent on LLM-mediated judgments and predefined rules. This is especially consequential for composite indices and socioeconomic outcomes, where proxies may be mathematically or institutionally entangled with targets even when their variable names differ. The FEMA results further show that intelligent selection can produce substantial target-level gains and losses, requiring uncertainty estimates and independent replication rather than reliance on aggregate means.

Finally, the epidemiological evidence is confined to one 2026 DRC outbreak and a limited number of forecast windows. Generalization across pathogens, reporting systems, mobility regimes, and outbreak geometries is therefore unresolved. The proposed combination of spatial transmission models for hotspot detection and mechanistic models for caseload allocation is plausible within the paper’s empirical framing, but it is not validated as an operational decision system.

Conclusion

PPE integrates LLM orchestration, provenance-aware geospatial data retrieval, foundation-model embeddings, leakage controls, and bounded AutoML into an end-to-end prediction system. Across health, environmental risk, food security, social vulnerability, and Ebola transmission tasks, it reports substantial gains over manual or simplified baselines, including 76.8% mean $R^2$ for 21 CDC indicators, 66.1% $R^2$ for Nigerian food-security downscaling versus 31.5%, and 83.3% Recall@10 for DRC outbreak hotspot prediction versus approximately 73%. The evidence supports the narrower claim that autonomous co-optimization of data selection and model configuration can improve heterogeneous geospatial workflows. It does not yet establish uniform robustness across spatial resolutions, causal structures, or epidemiological settings, particularly where labels are model-estimated or validation data are spatially correlated.