PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2608.26088v1cs.AIcs.LG

TL;DR

Planetary-scale geospatial prediction is bottlenecked by fragmented data and labor-intensive curation, fusion, and model selection. PPE autonomously converts natural-language queries into task-tailored predictive models by discovering multimodal data, integrating foundation-model embeddings, and optimizing model architectures. Across diverse tasks and geographies, it outperforms state-of-the-art or expert baselines, while remaining limited by frozen embeddings, spatial-resolution trade-offs, and unresolved causal-direction verification.

  • Problem

    Building high-fidelity geospatial models requires labor-intensive discovery, curation, fusion, and optimization across fragmented data ecosystems.

  • Method

    PPE translates natural-language queries into geospatial models by selecting multimodal covariates, fusing PDFM and AlphaEarth embeddings, and searching task-tailored model families with overfitting safeguards.

  • Results

    Across diverse tasks and geographies, PPE achieves expert-level or superior performance, including 76.8% mean R^2 for 21 US CDC health indicators, 66.1% versus 31.5% for Nigerian food-security downscaling, and 83.3% Recall@10 for DRC Ebola nowcasting.

  • Takeaways & Limitations

    PPE lowers the technical barrier to planetary-scale analytics by enabling rapid, customized deployment from natural-language specifications.

  • Takeaways & Limitations

    PPE currently relies on frozen foundation-model embeddings, while adaptive feature selection and formal verification of causal-direction filters remain open challenges.

Abstract

from arXiv · show

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 $R^2$ 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 ($R^2$ 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.

1. Introduction

Planetary analytics is constrained by fragmented geospatial data and labor-intensive workflows, while existing automation lacks specialized capabilities for discovering, curating, and modeling heterogeneous data. PPE addresses this bottleneck by translating natural-language queries into autonomous, task-tailored geospatial predictions across diverse benchmarks.

  • 1. Introduction: Geospatial predictive workflows can exceed 700 steps across data selection, curation, and model optimization, requiring specialized teams to navigate fragmented data ecosystems.These demands hinder rapid deployment for socio-environmental and epidemiological tasks.
  • 1. Introduction: Existing scientific automation is typically limited to clean, pre-curated tabular datasets or software engineering tasks and cannot autonomously curate large-scale heterogeneous geospatial data.It also lacks capabilities for fusing geospatial embeddings or parameterizing mechanistic epidemiological models.
  • 1. Introduction: PPE translates natural-language queries into executed geospatial models through three stages: Intelligent Data Selection, Multimodal Dataset Curation, and Automated Model Building and Prediction.LLM orchestrators interpret queries and select tools within predefined stage-specific tool sets.
  • 1. Introduction: PPE automatically identifies predictive task types and propagates task-specific constraints to downstream data retrieval and modeling workflows.Supported paradigms include spatial regression, super-resolution, spatial transmission modeling, and epidemiological nowcasting.
  • 1. Introduction: PPE combines foundation-model embeddings with on-the-fly statistical covariates and searches diverse model families while applying leakage mitigation and overfitting safeguards.The protocol includes regularized linear models, gradient-boosted trees, extreme gradient boosting, and multilayer perceptrons.
  • 1. Introduction: 12–94% relative R^2 improvement over standard baselines is reported across tasks, including near-doubling for Nigerian food-security downscaling and a 10.3 percentage-point nowcasting gain.The benchmark spans spatial regression, super-resolution, and epidemiological nowcasting across the Global North and Global South.

2. Results

Across mechanistic nowcasting, super-resolution downscaling, and spatial regression benchmarks, PPE combines automated data processing with multimodal features and generally outperforms baselines. Results span Ebola hotspot detection, Nigerian food-security mapping, and US social, health, and environmental indicators.

  • Mechanistic Nowcasting: 83.3% Recall@10 was achieved for DRC Ebola hotspot detection using fused epidemiological signals, geospatial covariates, and PDFM embeddings.This was a 10.3 percentage-point gain over the 73% state-of-the-art baseline.
  • Super-Resolution Downscaling: 66.1% R² was achieved for Nigerian food-security downscaling, compared with 31.5% for the macro-covariate and interpolation baseline.The selected feature combination used gradient boosting with temporal macro-covariates, nighttime lights, and vegetation signals.
  • Super-Resolution Downscaling: The Nigerian model predicted food-insecurity prevalence across all 774 LGAs from ground truth available for 30 surveyed states.The comparison used ADM1 state-level information to produce ADM2 LGA-level predictions in areas without ground surveys.
  • Spatial Regression: Social Vulnerability Index performance reached 67.6% R² versus 48.2% for the comparison baseline on a standout single-target result.The reported difference corresponds to a +40.0% gain.
  • Spatial Regression: 66.2% R² was achieved by the complete PPE pipeline on the SVI benchmark, exceeding standalone PDFM embeddings at 58.6% and explicit covariates at 51.6%.The full configuration combined covariates, PDFM, AEF, and intelligent selection.

3. Discussion

Across spatial regression, super-resolution, and epidemiological nowcasting, PPE’s gains are attributed to multimodal fusion, autonomous data curation, and joint optimization of data and models. The evaluations also reveal a fine-scale noise trade-off and remaining limits around frozen embeddings, causal filters, and outbreak generalization.

  • Key insights: Multimodal fusion consistently outperforms either explicit geospatial covariates or foundation-model embeddings alone across benchmarks.The reported gains depend on feature engineering and leakage prevention.
  • Key insights: Autonomous data curation assembles task-specific covariates that rival or exceed expert-constructed sets, especially in data-scarce settings.Localized proxy discovery is highlighted for Nigeria food security and DRC Ebola.
  • Key insights: Joint optimization of curated features and model architectures is the primary driver of PPE’s performance advantage.Ablations indicate that either intelligent data curation or automated model selection alone does not achieve peak performance.
  • Spatial regression: 76.8% mean R² on 21 CDC health indicators exceeded the 60% expert-pipeline baseline, while FEMA reached 64.9% versus 60% and SVI reached 66.2% across five themes.These results show consistent performance lifts across public-health, environmental-risk, and socioeconomic domains.
  • Super-resolution downscaling: 66.1% versus 31.5% R² doubled baseline accuracy for Nigerian LGA-level food-security mapping, but adding high-resolution AlphaEarth features reduced SVI downscaling from 52.0% to 40.1% R².The SVI result indicates a noise-resolution trade-off in fine-scale administrative downscaling.
  • Limitations and future work: PPE currently uses PDFM and AlphaEarth as frozen feature extractors, while causal-direction verification and broader epidemiological validation remain open challenges.The paper specifically notes the need for regularization during fine-tuning and validation across pathogens, geographies, and surveillance infrastructures.

4. Methods

PPE autonomously discovers, prioritizes, and assembles multimodal geospatial features, then applies leakage-controlled, task-specific modeling workflows for spatial prediction and epidemiological nowcasting.

  • Intelligent Data Selection: PPE dynamically identifies, retrieves, scores, and assembles geographically and temporally relevant covariates into a join-ready multimodal DataFrame.The process begins from the user’s natural-language predictive query and labeled data constraints.
  • Intelligent Data Selection: Signal discovery distinguishes direct variables from literature-validated proxy signals before repository and open-web retrieval.Established sources include Data Commons, Google Earth Engine, and Google Maps Platform Insights; unresolved signals trigger programmatic web discovery.
  • Multimodal Dataset Curation: PPE fuses tabular geospatial covariates with pretrained foundation-model embeddings representing socio-demographic, physical, industrial, and ecological properties.The feature set combines automatically selected covariates with location-based representations from PDFM and AlphaEarth.
  • Multimodal Dataset Curation: A Feature Gate excludes target components, shared-survey variables, downstream effects, and features from after the prediction window to prevent leakage.Missingness filtering and split-isolated imputation further restrict feature construction to information available without validation or test contamination.
  • Modeling Objectives: The general training strategy minimizes a loss over observed feature-target pairs and applies the optimized model to infer predictions at target locations.This formulation underlies the prediction tasks, while spatial regression trains on observed locations and predicts missing values at the same geographic granularity.
  • Modeling Objectives: Epidemiological nowcasting predicts the seven-day increase in caseloads using pandemic-specific signals, with optimization performed across spatial regions and expanding historical time folds.The nowcasting target is the difference between future and current confirmed caseloads at each location.

5. Conclusion

PPE translates natural-language queries into geospatial predictive models and addresses the labor-intensive process of discovering, curating, and fusing heterogeneous data with suitable architectures. Across evaluated tasks, it achieves expert-level or superior performance while supporting broader access to geospatial prediction.

  • PPE translates natural-language queries into executed geospatial predictive models spanning regression, downscaling, transmission modeling, and nowcasting.
  • PPE addresses the expertise-dependent bottleneck of discovering, curating, and fusing heterogeneous geospatial data with domain-appropriate modeling architectures.
  • 76.8% mean R^2 versus 60% for the expert baseline on 21 US CDC health indicators, 66.1% versus 31.5% for Nigerian food-security downscaling, and 83.3% Recall@10 for DRC Ebola nowcasting.
  • PPE makes predictive modeling more accessible to researchers, humanitarian organizations, and policymakers without specialized engineering teams.

A. Benchmarks •

The benchmark materials cover US socioeconomic, health, and environmental-risk indicators, Nigerian food-security downscaling, and DRC Ebola nowcasting. The DRC workflow combines epidemiological, demographic, socioeconomic, spatial, and temporal processing before model fitting.

  • US benchmarks include county- and ZIP-code-level Social Vulnerability Index metrics, 21 census-tract CDC health indicators, and FEMA environmental-risk indices.
  • Nigerian food-security evaluation uses food-price, acute-food-insecurity, precipitation, vegetation, fatality, and processed-news data, training at ADM1 and evaluating at ADM2.
  • The DRC benchmark simulates the 2026 Bundibugyo Ebola outbreak across 519 health zones using official WHO surveillance data.
  • Epidemiological inputs include daily confirmed cases, deaths, contact tracing, and manually transcribed situation-report indicators, with three absent health zones added from reports.
  • Raster and point covariates are aggregated to health zones, standardized before fitting, and accompanied by log-transformed population counts.
  • Nowcasted incidence is adjusted for reporting delays, convolved with Gamma generation-time distributions, and binned weekly for one- and two-week horizons.

B. Geospatial Covariates and Foundation Models

The paper documents inventories of geospatial covariates and geospatial foundation models used to support PPE’s multimodal prediction workflow. The supplied materials identify these inventories but do not enumerate their entries.

  • The paper directs readers to Tables 9 and 10 for detailed inventories of geospatial covariates and foundation models.
  • Table 9 is the inventory of geospatial covariates used by the study.
  • Table 10 is the inventory of geospatial foundation models used by the study.

C. Intelligent Data Selection Methodology & Rubrics

PPE’s intelligent data-selection methodology uses a provenance-first prioritization rubric to rank candidate datasets. The supplied appendix materials identify a five-dimension scoring rubric and its highest-weighted criteria.

  • C. Intelligent Data Selection Methodology & Rubrics: The Intelligent Data Selection Methodology & Rubrics section describes a scoring approach for ranking candidate datasets.
  • C. Intelligent Data Selection Methodology & Rubrics: The Provenance-First Prioritization sub-stage uses a five-dimension scoring rubric.
  • C. Intelligent Data Selection Methodology & Rubrics: Provenance and license compliance receive the highest rubric weight, at 5x.
  • C. Intelligent Data Selection Methodology & Rubrics: The rubric prioritizes covariates from openly licensed sources.
  • C. Intelligent Data Selection Methodology & Rubrics: The rubric also prioritizes institutionally authoritative sources.
  • C. Intelligent Data Selection Methodology & Rubrics: Table 11 details the five-dimension scoring rubric used in Provenance-First Prioritization.

D.1. Autonomous AI Agents for Science

The paper situates PPE within autonomous LLM-based scientific systems and presents intelligent data selection as a research-grounded pipeline for heterogeneous geospatial data retrieval.

  • Prior autonomous LLM systems use reasoning, planning, self-refinement, tool use, and multi-agent architectures for scientific discovery and data science.
  • The paper’s title identifies PPE as an autonomous geospatial prediction system combining intelligent data selection with foundation model embeddings.
  • PPE adds a research-grounded intelligent data selection pipeline that dynamically fetches data from heterogeneous repositories.

D.2. Geospatial Foundation Models

PPE combines geospatial foundation-model embeddings with task-specific prediction workflows addressing socioeconomic, environmental, and epidemiological settings. Its outbreak application targets surveillance lag and anisotropic transmission in spatial nowcasting.

  • Geospatial Foundation Models: PDFM produces 330-dimensional to 512-dimensional embeddings for global administrative boundaries, representing latent socioeconomic, demographic, mobility, and health patterns.
  • Geospatial Foundation Models: These foundation-model embeddings serve as compact spatial features for downstream regression tasks.
  • Geospatial Prediction Tasks: Geospatial prediction workflows include spatial regression, super-resolution downscaling, epidemiological nowcasting, and spatial transmission modeling.
  • Epidemiological Nowcasting: Real-time epidemic tracking must address surveillance lag and anisotropic transmission along mobility corridors rather than uniform spatial diffusion.
  • Epidemiological Nowcasting: The Bundibugyo outbreak task predicts which currently uninfected health zones will experience new cases within the following week, evaluated with Recall@10.
  • Super-Resolution Downscaling: Nigeria downscaling estimates state-level WFP HungerMap FCG and rCSI indicators at LGA-level 1km² hexagonal grids for targeted aid distribution.

SEIR vs Spatial Transmission Regression

The paper compares Spatial Transmission Regression with mechanistic SEIR Bayesian modeling and evaluates FEMA risk predictions through feature ablations, including intelligent data selection.

  • Model Comparison: Table 12 compares Spatial Transmission Regression with SEIR Bayesian Modeling.
  • Model Comparison: Spatial Transmission Regression achieves Test Top-10 accuracy of 0.8 for static vulnerability ranking across regional health zones in Ituri and Nord-Kivu.
  • Evaluation Metrics: The comparison evaluates static vulnerability ranking with Top-10 accuracy and dynamic caseload trajectory tracking with RMSE.
  • FEMA Ablations: The ablation comparison first removes intelligent Data Commons selection and then evaluates modified feature suites and metrics after adding discovered intelligent data selection.
  • FEMA Ablations: FEMA ablations examine PDFM, AEF, and Data Commons features across socioeconomic, composite, atmospheric, climatological, geophysical, and hydrological targets.
Loading 2608.26088v1…