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
cs.AI, cs.LG
2026-08-27
PPE auto-builds geospatial models from a text query. CDC health mean R² goes from 60.0% to 76.8%; Nigeria food-security downscaling from 31.5% to 66.1% R².
Building a map that a relief agency can actually act on still takes weeks. The covariates live in Data Commons, Earth Engine, national portals, and satellite products. Someone has to pick proxies, align administrative units, fuse them with geospatial foundation embeddings such as PDFM and AlphaEarth, block spatial leakage, then tune a model. One epidemiological nowcasting workflow in this paper unpacks into more than 700 steps. Off-the-shelf AutoML and science agents mostly run on tables that are already clean. They do not discover planetary data, align heterogeneous embeddings, or parameterize a transmission model.
Google Research packages that pipeline as the Planetary Prediction Engine (PPE) under Google Earth AI. A natural-language query comes in; a trained geospatial model and a report come out. The Institut National de Recherche Biomedicale (INRB) in the DRC collaborated on nowcasting the 2026 Bundibugyo Ebola outbreak.
Three stages, each with its own LLM orchestrator at temperature 0. Stages pass DataFrame and GeoJSON artifacts as opaque handles. Raw tables never enter the prompt, which keeps context windows intact and keeps data selection isolated from training.
Task type is locked in stage 1: spatial regression, super-resolution downscaling, spatial transmission, or epidemiological nowcasting. The search covers Ridge / Lasso / ElasticNet, HistBoost, XGBoost, and Keras MLPs. An overfitting guard inspects n, p/n, and spatial grouping before training; high risk biases the search toward strongly regularized linear models and shallow trees. If validation collapses, the agent is allowed one more conservative restart. Nigeria food security has 30 states and a 330-dim PDFM vector. That guard is written for exactly this small-n, high-p regime.
Ablations sit on four rungs: a human expert pipeline or public SOTA, covariates only, embeddings, and the full stack (covariates + embeddings + intelligent selection).
| Task | Metric | PPE | Baseline |
| 21 US CDC health indicators, census tract | Mean R² | 76.8% | Expert pipeline 60.0% |
| FEMA National Risk Index, 20 targets | Mean R² | 64.9% | Expert 59.9% |
| US SVI, county | Mean R² | 66.2% | PDFM only 58.6% |
| Nigeria food security, state → LGA | R² (state LOSO) | 66.1% | Macro + interpolation 31.5% |
| 2026 DRC Bundibugyo Ebola, 519 health zones | Recall@10 | 83.3% | Bayesian SOTA 73% |
The CDC numbers show that embeddings alone do not carry the gain. PDFM scores 59.7%, PDFM+AlphaEarth 61.8%, sitting on the expert baseline. Adding Data Commons covariates and intelligent selection jumps to 76.8%. The lift is the auto-discovered statistical covariates.
Nigeria FCG is trained on 30 states over 40 months and scored on 581 LGAs. Vegetation features alone take state-level R² from 31.5% to 60.1%; the full stack reaches 66.1%. LGA labels are independent MRP estimates, held out during training; MAE there is 10.0% versus 13.6% for the baseline. The winning recipe was gradient boosting with temporal macros, nighttime lights, and vegetation. Extra signals hurt. SVI downscaling from county to ZIP/ZCTA reaches 37.6% R² against an 11.0% macro-interpolation baseline. That is a large relative jump and still a modest absolute score, which is what cross-boundary heterogeneity looks like.
Ebola is a one-week rolling forecast. Across five evaluation weeks, 15 of 18 newly invaded health zones landed in the top 10. Covariates alone reach 77.8%; the full stack 83.3%, with a 95% CI of [60.8, 94.2]. Eighteen events make that interval wide. In the appendix, spatial transmission regression ranks new hotspots better (test top-10 accuracy 0.8), while SEIR nowcasting tracks caseload better (test RMSE 0.92 versus 6.08). A full nowcasting run took 793 steps, about 55 minutes across three sessions.
This is a natural-language on-ramp for people who cannot assemble Earth Engine pipelines and leakage-safe validation by hand. In humanitarian and public-health work, a multi-week modeling cycle is the failure mode. What transfers is the selection protocol, the Feature Gate, and the overfitting guard, not a particular estimator. PDFM is still pre-general-availability plus research use; AlphaEarth is public. The whole machine is hard to copy outside Google. The recipe that can travel is: lock the task type, retrieve literature-backed proxies, and treat tables and frozen embeddings as separate objects.
For teams already doing geospatial ML, this is incremental automation. The model zoo stops at Ridge, XGBoost, and MLP. The claim is joint search over data and models, plus leakage control, not a new predictor.
The paper lists four: embeddings stay frozen; adding high-resolution AlphaEarth onto a PDFM stack in SVI downscaling drops R² from 52.0% to 40.1%, so fine satellite features can act as noise; the causal-direction filter is not formally verified; epidemiology is a single outbreak.
A few more points do not fully land. CDC and FEMA spatial regression follow the public SOTA protocol with an 80:20 random split on 84k census tracts. Spatial autocorrelation inflates those scores; the paper implements Spatial Group Split and does not use it on the headline table. The Ebola Recall@10 interval is wide enough that it nearly touches the baseline, so +10.3 points should not be read as a stable win. The abstract's "doubled accuracy" for Nigeria is state-level LOSO R²; the LGA number on offer is MAE. Do not cite them as one result. The Feature Gate is an LLM judging four rules, and bidirectional targets will slip through. Maps Platform Insights is GCP commercial billing, so a from-scratch reimplementation will be missing a slice.