← Nicholas Gunner

NASA FINESST · ROSES F.5 · Earth Science

Adapting Geospatial Foundation Models to Working Farms

Federated, multi-source ground truth for agroecosystem science — turning NASA's Prithvi foundation model into something that actually works on real, heterogeneous farms.

Carbon Cycle & Ecosystems New Observing Strategies Surface Biology & Geology NASA Acres

Future Investigator: Nick Gunner · CAIR Lab, Cornell AgriTech  ·  Advisors: Dr. Yu Jiang, Dr. Katie Gold, Dr. Sara Emery, Dr. Terry Bates  ·  Draft for advisor review

How to read this: follow the spine from the problem down to the payoff. The three colored cards in the middle are the engine of the proposal — the objectives. Click any card to open its full detail, hypothesis, and methods.

The logic

The engine — three objectives

Each objective feeds the next

Get the right labels cheaply → pool what's learned across farms without sharing data → use the adapted model to do real agroecosystem science.

↻ It's a loop The science in O3 reveals where the model is still uncertain — which sends O1 back out to collect the next, smartest labels.

Three years, three publications

Each phase yields at least one publication, mapping directly onto the three objectives and their hypotheses.

Year 1 · Stand it up

Pipeline + active learning

Stand up the multi-source ground-truth pipeline (robots + human mobile) on Every.Farm; fine-tune Prithvi baselines; build and test the cost-aware active-learning agent; establish cover-crop replicate plots with variable-rate seeding.

Pub 1: Cost-aware, multi-source active ground-truthing for foundation-model adaptation
Tests H1
Year 2 · Federate

Multi-farm, multi-task tuning

Deploy federated fine-tuning of Prithvi across all three sites and crops; benchmark against single-site, single-task, and centrally-pooled models; characterize robustness to cross-site (non-IID) heterogeneity.

Pub 2: Federated multi-farm/multi-task adaptation of a geospatial foundation model
Tests H2
Year 3 · Discover

Science + scale to satellite

Apply adapted models to agroecosystem science: flagship SGH test in polycultures via spectral diversity, plus transfer to grape/apple disease, nutrient, and yield; scale signals to satellite (HLS / SBG-class); release the open framework.

Pub 3: Biodiversity–function science and cross-task transfer from a federated geoFM
Tests H3

Why this is feasible — what's already in hand

This proposal builds on infrastructure, partnerships, and a farm network that already exist.

✓

A multi-source labeling platform

Every.Farm — built with Cornell AgriTech — already collects human (mobile) and sensor data, with the orchestration, geolocation, and provenance backbone.

✓

Field robots, operating now

Our lab currently runs several autonomous robotic data-collection units at CLEREL and Cornell AgriTech (Geneva). The embodied sensing channel is established, not hypothetical.

✓

Heterogeneous, multi-crop access

Established collaborations provide three farms spanning polycultures, grapes, and apples — plus my 10-acre field for the designed polyculture experiment.

✓

Experimental throughput

Year-1 acquisition of a tractor, roll-crimper, and no-till drill with variable-rate seed hoppers enables hundreds of designed replicate plots.

✓

Open methods & models

Prithvi and HLS are open; parameter-efficient and federated fine-tuning of geoFMs are demonstrated in the 2025 literature; spectral diversity as a biodiversity proxy is established across biomes.

✓

Regional precedent

Strong regional track record for specialty-crop disease sensing with NASA imagery (e.g., Gold Lab, Cornell AgriTech).

Key references (selected starting set — to be expanded & verified)
  1. Jakubik, J., et al. (2023). Foundation Models for Generalist Geospatial Artificial Intelligence (Prithvi). arXiv:2310.18660.
  2. Prithvi-EO-2.0 (2024). A Versatile Multi-Temporal Foundation Model for Earth Observation Applications. arXiv:2412.02732.
  3. Parameter-Efficient Fine-Tuning for Geospatial Foundation Models (2025). arXiv:2504.17397.
  4. Federated fine-tuning of vision / remote-sensing foundation models (2025). Privacy-preserving and uncertainty-weighted federated approaches for Earth-observation tasks.
  5. Wang, R., et al. (2022). Plant beta-diversity across biomes captured by imaging spectroscopy. Nature Communications, 13, 2839.
  6. Cawse-Nicholson, K., et al. (2021). NASA's Surface Biology and Geology designated observable. Remote Sensing of Environment, 257, 112349.
  7. Brooker, R.W., et al. (2021). Facilitation and biodiversity–ecosystem function relationships in crop production systems. Journal of Ecology, 109(5), 2054–2067.
  8. Bertness, M.D., Callaway, R. (1994). Positive interactions in communities. Trends in Ecology & Evolution, 9(5), 191–193.
  9. Li, C., et al. (2023). Intercrop overyielding weakened by high inputs: a global meta-analysis. Agriculture, Ecosystems & Environment, 343.
  10. Gold, K.M., et al. (2020). Hyperspectral measurements enable pre-symptomatic detection and differentiation of plant disease. Remote Sensing, 12, 286.
  11. Le Moigne, J., et al. (2022). AIST: Combining New Observing Strategies and Analytics Frameworks to Build Earth System Digital Twins. IGARSS 2022 (IEEE).
  12. Reichstein, M., et al. (2019). Deep learning and process understanding for data-driven Earth system science. Nature, 566, 195–204.
  13. Spatially explicit active learning for crop-type mapping from satellite image time series (2024). PMC11014375.