NASA Space Apps 2026, challenge Adapting farms with NASA data. Two demo regions: Pinal County, Arizona (irrigated, aquifer under stress) and East Kazakhstan (rain-fed steppe, wheat-wheat-black fallow).
For one field, Furrow compares rotations on profit (as a P10–P90 range), soil loss, soil carbon and pumped water, replayed over 22 real seasons of NASA POWER weather. It lays out the trade-offs. The farmer and the agronomist make the decision.
ML (Prithvi-EO-2.0 crop map, SMAP downscaling) lives in a separate repository. Its contract is in data/ml/README.md.
uv venv --python 3.12 .venv && uv pip install --python .venv numpy pymoo h3 pillow pytest
.venv/bin/python pipeline/fetch.py # NASA POWER + SoilGrids per H3 cell, ~25 min, cached in data/raw
.venv/bin/python pipeline/fetch.py --smap # SMAP L4 root-zone moisture from NASA GIBS, ~1 min, data/raw/*.smap.json
EARTHDATA_TOKEN=... .venv/bin/python pipeline/fetch.py --imerg # GPM IMERG monthly, ~3 min, data/raw/*.imerg.json
.venv/bin/python pipeline/build.py # simulations, climate twin, NSGA-II -> web/public/data/*.json
.venv/bin/python pipeline/model.py # model self-check
.venv/bin/python -m pytest pipeline # priority numbers, SMAP check, ML contract
cd web && npm install && npm run dev # http://localhost:5173npm run build makes a static site in web/dist, with no backend.
| Part | File | Status |
|---|---|---|
| Regions, H3 grid | pipeline/regions.py |
done |
| NASA POWER daily weather, SoilGrids topsoil | pipeline/fetch.py |
runs, no auth |
| SMAP L4 root-zone moisture (check of the water balance) | pipeline/fetch.py --smap, build.py smap_check() |
runs, no auth |
| GPM IMERG Final monthly rain (second rain source) | pipeline/fetch.py --imerg, build.py imerg_check() |
runs, Earthdata token |
| HLS NDVI, SMAP, IMERG, GRACE trend, SRTM slope, NEX-GDDP-CMIP6 deltas | pipeline/gee_layers.py |
written, needs an Earth Engine project |
| Crop parameters, prices, strategies, N rates, market depth | pipeline/crops.py |
assumptions, for an agronomist to replace |
| Farmer priorities: water, market, nitrogen, crop prices | build.py farm(), web/src/App.tsx Priorities |
done |
| Water balance + yield, wind and water erosion, RothC | pipeline/model.py |
done |
| Time machine, climate twin, NSGA-II, carbon, water | pipeline/build.py |
done |
| Map with NASA GIBS layers, strategy cards, cockpit, assistant | web/src |
done |
| ML layers | separate repo → data/ml/<region>.json |
contract only |
- Water and yield: daily FAO-56 bucket (1.2 m), Hargreaves ET0, FAO-33 yield response (Ky), heat stress at
flowering, rotation credits (after a legume +10%, cereal after cereal −8%), snow trapped by stubble.
AquaCrop-OSPy can replace it per crop; the interfaces are
simulate()andrun_strategy(). - Wind erosion: RWEQ-style index (erodible fraction × gust excess cubed × bare × dry). Water erosion: RUSLE with R from annual rain, K from texture, fixed LS until SRTM is wired. Both compare rotations; they are not absolute measurements.
- Soil carbon: RothC-26.3, monthly, initialised from SoilGrids SOC. The pools start near equilibrium, so read the difference between strategies, not the absolute trend.
- Uncertainty: every number is a spread over 22 weather years and every rotation start year. The climate twin shifts the record by ΔT/ΔP; the Earth Engine script computes the NEX-GDDP-CMIP6 SSP2-4.5 and SSP5-8.5 deltas per cell.
- Carbon credits: RothC difference × 44/12, 20% buffer, $15–35/t. Real payment needs measured samples and a registry methodology (for example Verra VM0042).
- Farmer priorities: four numbers per strategy, weighted by the farmer's sliders. Profit is the median at the
chosen crop prices (±30%, revenue moves, costs stay). Water is pumped mm (Arizona) or the share of crop water
demand not met (Kazakhstan). Market is the share of revenue from crops with a deep market. Nitrogen is mineral N
bought, kg/ha a year, with 30 kg credited after a legume. N rates and market depth are assumptions in
crops.py. - SMAP check: the model's root-zone water at the reference field against SMAP L4 root-zone moisture on the next day, 2015–2024. SMAP values come from NASA GIBS images decoded through GIBS's published colour table (0.003 m³/m³ bins), so no Earthdata login is needed. Correlation is reported raw and with the seasonal cycle removed (about 0.6 and 0.55 in both regions). It checks the model; it does not drive it.
- GPM IMERG: IMERG Final monthly totals (GPM_3IMERGM v07, GES DISC OPeNDAP) under the cells of each POWER box.
The reference field is rerun with POWER's daily rain scaled to IMERG's monthly totals; the climate section shows
both. Headline numbers stay on POWER. At the reference fields IMERG sees 14% more rain in Pinal County and 12%
more in East Kazakhstan (up to 30% in other East Kazakhstan boxes); the strategy ranking does not change. The token is read from
EARTHDATA_TOKENand never written to the repo. The GIBS IMERG image on the map is a rate snapshot, not a daily total, so it is used only as a picture. - GRACE-FO: on the map through NASA GIBS. A 300 km mascon cannot see a farm; the Earth Engine script gives the cell trend in cm/yr so pumping savings sit on the same scale.
SMAP pixels are 9 km: the app never calls them your field's moisture. Prices and costs are round assumptions. The four strategies per region are hypotheses for an agronomist to test, not recommendations.