Model card
What the forecast is and how it was tested.
30e31365a5780baa49d74f851486fcd41a008544.What it forecasts
Monthly depth to groundwater across the Central Valley. It is trained on 351,086 readings from 9,960 monitoring wells, Jan 2005 to Sep 2026. Months 1–24 are tested. Later months are a what-if.
A scenario sets precipitation and surface-water supply. Pumping is not an input. The model learns its effect indirectly, from crop demand, precipitation and the surface water available.
The model
A graph neural network over the well network, blended with gradient-boosted trees on the same well histories. A scenario moves the forecast by the network's difference between the scenario and a normal year.
Inputs
| Input | Source |
|---|---|
| Depth to groundwater, monthly | DWR Periodic Groundwater Level Measurements; USGS CVHM2 |
| Precipitation | PRISM Group, Oregon State University |
| Mean temperature | PRISM Group, Oregon State University |
| Vegetation greenness (NDVI) | NASA MODIS |
| Evapotranspiration | OpenET ensemble (CC BY 4.0) |
| Water-year index, % of normal | DWR Bulletin 120 / CDEC |
| CVP and SWP contract allocations | Reclamation and DWR |
| Share of permanent crops | USDA NASS Cropland Data Layer |
| Share of annual crops | USDA NASS Cropland Data Layer |
| Share fallow | USDA NASS Cropland Data Layer |
| River flow vs. normal | USGS NWIS |
| Project canal flow vs. normal | CDEC |
| Months since the well's last reading | derived from the readings |
Data credits:
- Groundwater levels: California DWR, Periodic Groundwater Level Measurements (data.cnra.ca.gov, public domain); USGS CVHM2 observations
- Precipitation and temperature: PRISM Group, Oregon State University, https://prism.oregonstate.edu, accessed 2026-10-22
- Evapotranspiration: OpenET ensemble v2.1 (https://etdata.org), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/); sampled at the wells
- Vegetation (NDVI): NASA MODIS MOD13A2 v061 (Didan, K., doi:10.5067/MODIS/MOD13A2.061), via NASA AppEEARS
- Crop cover: USDA NASS Cropland Data Layer
- River flows: USGS NWIS. Project flows and water-year indices: California DWR CDEC. Allocations: U.S. Bureau of Reclamation (CVP) and California DWR (SWP)
- GSA boundaries (i03) and groundwater basins (i08 Bulletin 118, critically overdrafted basins): California DWR, data.cnra.ca.gov, Creative Commons Attribution (http://www.opendefinition.org/licenses/cc-by)
- Ground elevation: USGS 3DEP 1/3 arc-second DEM and DWR surveyed station elevations. Data available from U.S. Geological Survey, National Geospatial Program. Recharge soils: Soil Survey Staff (2020). Gridded National Soil Survey Geographic (gNATSGO) Database for the Conterminous United States. USDA Natural Resources Conservation Service. https://nrcs.app.box.com/v/soils. Canals: (c) OpenStreetMap contributors (ODbL)
- NOAA outlook scenario: NOAA Climate Prediction Center long-lead outlooks. Disadvantaged communities: California DWR i16 (2023), CC BY. Domestic wells: DWR Well Completion Reports; dry wells: DWR Dry Well Reporting System
Accuracy
Tested on years the model never saw, given the weather and water supply that came. Typical error at a well, main aquifer (MAE; RMSE in brackets):
| Test | 1 month | 6 months | 12 months | 12 months vs. same month last year |
|---|---|---|---|---|
| Quiet years (Oct 2024–Sep 2026) | 4.2 ft (10.7 ft) | 6.2 ft (16.2 ft) | 7.0 ft (18.4 ft) | 4.4 ft vs. 4.3 ft |
| Drought (Oct 2020–Sep 2022) | 5.9 ft (14.2 ft) | 8.3 ft (17.6 ft) | 8.9 ft (20.2 ft) | 6.1 ft vs. 7.3 ft |
MAE is the average miss in feet. RMSE squares each miss before averaging, so large misses count more: it runs well above the MAE because a few percent of wells miss by tens of feet. The last column compares the model with last year's reading for the same month, on the wells that have one.
On the wells the map is drawn from
Most wells on the map were last read 3–12 months before the forecast starts. Their errors are larger. The usual score uses wells with a reading in the same month a year earlier. The last column weights the map's wells as on today's map (0 months 13%, 1-2 months 13%, 3-6 months 59%, 7-12 months 15% by months since the last reading).
| Test | Month | Usual score | Map's wells, average mix | Map's wells, today's mix |
|---|---|---|---|---|
| Quiet | 1 | 3.2 ft | 4.9 ft | 6.6 ft |
| Quiet | 3 | 3.5 ft | 5.3 ft | 5.8 ft |
| Quiet | 6 | 3.7 ft | 6.0 ft | 7.1 ft |
| Quiet | 12 | 4.0 ft | 7.1 ft | 10.3 ft |
| Drought | 1 | 4.2 ft | 6.6 ft | 7.8 ft |
| Drought | 3 | 5.0 ft | 7.1 ft | 7.4 ft |
| Drought | 6 | 5.6 ft | 7.9 ft | 8.8 ft |
| Drought | 12 | 5.6 ft | 9.2 ft | 14.0 ft |
Year 2
Months 13–23, typical error at a well:
| Test | Typical error |
|---|---|
| Quiet | 8.9 ft |
| Drought | 10.3 ft |
| Confirmation (Oct 2022–Sep 2024) | 10.2 ft |
GSA averages
The forecast from Sep 2025 against what happened, per GSA (193 GSAs; model trained through Sep 2024). Median across GSAs:
| 1 month | 6 months | 12 months | |
|---|---|---|---|
| Typical miss | 2.3 ft | 2.9 ft | 3.1 ft |
| Bias (+ = too deep) | +0.7 ft | +1.0 ft | +0.6 ft |
Drought check
With no precipitation, less surface water and the same crops, levels must go down. On 2,556 measured wells, 0.0% went the wrong way. After 12 months the drought run is 17.4 ft deeper than normal on average (13.9–21.4 ft across the 6 copies).
Within a real drought, levels can rise early. 3.0% of wells do so in the model.
Limitations
- No pumping data. The model infers pumping from crops, precipitation and surface water.
- Wells read rarely are forecast less well (see the map's wells above).
- Kettleman Plain is not forecast. The Kern banking area is less reliable.
- Shallow upper-zone (perched) wells are left out of the surface.
- Months 25 and later are a what-if, not tested.