prucodes/ap-groundwater-fusion
Domain:
environment and energygeospatialCreator:
pru
Host:
Andhra Pradesh groundwater fusion prototype — APWRIMS sensor history fused with NASA GRACE-DA, CHIRPS and TerraClimate satellite context. Modelled mandal-level nowcasts with explicit uncertainty. Prototype, not official.
# Mandal-Level Groundwater Fusion Layer for Andhra Pradesh
This repository is a data-first proof of concept for combining APWRIMS-format
groundwater observations with regional NASA/NDMC GRACE-DA and climate context.
The active Phase 0 output separates measured observations, modelled nowcasts,
unreleased forecasts, external signals, data completeness and neutral monitoring
flags.
APWRIMS-format observations are the measured historical source. Their current
browser-session sample has authorization pending and is not described as an
official export. NASA/NDMC GRACE-DA is supporting regional model-assimilated
context and must not be interpreted as a direct mandal-level groundwater-depth
measurement. Public boundary/name datasets are prototype-only until replaced by
verified official boundaries and identifiers.
The `app/` directory contains the static prototype UI. It reads generated JSON from `app/data/` and does not add a backend, database, or official-result claims.
### Active data pipeline (Phase 0 / contract V2)
The dashboard's V2 records are produced by
`phase3_levels/build_real_app_data.py`, which delegates to the fail-closed Phase
0 publisher and combines:
- the **APWRIMS depth history** (`phase3_levels/apwrims/apwrims_gw_history.csv`) — a **browser-session research sample (authorization pending)**, not an official APWRIMS export;
- the **holdout-safe nowcast output** (`phase3_levels/outputs/mandal_nowcasts_v2.json`), whose latest target rows are excluded from fitting;
- **NASA GRACE-DA / CHIRPS / TerraClimate** district signals.
Active application artifacts are
`app/data/mandal_groundwater_records_v2.json`,
`app/data/mandal_observation_series_v2.json`,
`app/data/model_card.json`, and `app/data/dataset_manifest.json`. The manifest
contains canonical counts, validity periods, hashes and active/legacy lifecycle
status.
The temporal holdout metric evaluates lag-eligible nowcasting/gap filling; it is
not sensorless spatial accuracy. Whole-mandal estimation is …