# TORRENT generic tools

Standalone utilities for working with (and extending) the TORRENT dataset.
They talk only to public agency services — USGS NWIS, USGS NLDI, and the Iowa
State MT MRMS archive — and never require the authors' build environment.

Requirements: Python ≥3.9, `pandas`, `requests`; `download_forcing.py` also
needs `rasterio` (GDAL with the GRIB driver), `shapely`, `numpy`.

| Tool | Purpose |
|---|---|
| `find_gauges.py` | List stream gauges near an episode (by TORRENT episode ID or lat/lon) and optionally check NWIS discharge/stage availability over the event window — the candidate-generation step as a utility. |
| `download_streamflow.py` | Fetch the NWIS IV discharge record for a gauge over the standard window (begin −24 h … end +72 h) → `streamflow.csv`. |
| `get_watershed.py` | Retrieve the NLDI contributing watershed for a gauge → `watershed.geojson` (retries transient NLDI failures). |
| `download_forcing.py` | Download 2-min MRMS PrecipRate for a window, clip to a watershed, write the basin-mean series → `mrms_2min_precip_basin_mean.csv`. Use it to add forcing to the 115 light packages. |
| `build_testbed.py` | Compose the above into a candidate testbed package (watershed + streamflow + summary) for any episode–gauge pair. |

## Typical session

```bash
# which gauges could observe this episode?
python find_gauges.py --episode FF_2025_07_TX_ep002 --radius 50 --check-iv

# build a candidate package for one of them
python build_testbed.py --episode FF_2025_07_TX_ep002 --gauge 08165500 --outdir ./my_testbeds

# add forcing (also works for the released light packages)
python download_forcing.py \
    --watershed ./my_testbeds/FF_2025_07_TX_ep002/08165500/watershed.geojson \
    --begin 2025-07-04T02:00 --end 2025-07-04T17:41 --cache ./mrms_cache \
    --out ./my_testbeds/FF_2025_07_TX_ep002/08165500/mrms_2min_precip_basin_mean.csv
```

Notes
- Basin-mean values from `download_forcing.py` agree with the released series to
  within a few percent (mask-rasterization details differ slightly from the
  release pipeline).
- Packages built this way are **candidates**: apply the screening criteria in
  `code/pipeline` (paper Sect. 3.4) before comparing them with benchmark
  testbeds.
- NLDI throttles around 800 requests/hour; most failures are transient — retry.