BAIGroup/OlmoEarth-v1-Potomac-Sewage-Spill-2026

OlmoEarth-v1-Potomac-Sewage-Spill-2026

Validated, time-aligned reference dataset supporting GeoAI tracking of the 2026 Potomac River sewage spill (Glen Echo, MD — 240–300 million gallons released from the 72-inch Potomac Interceptor on January 19, 2026). Provides event timeline, hydrologic context, monitoring station locations, and AOI corridor polygons for AI2's downstream Sentinel-1/2 plume detection.

8 USGS gauges · 846 daily flow records · 4 CBP/CBF stations in AOI · 37 NHD waterbody polygons · 4-phase timeline · Dec 2025 – Mar 2026

v4 (April 2026) structural fixes:

• —Replaced bbox AOIs with 37 NHD high-resolution waterbody polygons (Potomac mainstem, Anacostia, upper Bay; all >=1 km²)
• —Filtered the broad CBP integration manifest (was 3,933 keyword-matched stations) to 4 stations with verified coordinates within the AOI
• —Added reference event logs for control windows in 2023–2024 and 2024–2025 (no spill) — enables false-positive characterization for plume detection
• —Documented the JBA jet fuel leak (Mar 23, 2026 – ongoing) as a known confounding event affecting tidal_potomac_lower and upper_chesapeake_bay AOIs from late March onward (see docs/confounding_events.md). Five entries appended to event_log.csv with event_type = confounding_external. No time-series data has been modified.

Per the case study scope: BAI prepares time-aligned reference data; the OlmoEarth team analyzes Sentinel-1/2 imagery to detect plume formation, transport, and recovery. BAI does not perform satellite processing or modeling. This is an event reference / validation dataset, not a training set.

Quick Start

Two formats are provided:

Option A — OlmoEarth Studio Web UI (studio/import_corridor.json)

Upload via Studio's "Import your own dataset" feature for visualization and embedding extraction at the corridor stations.

import json
with open("studio/import_corridor.json") as f:
    data = json.load(f)

Option B — Pipeline / rslearn (rslearn/annotation_features.geojson)

For direct use with olmoearth_run prepare_labeled_windows or any rslearn-compatible workflow.

import json, pandas as pd

flow = pd.read_parquet("data.parquet")           # daily streamflow, phase-tagged
timeline = json.load(open("timeline.json"))      # 4-phase chronology
spill = json.load(open("spill_metadata.geojson"))  # spill site + receiving waters
stns = json.load(open("monitoring_stations.geojson"))  # 8 USGS gauges

Files

Schemas

rslearn / pipeline schema (rslearn/annotation_features.geojson)

{
  "type": "Feature",
  "geometry": {"type": "Point", "coordinates": [lon, lat]},
  "properties": {
    "oe_labels": {"category": 1},
    "oe_start_time": "2026-01-19 00:00:00+00:00",
    "oe_end_time": "2026-01-24 23:59:59+00:00",
    "oe_annotations_task_id": "uuid",
    "site_no": "01646500",
    "phase": "active_release",
    "role": "near_spill"
  }
}

Category encoding: 0 = prespillbaseline, 1 = activerelease, `2` = bypassperiod, 3 = post_recovery.

Studio import schema (studio/import_corridor.json)

{
  "type": "Feature",
  "geometry": {"type": "Point", "coordinates": [lon, lat]},
  "properties": {
    "task_name": "near_spill",
    "observation_time": "2026-01-19T00:00:00Z",
    "sample_category": "active_release",
    "sample_number": 1,
    "sample_true_false": true
  }
}

Field mapping:

• —sample_category → phase label (pre_spill_baseline, active_release, bypass_period, post_recovery)
• —sample_number → phase index (0–3)
• —sample_true_false → true if not pre-spill baseline

These names are required for Studio's auto-detection of training fields.

Event Summary

Phases

AOI Corridor

Streamflow Coverage

Source: USGS NWIS daily values service (parameter 00060, statistic 00003).

Event Intelligence Layer

event_log.csv and event_log.geojson provide a day-level operational record from DC Water updates spanning January 19 – April 28, 2026. Used as an event intelligence / narrative explanation layer, not analytical input — gives temporal alignment for satellite acquisitions and operational context for plume detection signals.

Source: DC Water Potomac Interceptor Updates.

Schema: date, event_type, description, phase, source_url.

Caveats

• —Not a training dataset. Only 32 station-phase records — far below any fine-tuning threshold. Designed for validation and context.
• —AOI polygons are bbox rectangles, not actual watershed/Bay polygons. Replace with NHD/watershed shapes for production-grade analysis.
• —CBP water quality data not auto-fetched. Available manually via datahub.chesapeakebay.net; cited as a source but not bundled here.
• —One station (01651000) had no NWIS data in the event window; retained in metadata for completeness.
• —Plume detection is out of scope. AI2's OlmoEarth team performs Sentinel-1/2 analysis to produce plume maps, footprint, duration, and recovery timeline.

Known Confounding Events

A separate environmental incident overlaps with this dataset's window:

• —JBA jet fuel leak — discovered March 23, 2026 on Piscataway Creek (Joint Base Andrews); active containment and cleanup ongoing through May 2026+. Piscataway Creek drains into the Potomac within tidal_potomac_lower AOI and downstream effects could reach upper_chesapeake_bay.

Any satellite-detected anomaly in those AOIs from late March 2026 onward could be from either the sewage spill recovery or the JBA leak. We cannot separate the two from this dataset alone.

See `docs/confounding_events.md` for full details and reasoning. Five entries added to event_log.csv with event_type = "confounding_external". No time-series data has been modified — this is documentation only.

Sources

All data is public domain.

• —USGS NWIS — waterdata.usgs.gov
• —DC Water — Potomac Interceptor Collapse
• —Virginia Department of Health — Potomac Sewage Spill News
• —Maryland DEP / MDE — Potomac Interceptor Sewer Overflow
• —DC DOEE — Potomac Interceptor Update and FAQs
• —Chesapeake Bay Program Data Hub — datahub.chesapeakebay.net

Related

• —allenai/OlmoEarth-v1-Base
• —olmoearth_projects
• —rslearn
• —OlmoEarth Studio docs
• —2imi9/OlmoEarth-v1-FT-Karst-Groundwater-Base — sister dataset (PA karst FT)
• —2imi9/OlmoEarth-v1-Chesapeake-Bay-Nutrient-Loads — sister dataset (Chesapeake nutrient loads)

Citation

@dataset{qi2026potomac_spill,
  author    = {Qi, Ziming and BAI Group},
  title     = {OlmoEarth-v1-Potomac-Sewage-Spill-2026: Event Reference Dataset for Plume Tracking},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/2imi9/OlmoEarth-v1-Potomac-Sewage-Spill-2026}
}

License

Apache 2.0. USGS source data and agency reports are public domain.