OlmoEarth-v1-Chesapeake-Bay-Nutrient-Loads

Annual and monthly water-quality loads (Total Nitrogen, Total Phosphorus, Suspended Sediment) at 121 USGS Nontidal Network monitoring stations across the Chesapeake Bay watershed (2014–2023). Built for use with OlmoEarth-v1-Base embeddings + classical regression, with optional fine-tuning support via quantile-binned classes.

14,166 monthly records · 1,208 annual records · 121 stations · 118 watershed polygons · 6 sub-basins · 4 equal-frequency classes

v4 (April 2026) structural fixes:

• Replaced bbox AOIs with USGS NLDI contributing-area polygons (one per station, 118/121 with valid upstream basin)
• Re-binned categories using equal-frequency quartiles per metric (was watershed-relative quantiles → produced majority-class bias)
• Added predefined spatial_fold (leave-one-sub-basin-out) and random_fold columns for honest cross-validation

---

Quick Start

Two formats are provided for two upload paths:

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

Upload one metric at a time (TN, TP, or Sediment). Studio auto-detects sample_category, sample_number, sample_true_false.

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

Option B — Pipeline / rslearn (rslearn/*_features.geojson)

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

import json
with open("rslearn/tn_features.geojson") as f:
    annotations = json.load(f)
with open("rslearn/annotation_task_features.geojson") as f:
    tasks = json.load(f)

---

Files

Path	Records	Purpose
data.parquet	14,166	Monthly wide-format master (equal-freq classes + cv_fold cols)
data_annual.parquet	1,208	Annual wide-format (doc-spec primary)
stations.geojson	121	NTN station points + summary stats
watersheds.geojson	118	NLDI contributing-area polygons per station (replaces bbox AOIs)
predefined_splits/spatial_loo.csv	121	Per-station spatial_fold (0-5) + random_fold for CV
tabular/loads_wide.csv	14,166	Monthly CSV mirror
tabular/loads_wide_annual.csv	1,208	Annual CSV mirror
tabular/loads_long.csv	39,699	Long format (one row per station-month-metric)
rslearn/{tn,tp,sediment}_features.geojson	~13K each	Classification labels (oe_labels.category 0–3), monthly
rslearn/{tn,tp,sediment}_regression.geojson	~13K each	Regression labels (oe_labels.value), monthly
rslearn/{tn,tp,sediment}_features_annual.geojson	~1.1K each	Annual classification
rslearn/{tn,tp,sediment}_regression_annual.geojson	~1.1K each	Annual regression
rslearn/annotation_task_features.geojson	6 polygons	Sub-basin AOI bboxes
studio/import_{tn,tp,sediment}.{geojson,json}	~13K each	Studio Web UI format, monthly
studio/import_{tn,tp,sediment}_annual.{geojson,json}	~1.1K each	Studio Web UI format, annual
audit_output/ASSESSMENT.md	—	Criteria audit

---

Schemas

rslearn classification (rslearn/*_features.geojson)

{
  "type": "Feature",
  "geometry": {"type": "Point", "coordinates": [lon, lat]},
  "properties": {
    "oe_labels": {"category": 0},
    "oe_start_time": "2014-01-01 00:00:00+00:00",
    "oe_end_time": "2014-01-31 00:00:00+00:00",
    "oe_annotations_task_id": "uuid",
    "STAID": "01491000",
    "metric": "tn_load",
    "year": 2014,
    "month": 1,
    "value": 88.1,
    "flow_normalized": true,
    "class_name": "reference_baseline"
  }
}

Studio import schema (studio/import_*.json)

{
  "type": "Feature",
  "geometry": {"type": "Point", "coordinates": [lon, lat]},
  "properties": {
    "task_name": "patuxent_choptank",
    "observation_time": "2014-01-01T00:00:00Z",
    "sample_category": "reference_baseline",
    "sample_number": 88.1,
    "sample_true_false": true
  }
}

Field mapping:

• sample_category → 4-category label (reference_baseline, low_stress, moderate, high_stress)
• sample_number → continuous load value (regression target)
• sample_true_false → flow-normalized flag

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

---

Categories

Quantile-binned per metric. reference_baseline (bottom 10%) serves as the negative reference.

ID	Tag	Range	Count	%
0	reference_baseline	bottom 10%	3,982	10.0%
1	low_stress	10–25%	5,949	15.0%
2	moderate	25–75%	19,848	50.0%
3	high_stress	top 25%	9,920	25.0%

Counts shown are monthly across all three metrics combined.

---

AOI Sub-basins

Sub-basin	Longitude	Stations
susquehanna	-78.5 to -74.8	60
potomac	-79.5 to -76.6	26
james	-81.3 to -75.9	14
rappahannock	-78.3 to -76.2	6
york	-78.4 to -76.2	6
patuxent_choptank	-77.4 to -75.5	9

Six sub-basin bbox AOIs covering the Chesapeake watershed.

---

Studio Configuration

Setting	Value
Data sources	Sentinel-1 (VV/VH) + Sentinel-2 (12 bands)
Time range	2014-01-01 to 2023-12-31
Granularity	Monthly (primary) / Annual (doc-spec)
Train / val split	Spatial leave-out by sub-basin recommended
sample_number range	Set to 0 – 100,000,000 when uploading multiple metrics

Upload tip: Studio auto-locks sample_number range based on the first upload. Loosen to 0–100M before uploading a second metric, OR use one Studio project per metric.

---

Caveats

• Class boundaries are watershed-relative quantiles, not absolute thresholds. A high_stress load in one basin may differ in absolute magnitude from another.
• Bbox AOIs, not actual watershed polygons. Adequate for windowing imagery; replace with HUC-8 shapes for production-grade analysis.
• Reference_baseline is value-based, not site-based — marks low-load observations, not pristine reference watersheds.
• Monthly granularity is a scaled format; the case study specifies annual scale, so data_annual.parquet is the primary doc-compliant deliverable.
• Flow-normalized vs raw loads — flag in *_flow_norm columns indicates which.

---

Sources

All data is public domain.

• USGS Chesapeake Bay Loads & Trends — Annual Loads Table and Monthly Loads Table (2024 release, filtered to 2014–2023)
• USGS NWIS — Station metadata via waterservices.usgs.gov
• Chesapeake Bay Program Data Hub — datahub.chesapeakebay.net (NTN reference)

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-Potomac-Sewage-Spill-2026 — sister dataset (event reference)

Citation

@dataset{qi2026chesbay,
  author    = {Qi, Ziming and BAI Group},
  title     = {OlmoEarth-v1-Chesapeake-Bay-Nutrient-Loads: Analysis-Ready Loads Dataset for GeoAI Modeling},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/2imi9/OlmoEarth-v1-Chesapeake-Bay-Nutrient-Loads}
}

License

Apache 2.0. USGS source data are public domain.