OlmoEarth-v1-FT-Karst-Groundwater-Base

Fine-tuning dataset for OlmoEarth-v1-Base to detect karst terrain features in Pennsylvania from Sentinel-1 SAR backscatter and Sentinel-2 optical imagery.

47,020 labeled points · 4 classes · ~50,000 km² · 3 AOI regions

---

Quick Start

import json

with open("annotation_features.geojson") as f:
    annotations = json.load(f)

with open("annotation_task_features.geojson") as f:
    tasks = json.load(f)

To use with OlmoEarth Studio, upload both GeoJSON files in your project's annotation step. The pipeline pulls Sentinel-1/2 imagery automatically.

---

Files

File	Records	Purpose
annotation_features.geojson	47,020 points	Canonical annotations (OlmoEarth oe_* schema) — use this for Studio upload
annotation_task_features.geojson	3 polygons	AOI region boundaries
data.parquet	47,020 rows	Flat tabular preview of annotations (for HF Data Studio / quick analysis)

---

Schema

Each annotation feature:

{
  "type": "Feature",
  "geometry": {"type": "Point", "coordinates": [lon, lat]},
  "properties": {
    "oe_labels": {"category": 0},
    "oe_start_time": "2020-01-01 00:00:00+00:00",
    "oe_end_time": "2025-12-31 00:00:00+00:00",
    "oe_annotations_task_id": "uuid",
    "confidence": 0.83,
    "quality_flags": ["on_carbonate", "dense_cluster", "source_pageode"]
  }
}

---

Classes

ID	Tag	Count	%
0	surface_depression	16,295	34.66%
1	sinkhole	3,810	8.10%
2	surface_mine	1,916	4.07%
3	other	24,999	53.17%

other represents non-karst terrain inside the AOI (sampled from non-carbonate bedrock). 191 cave records are included in the sinkhole class and tagged with merged_from_cave: true.

---

AOI Regions

Region	Longitude	Area	Annotations
pa_karst_east	-76.2 to -75.0	~12,000 km²	16,009
pa_karst_central	-77.4 to -76.2	~16,000 km²	14,841
pa_karst_west	-78.7 to -77.4	~16,000 km²	16,170

Covers the full Pennsylvania karst belt across 3 task regions, each within OlmoEarth Studio's 20,000 km² limit.

---

Per-Label Confidence

Every annotation carries a confidence score (0-1) and quality_flags list.

Signals:

Signal	Weight	Description
Bedrock alignment	0.5	Karst classes on carbonate; "other" off carbonate
Local class density	0.3	Same-class neighbors within 500m
Source provenance	0.2	PaGEODE (0.2) vs auto-generated (0.1)

Distribution:

Class	Mean	High-confidence (≥0.8)
surface_depression	0.83	73.9%
sinkhole	0.77	28.3%
surface_mine	0.76	40.3%
other	0.69	26.3%

Quality flag values: on_carbonate, off_carbonate_karst, dense_cluster, isolated, auto_generated, source_pageode, merged_from_cave.

Recommended uses:

• Filter: train only on confidence >= 0.8
• Weight: use confidence as per-sample loss weight
• Triage: prioritize manual review of low-confidence labels

---

Studio Configuration

Setting	Value
Data sources	Sentinel-1 (VV/VH) + Sentinel-2 (12 bands)
Time range	2020-01-01 to 2025-12-31
Grid size	2 km
Window resolution	10 m
Window buffer	31 px (~310 m)
Train / val split	75% / 25% spatial split (128px grid)

---

Caveats

• "Other" class is auto-generated from non-carbonate bedrock polygons; pending domain review.
• Cave merged into sinkhole — 191 records traceable via merged_from_cave: true.
• 2,768 outside-carbonate karst labels retained pending review (some may be valid non-carbonate features).
• Class imbalance — use class weighting or oversampling for sinkhole (8%) and surface_mine (4%) during training.

---

Sources

All data is public domain.

• PaGEODE Karst Features — PA DCNR (last updated 2026-02-24)
• PA Bedrock Geology — PA DCNR (DCNR_Bedrock_Geology_Pa_Geologic_Units52023)
• Satellite imagery — Sentinel-1/2 via OlmoEarth Studio (Microsoft Planetary Computer)

Related

• allenai/OlmoEarth-v1-Base
• olmoearth_projects
• rslearn
• OlmoEarth Studio docs

Citation

@dataset{qi2026karst,
  author    = {Qi, Ziming and BAI Group},
  title     = {OlmoEarth-v1-FT-Karst-Groundwater-Base: Pennsylvania Karst Features for OlmoEarth Fine-Tuning},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/2imi9/OlmoEarth-v1-FT-Karst-Groundwater-Base}
}

License

Apache 2.0. PaGEODE and PA Bedrock source data are public domain.