galamsey-unified-decisions

Hand-labeled corpus of 250 Sentinel-2 tiles over Ghana, each annotated with one of five on-orbit satellite-tasking actions plus a per-pass scalar context block (downlink budget, mission priors, structured neighbor summary). Used to train samwell/galamsey-unified-v3, a 450M unified vision-language model that picks bandwidth-aware actions per tile.

The corpus is the data half of the GalamseyWatch project - an agentic Earth-observation system for detecting illegal small-scale gold mining ("galamsey") in southwestern Ghana.

What's in here

Each row in labels.jsonl corresponds to one Sentinel-2 patch (1.28 km × 1.28 km, 10 m/pixel) over a hand-curated AOI in Ghana, with:

• Two image composites in images/<coord_id>/:
  • rgb.png - natural-color RGB (bands B4 + B3 + B2)
  • swir.png - SWIR false-color (bands B12 + B11 + B8); bright SWIR indicates exposed soil and mining disturbance
• Per-tile metadata: coord_id, lon, lat, stratum, mission_priors (AOI-specific operational note)
• Per-pass scalar context: budget_remaining_kb, budget_total_kb, prior_tiles_downlinked, cloud_cover, captured_at, tile_imagery_issue, and a neighbor_summary JSON object describing what each adjacent tile decided this pass
• A target action (the label): one of five tool names plus an optional reason

Action vocabulary

Action	Meaning
discard	skip the tile entirely (forest, water, urban, cloud, no signal)
flag_for_review	log as text only - no image downlink, cheap follow-up
request_higher_resolution	request a higher-res recapture next pass
request_neighbor_tile	fetch the adjacent tile in a given direction
downlink_now	spend the downlink budget to send this tile to ground

Class distribution (250 labeled rows)

Action	Count	%
discard	146	58 %
flag_for_review	53	21 %
downlink_now	47	19 %
request_higher_resolution	4	2 %
request_neighbor_tile	0	0 %

Discard dominates because Ghana's actual surface composition is mostly forest, water, savanna, urban, and farmland - the deterministic sampler reflects that base rate. The request_neighbor_tile class is structurally hard to elicit from naturalistic Sentinel-2 imagery and is absent from the labels (open follow-up).

How it was built

Sampling. A deterministic stratified sampler (seed = 42) over 15 hand-curated AOIs across Ghana spanning all relevant strata:

• Mining hotspots: Bibiani, Pra basin (Bogoso), Ankobra basin (Prestea), Obuasi, Asutifi
• Forest reserves: Atewa, Kakum, Bia
• Water: Lake Bosumtwi (crater lake), Lake Volta
• Urban: Accra, Kumasi
• Agricultural mosaic: Northern savanna, central farmland (cocoa belt)
• Cloud-prone edge: Coastal Axim

The sampler picks a stratum per fixed weights, an AOI within the stratum, and adds Gaussian jitter within each AOI's radius. Reproducible end-to-end via training/scripts/fetch_unified_corpus.py.

Imagery fetch. RGB + SWIR composites pulled from the DPhi SimSat simulator (Sentinel-2 imagery served as a satellite-pass simulator); sequential to be polite to the upstream service.

Labeling. Hand-labeled over multiple sessions, following the labeling protocol locked in three validation rounds (documented in docs/UNIFIED_VLM_VALIDATION.md). For each tile the labeler reads both image composites plus the scalar context, applies the disambiguation rules from the labeling system prompt (e.g. "SWIR brightness in the absence of exposed-soil patterns is more likely infrastructure, not mining", "rectilinear field patterns are agriculture, not mining"), and emits the action with a reason string.

Usage

from datasets import load_dataset
ds = load_dataset("samwell/galamsey-unified-decisions", split="train")
row = ds[0]
print(row["coord_id"], row["label"]["action"], row["context"]["budget_remaining_kb"])
# row["rgb"] and row["swir"] are PIL Images

For the train/eval splits used in samwell/galamsey-unified-v3 training (151 train, oversampled to 327, plus 99 held-out eval), see training/scripts/build_unified_v2_sft_dataset.py and training/scripts/build_expanded_eval_dataset.py in the GalamseyWatch repo.

Recurring data anomalies

Two SimSat-side defects are documented in the labels (tile_imagery_issue field is null in the JSONL because the synthesis pipeline didn't catch them - the labeler-side reasoning string flags them):

• cloud_cover = 1.403499 appears on every Coastal Axim tile in the corpus (18 tiles total). The expected range is [0, 1]; the bit-identical out-of-range value across all 18 tiles indicates a deterministic upstream pipeline bug specific to this AOI's scene definition. Mitigation: clamp cloud_cover to [0, 1] before consuming the field.
• Partial-tile black-bottom imagery appears on 5 Kakum tiles (u0126, u0148, u0192, u0212, u0235). Bottom 40-60 % of the tile is solid black; top portion is valid forest. Also AOI-specific. Mitigation: detect via pixel statistics (fraction of RGB = 0,0,0) and treat as a tile_imagery_issue.

Both are useful as out-of-distribution / data-quality test cases for downstream consumers.

Limitations

• Same-distribution. The 99-tile held-out eval is drawn from the same 15 AOIs as the train set, just at indices the train set never saw. Robustness across the rest of Ghana is not proven.
• Ghana-specific. All AOIs are in Ghana; geographies, land cover, and mining patterns elsewhere may differ.
• Class imbalance. Two of the five classes (request_higher_resolution, request_neighbor_tile) are essentially absent (4 / 0 examples respectively). Models trained on this corpus will not learn those actions without additional hand-constructed examples.
• Single labeler. Each labeling round was produced by a single labeler, not multi-rater. Inter-rater reliability is not measured.

License and citation

Released under CC-BY-SA-4.0, matching the upstream perception dataset (SmallMinesDS, Ofori-Ampofo et al., 2025).

@misc{galamsey_unified_decisions_2026,
  author = {Donkor, Samuel},
  title = {galamsey-unified-decisions: hand-labeled Sentinel-2 tiles for on-orbit satellite-tasking decisions},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/samwell/galamsey-unified-decisions}
}

Upstream Sentinel-2 imagery is © European Union / Copernicus Programme. The SimSat simulator is © DPhi Space.