DefendIntelligence/vessel-detection-labeled-patches

Vessel Detection Labeled Patches

Validated/confirmed satellite image patches exported from the military-boat-detection review workflow.

Contents

• images/: patch images.
• metadata.csv: one row per patch, compatible with Hugging Face image-folder metadata.
• metadata.jsonl: rich patch metadata with nested objects.
• annotations.csv: one row per vessel annotation.
• annotations.jsonl: JSONL version of the object annotations.
• labels/: YOLO-format labels. Hard negatives have empty .txt files.
• classes.txt: YOLO class names. Class 0 is boat.
• splits/: deterministic train/validation/test image lists.
• dataset_summary.json: export counts and provenance.

Selection

This export includes non-synthetic captures whose review status is reviewed or reviewed_local in data/app_index.sqlite3. Items still marked need_review are not included.

Generated at: 2026-05-01T11:49:03+00:00

Counts

• Patches: 1920
• Vessel object annotations: 1347
• Sample types: {'hard_negative': 844, 'boat': 1076}
• Source datasets: {'sentinel-1-vv': 1, 'sentinel-2-rgb': 1917, 'sentinel-2-swir': 2}
• Review statuses: {'reviewed': 301, 'reviewed_local': 1619}
• Splits: {'train': 1516, 'test': 219, 'validation': 185}
• Vessel types: {'porte-conteneur': 531, 'inconnu': 816}

Coordinates

Patch coordinates are provided as geographic bounding boxes and center points: patch_west, patch_south, patch_east, patch_north, patch_latitude, patch_longitude.

Boat coordinates are provided in annotations.csv / annotations.jsonl: boat_latitude, boat_longitude, geographic boat bbox, and pixel-space bbox on the patch: x_min, y_min, x_max, y_max, x_center, y_center, width, height, plus normalized YOLO values.

Labels

sample_type is boat for positive patches and hard_negative for confirmed negative patches. The original capture label is kept in label.