README.md

metadata

pretty_name: SEN2MS 5-band Vessel Detection
task_categories:
  - object-detection
tags:
  - remote-sensing
  - sentinel-2
  - vessel-detection
  - geospatial
  - parquet
license: mit
size_categories:
  - 1K<n<10K

sen2ms-optical

Dataset Summary

This dataset contains 5-band Sentinel-2 patches exported from SEN2MS for vessel detection.

• Images: 1161
• Bounding boxes: 1553
• Empty images: 300
• Non-empty images: 861
• Export errors: 0

Data Layout

• images.parquet: image-level manifest
• objects.parquet: object-level annotations
• images/: .npy tensors referenced by image_relpath

Image Format

• Spatial size: 128 x 128
• Channels: 5
• Band order: B04,B03,B02,B08,B11
• Image file format: .npy

Visual Preview

Three example rows are shown below: samples with ships, samples without ships, and ship samples with bounding-box overlays.

Splits

• train: 929 images
• val: 232 images

Geographic Coverage

• DEN: 464 images
• EXT: 100 images
• USA: 597 images

Schema

images.parquet columns

• image_id
• image_relpath
• height
• width
• channels
• band_order
• num_boxes
• has_objects
• country
• format
• task
• split

objects.parquet columns

• image_id
• bbox_idx
• class_id
• class_name
• x_center
• y_center
• bbox_width
• bbox_height
• x_min
• y_min
• x_max
• y_max
• img_width
• img_height

Intended Use

This dataset is intended for vessel detection experiments on optical remote sensing imagery.

Limitations

• Bounding boxes are derived from the exported YOLO labels.
• Empty images are included in the manifest.
• Review geographic balance and annotation quality before benchmarking.