Venezuela Earthquake 2026 Landcover

Description

This dataset contains 216 manually annotated very high resolution (0.5 m) satellite image tiles covering La Guaira, Caracas, and Catia La Mar, Venezuela. The imagery was acquired by the SkySat and Pelican constellations on 26–27 June 2026, two to three days after the earthquake of 24 June 2026. Each 512 × 512 pixel tile was annotated for semantic segmentation into five land-cover classes: building, road, vegetation, bare_soil, and water. The release includes RGB image tiles, annotation masks, per-tile provenance and image-quality records, machine-readable metadata, and a written annotation protocol. Pixel value 0 represents unannotated regions and should be treated as an ignore value during training and evaluation.

Steps to reproduce

1. Access the post-event SkySat and Pelican imagery from the Planet Crisis Response Program collection for the 24 June 2026 Venezuela earthquake. Use the three-band visual asset from the scenes covering La Guaira, Caracas, and Catia La Mar. 2. Read each source scene using its one-eighth-resolution overview. Divide the overview into blocks corresponding to a 512 × 512 pixel grid at full resolution. 3. Calculate edge density for each block as the sum of the mean absolute first differences along the horizontal and vertical image axes. Rank the blocks in descending order to prioritise structurally complex built-up areas. 4. Retrieve the highest-ranked candidate blocks from the source imagery at full resolution. Retain no more than 45 tiles from each source scene. 5. Screen candidate tiles using the following criteria: valid-pixel fraction ≥ 0.92, image intensity variability σ ≥ 13, Laplacian variance ≥ 35, and mean brightness between 25 and 235. Reject cloud-contaminated candidates where mean brightness > 205 and σ < 28. Tiles with more than 30% opaque cloud cover are excluded during visual review. 6. Manually annotate the retained 512 × 512 pixel tiles in Roboflow using polygon delineation. Assign five land-cover classes: bare_soil, building, road, vegetation, and water. Follow the rules provided in annotation_protocol.md. 7. Export the annotation masks using pixel values 1–5 for the five classes. Reserve value 0 for unannotated pixels. Verify that all mask values fall within the range 0–5 and record per-tile provenance and image-quality information.

Institutions

• Binus University

  Jakarta, Jakarta

Categories

Licence