mjf-su/VIS-SAR-Sat-Dataset

VIS–SAR Satellite Dataset

Spatially- and temporally-aligned visible (optical) ↔ SAR satellite image patches for cross-modal representation learning (CLIP-style SAR↔visible alignment).

Contents

• ~3.68M paired patches (this is a snapshot of an in-progress 10M build).
• shards/ — WebDataset .tar shards. Each sample key K has three members:
  • K.vis.npy — visible/optical, uint8 [3, 224, 224] (3 bands)
  • K.sar.npy — SAR, uint8 [2, 224, 224] (VV, VH)
  • K.json — per-sample metadata (AOI, biome, region, lat/lon, acquisition times, CRS, transform, SAR scaling, cloud cover, Δt hours)
• index.parquet — random-access offset index, one row per sample: key, shard, vis_offset, vis_size, sar_offset, sar_size, ground_tile_id, aoi_id, biome, region. A worker can seek+read a single sample without scanning the shard.
• manifests/ — per-shard metadata sidecars. manifest.parquet — consolidated metadata.

Decoding (uint8 → physical)

Values are quantized from normalized ranges to uint8 (4× smaller than float32, retrieval-neutral):

• Visible: physical reflectance clipped to [0, 30000] → [0,1] → uint8. Decode: x/255.0 gives normalized [0,1].
• SAR: linear amplitude → dB, clipped to [-40, 5] → [0,1] → uint8. Decode: x/255.0 gives normalized [0,1]; dB = x/255*45 - 40.

Training

Directly consumable by the multi-modal CLIP training pipeline: point it at the shard directory and the offset index —

--shards_dir  <download>/shards
--index_path  <download>/index.parquet

The loader (tar_dataset) auto-detects the uint8 dtype and rescales to [0,1] on read (no re-normalization needed). Use ground_tile_id for leak-free train/val splits (group by tile so the same ground location never spans splits).

Notes

• Patches are 224×224 at a common spatial grid (VIS and SAR resampled to the same grid).
• SAR penetrates cloud/night, so VIS↔SAR pairs are matched within a Δt tolerance (see dt_hours in metadata).