satellite-imagery-finder

name	satellite-imagery-finder
description	Use when the user wants satellite scenes — phrases like "Sentinel-2 imagery of X", "least cloudy scene over Y", "SAR coverage of an oil spill", "Landsat-Sentinel timeseries", "land cover classification". Encodes which AQUAVIEW collection matches which question (optical / SAR / harmonized / land cover) and the right CQL2 patterns for cloud cover and SAR polarization filtering.

Satellite imagery finder

AQUAVIEW exposes four global satellite imagery collections. Picking the wrong one wastes time. Use this decision tree:

User wants…	Collection	Why
Optical RGB / NDVI / coastal water color (10–20m)	sentinel-2-l2a	10m bands, surface reflectance, frequent revisit
Multi-temporal Landsat-Sentinel harmonized stack	hls2-l30	HLS L30 = harmonized 30m, perfect for change detection
Radar / cloud-penetrating / oil slicks / ship detection / sea ice	sentinel-1-grd	C-band SAR, works through clouds and at night
Land cover / coastline / habitat classification	esa-worldcover	Pre-classified 10m global land cover

See mission-cheatsheet.md for full mission-specific guidance (revisit, polarization modes, band definitions).

Standard query pattern

For a single best-scene lookup:

search_datasets(
  collections=["sentinel-2-l2a"],
  bbox=<bbox from user>,
  datetime="<start>/<end>",
  filter="properties.eo:cloud_cover < 20",
  filter_lang="cql2-text",
  sort_by="properties.eo:cloud_cover ASC",   # least-cloudy first
  limit=10
)

For SAR (always cloud-free, so no cloud filter):

search_datasets(
  collections=["sentinel-1-grd"],
  bbox=<bbox>,
  datetime="<start>/<end>",
  filter="properties.sar:instrument_mode = 'IW'",   # IW = Interferometric Wide (most common ocean mode)
  filter_lang="cql2-text",
  limit=10
)

For an HLS time-aligned stack:

search_datasets(
  collections=["hls2-l30"],
  bbox=<bbox>,
  datetime="<start>/<end>",
  filter="properties.eo:cloud_cover < 30",
  filter_lang="cql2-text",
  limit=50
)

Cloud cover guidance

• < 5% — pristine, rare, only useful for "best possible scene" requests
• < 10–20% — typical "good scene" threshold for most user requests
• < 30–40% — acceptable for change detection where you can mask clouds
• No filter — for SAR or for users who want the full timeseries with cloud masking handled downstream

SAR polarization quick reference

• VV — single-pol, common for ocean / wave / wind retrieval
• VH — cross-pol, sensitive to volume scattering (vegetation, ice roughness)
• VV+VH — dual-pol IW mode, most common for Sentinel-1 GRD over oceans
• HH / HH+HV — used at high latitudes (sea ice) and over very rough ocean

For oil-slick detection and ship detection, prefer dual-pol VV+VH IW scenes.

Always pair with bbox-from-region

If the user gave a region name (not coordinates), call into the bbox-from-region skill first. Don't ask Sentinel for "the Mid-Atlantic Bight" with no bbox — you'll get the entire L2A archive paginated forever.

Output formats

• Sentinel-2 L2A and HLS scenes have assets for individual band TIFFs (B02, B03, B04, B08, etc.). Tell the user to download the bands they need rather than the entire scene.
• Sentinel-1 GRD has VV / VH GeoTIFF assets and a calibration metadata XML.
• ESA WorldCover has a single map asset (classified GeoTIFF).