aws glo30 CopernicusGLO30

rslearn.data_sources.aws_glo30.CopernicusGLO30

Copernicus GLO-30 DEM (30m) elevation data, served directly from the original public AWS S3 bucket at s3://copernicus-dem-30m (no credentials required).

The data is split into 1x1-degree Cloud-Optimized GeoTIFF tiles covering global land areas, so elevation is read on demand at materialize time and there is no ingest step. Tile paths are constructed deterministically from latitude/longitude; the bucket's tileList.txt index is downloaded and cached so that we only create items for tiles that actually exist.

This data source provides the raw elevation band only. Slope and aspect are derived on the fly with rslearn.train.transforms.terrain.ElevationToSlopeAspect, matching how other rslearn data sources materialize raw bands and leave derived quantities to transforms.

Configuration

{
  "class_path": "rslearn.data_sources.aws_glo30.CopernicusGLO30",
  "init_args": {
    // Directory to cache the tileList.txt index.
    "metadata_cache_dir": "cache/glo30",
    // Band name to use when the layer config is unavailable (default: elevation).
    "band_name": "elevation"
  },
  // Recommended query configuration.
  "query_config": {
    "space_mode": "MOSAIC",
    "max_matches": 1
  },
  "ingest": false
}

Available Bands

The data source should be configured with a single band set containing a single band, the elevation band (named elevation by default), in meters. The data type should be float32.

{
  "type": "raster",
  "band_sets": [{
    "bands": ["elevation"],
    "dtype": "float32"
  }],
  "data_source": {
    "class_path": "rslearn.data_sources.aws_glo30.CopernicusGLO30",
    "init_args": {"metadata_cache_dir": "cache/glo30"},
    "query_config": {
      "space_mode": "MOSAIC",
      "max_matches": 1
    },
    "ingest": false
  }
}

Items from this data source do not come with a time range (the DEM is static).

Deriving Slope and Aspect

Use ElevationToSlopeAspect in the model's transform pipeline to turn the elevation band into elevation/slope/aspect channels. Because the transform runs after materialization, the image is already in the window's projection, so pass the window's pixel size in meters:

{
  "class_path": "rslearn.train.transforms.terrain.ElevationToSlopeAspect",
  "init_args": {
    // The window resolution in meters.
    "pixel_size_m": 10,
    "input_selector": "elevation",
    // Any subset/order of elevation, slope, and aspect.
    "bands": ["elevation", "slope", "aspect"]
  }
}

The emitted bands are:

• elevation — raw DEM value in meters
• slope — terrain slope in degrees [0, 90)
• aspect — compass direction of steepest descent in degrees [0, 360), -1 for flat

Notes

• Only land tiles are published. Because we filter items against tileList.txt, windows over ocean simply match fewer (or no) tiles instead of failing to read.
• This data source supports direct materialization only. Setting "ingest": true raises an error.
• Slope and aspect use central differences in the interior and one-sided differences on the image border, so the outermost pixel ring of each window is less accurate.
• Nodata is assumed to be NaN elevation, which produces NaN slope and aspect for that pixel and its four orthogonally adjacent pixels.
• See also rslearn.data_sources.planetary_computer.CopDemGlo30 for the same dataset served via Microsoft Planetary Computer's STAC API.