Added AlphaEarth datasource

[robmarkcole]

• Added a new AlphaEarth datasource backed by the public Source Cooperative STAC GeoParquet index, with direct materialization support, optional dequantization, and minimal new dependencies.
• Documented and supported use of use_all_bands_in_order_of_band_set_idx so AlphaEarth embeddings can be consumed in model configs without repeating band names.
• Added datasource-aware default nodata handling during materialization. AlphaEarth now defaults to -2.0 in dequantized mode and -128 in raw mode when nodata_vals is not specified.
• Updated AlphaEarth docs to recommend ingest: false because annual TIFFs are large, and clarified dequantized embedding semantics, including that exact unit norm is only approximate after quantization and that optional L2 re-normalization may be useful for cosine-based workflows.
• Added unit and integration tests covering generated band names, model-input band resolution, datasource default nodata behavior, and AlphaEarth item lookup, ingest, and raster reads
• Extended raster band-set configuration so consecutive band names can be generated from num_bands with a custom prefix, start index, and zero padding, allowing AlphaEarth’s 64 bands to be configured without listing them explicitly.

// Before
{
  "type": "raster",
  "band_sets": [{
    "dtype": "float32",
    "bands": [
      "A00", "A01", "A02", "A03", "A04", "A05", "A06", "A07",
      "A08", "A09", "A10", "A11", "A12", "A13", "A14", "A15",
      "A16", "A17", "A18", "A19", "A20", "A21", "A22", "A23",
      "A24", "A25", "A26", "A27", "A28", "A29", "A30", "A31",
      "A32", "A33", "A34", "A35", "A36", "A37", "A38", "A39",
      "A40", "A41", "A42", "A43", "A44", "A45", "A46", "A47",
      "A48", "A49", "A50", "A51", "A52", "A53", "A54", "A55",
      "A56", "A57", "A58", "A59", "A60", "A61", "A62", "A63"
    ]
  }]
}

// After
{
  "type": "raster",
  "band_sets": [{
    "dtype": "float32",
    "num_bands": 64,
    "band_prefix": "A",
    "band_zero_pad": 2
  }]
}

Example config

        "alphaearth": {
            "type": "raster",
            "band_sets": [
                {
                    "dtype": "float32",
                    "num_bands": 64,
                    "band_prefix": "A",
                    "band_zero_pad": 2
                }
            ],
            "data_source": {
                "class_path": "rslearn.data_sources.alphaearth.AlphaEarth",
                "init_args": {
                    "metadata_cache_dir": "cache/alphaearth",
                    "index_url": "https://data.source.coop/tge-labs/aef/v1/annual/aef_index_stac_geoparquet.parquet",
                    "apply_dequantization": true
                },
                "ingest": false
            }
        }
    },

Materialised raster

[favyen2]

The existing data source https://github.com/allenai/rslearn/blob/master/rslearn/data_sources/aws_google_satellite_embedding_v1.py also pulls from the Source Cooperative data but using it from their AWS S3 bucket, is there a substantial difference with this one? For the bands, since we have num_bands already I think it would make sense to align the band names with the ones automatically generated by num_bands for these embedding datasets instead of adding the band_prefix/band_zero_pad options since the latter would add more things that users would need to figure out how to configure in order to use the data source.

[robmarkcole]

I wasn't aware of the existing data source, will switch to using that (although no immediate plans to use it again). Thanks