Alphaearth Foundations Core

The Google Satellite Embedding dataset provides 64-dimensional geospatial

embeddings representing the semantic characteristics of Earth's surface at

10-meter resolution.

Specifications

• Dataset ID: GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL
• Bands: 64-dimensional float embeddings, named A00 through A63.
• Unit-Vector Guarantee: The bands are pre-normalized unit vectors (values

in range [-1, 1]).

See the catalog page for more details:

Satellite Embedding V1 (Annual)

Quantization & De-quantization (Under the Hood)

To store global high-dimensional embeddings efficiently, the dataset is internally quantized to 8-bit signed integers.

If you are accessing raw signed 8-bit integers in the raw collection

GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL_RAW or direct from the GCS bucket

gs://alphaearth_foundations/satellite_embedding/v1/annual/, the non-linear

de-quantization mapping used to reconstruct the native float values is:

$$v_{de\_quant} = \text{sign}(v_{raw}) \cdot \left(\frac{v_{raw}}{127.5}\right)^2$$

> [!NOTE] If reading raw COGs from the GCS bucket, mask the reserved no-data

> value -128 before applying this formula; otherwise it dequantizes to a value

> outside the valid [-1, 1] range.

def de_quantize(raw_image):
  # Mask the reserved no-data value -128 before de-quantizing
  raw_image = raw_image.updateMask(raw_image.neq(-128))
  return raw_image.float() \
                  .divide(127.5) \
                  .pow(2) \
                  .multiply(raw_image.signum())