rolfsimoes/samples_cerrado_aef

AlphaEarth embeddings of the Cerrado samples

AlphaEarth Foundations annual embeddings (64 dimensions, 10 m) at the points of two sample sets of the e-sensing/samples_cerrado dataset:

• labelled_samples_cerrado.parquet: 110,044 labelled samples, 11 classes, label years 2018, 2020, 2022 and 2024;
• validation_data_2018.parquet: 24,000 validation points of 2018, 8 classes.

The embeddings are used to compare global embeddings with regional self-supervised encoders of the Cerrado in the sitsfm experiments.

Files

File	Samples	Reading	Image year
alphaearth_labelled_centre_y.parquet	labelled	centre	Y
alphaearth_labelled_mean_y.parquet	labelled	mean	Y
alphaearth_labelled_centre_ym1.parquet	labelled	centre	Y-1
alphaearth_labelled_mean_ym1.parquet	labelled	mean	Y-1
alphaearth_validation2018_centre_y.parquet	validation	centre	2018
alphaearth_validation2018_mean_y.parquet	validation	mean	2018
alphaearth_validation2018_centre_ym1.parquet	validation	centre	2017
alphaearth_validation2018_mean_ym1.parquet	validation	mean	2017

Y is the label year of the sample. A "2017/2018" sample is labelled with its final year, 2018, so its Y-1 image is 2017.

• centre: the vector of the 10 m pixel that contains the point.
• mean: the mean of the 3 x 3 window around that pixel (30 m x 30 m, the size of a Landsat pixel), renormalised to unit length. Pixels with no data are left out (4 labelled and 5 validation points).

Structure

Each file is a sits tibble, as sits_get_data() returns it from sits_cube(source = "GOOGLE", collection = "ALPHAEARTH"): columns longitude, latitude, start_date, end_date, label, cube and time_series. The time series has one row (Index = 1 January of the image year) and the bands A00 to A63.

Rows follow the order of the source file, so row i is sample i of labelled_samples_cerrado.parquet or validation_data_2018.parquet.

library(sits)
samples <- sits_from_hf(
  repo = "e-sensing/samples_cerrado_aef",
  file = "alphaearth_labelled_centre_y.parquet",
  type = "dataset"
)

How the files were made

The windows were read in Google Earth Engine from GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL, on the native UTM grid of each tile (analysis/alphaearth_extract.py in sitsfm), and converted with analysis/alphaearth_samples.R. The centre vectors are equal to the values sits_get_data() returns for the same points. A point covered by two overlapping tiles keeps the complete window, then the pixel closest to the point.

The DATASET_VERSION of the images is 1.1 for 2017 and 1.0 for the other years read (2018, 2019, 2021, 2023).

Licence and attribution

The labels come from e-sensing/samples_cerrado (CC BY-NC 4.0). The embeddings come from the AlphaEarth Foundations Satellite Embedding dataset (CC BY 4.0):

"The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind."