Why Generate Embeddings?

Presto geospatial embeddings provide a compressed representation of Earth Observation data, enabling more efficient mapping and analysis. Embeddings are generated by using the Presto encoder to compress location information, optical imagery (Sentinel-2), radar imagery (Sentinel-1), climatology data (ERA5), and elevation data (SRTM) over the course of a year. Each embedding contains 128 features representing a single 10m² pixel on Earth. Embeddings can be used in place of raw Earth Observation data for various machine-learning tasks, such as classification, clustering, and anomaly detection. In order to use embeddings they must first be generated for the time frame and area of interest. We provide an open-source pipeline to do this using Vertex AI and Google Earth Engine.

Generating Embeddings

The embedding generation pipeline consists of two steps:

1. Deploying Presto to Google Cloud Vertex AI in Google Colab
   
2. Using the ee.Model.fromVertexAi function in Google Earth Engine
   1_Generate_Embeddings

Cost Considerations

While Google Colab and Google Earth Engine (non-commercial) are free to use, deploying Presto to Vertex AI and running large-scale inference incurs cloud costs.
The current cost formula is: $5.37 - $10.14 / 1000 km².

The cost formula is a range due to scaling in Google Cloud. Small regions will tend toward the higher range, and larger regions, such as all of Togo, will tend towards the lower range.

Case Study: Generating Embeddings for Togo

To test our pipeline, we generated embeddings for all of Togo (56,785 km²) for the time frame of March 2019 - March 2020. We will use labels collected in March 2020 as part of a Rapid Response crop mapping effort in Togo [2]. The embedding generation took 16 hours and cost $313.40. The final Togo asset size was 128.8 GB.

The asset can be accessed here:
https://code.earthengine.google.com/?asset=users/izvonkov/Togo/Presto_embeddings_v2025_06_19

It's important to note that Google Earth Engine has a default asset storage limit of 250 GB.
For larger assets, we recommend exploring Cloud GeoTiff-Backed Earth Engine Assets.

Sanity Checking Embeddings

To verify that the embeddings contain meaningful information we cluster the embeddings using k-means and visually compare the clusters to an existing land cover map (WorldCover 2020).
This requires just 9 lines of code: 2_Kmeans_clustering_embeddings

// Region of interest
var roi = ee.FeatureCollection("FAO/GAUL/2015/level2").filter(ee.Filter.eq('ADM0_NAME', 'Togo'));
Map.centerObject(roi, 7)

// Load and cluster embeddings
var embeddings = ee.Image("users/izvonkov/Togo/Presto_embeddings_v2025_06_19")
var training = embeddings.sample({region: roi, scale: 10, numPixels: 10000});
var trainedClusterer = ee.Clusterer.wekaKMeans(7).train(training);
var result = embeddings.cluster(trainedClusterer);
Map.addLayer(result.randomVisualizer(), {}, 'clusters');

// Display WorldCover
var WorldCover = ee.ImageCollection('ESA/WorldCover/v200').first().clip(roi)
Map.addLayer(WorldCover, {bands: ['Map']}, 'WorldCover')
            

We see that the clusters visually align well with the land cover map. In particular, the green embedding cluster correlates very closely with built-up areas in the land cover map. This suggests that the embeddings capture meaningful information about the land cover in Togo and can be used for further mapping and analysis.

Why use embeddings for mapping?

Embeddings are a compressed version of processed Earth Observation data. So using embeddings for mapping will be more computationally efficient compared to using traditional Earth Observation data. Using embeddings also makes it possible to skip data processing (such as cloud processing) altogether.

How does it work?

A traditional mapping approach usually involves the following:

1. Loading Earth Observation data for the region of interest.
2. Loading labeled points representing the classes to be mapped.
3. Creating a training dataset by combining labeled points with Earth Observation data.
4. Training a classifier (e.g. Random Forest) on the training dataset.
5. Classifying all the Earth Observation data using the trained classifier.

1. Loading embeddings for the region of interest.
2. Loading labeled points representing the classes to be mapped.
3. Creating a training dataset by combining labeled points with embeddings.
4. Training a classifier (e.g. Random Forest) on the training dataset.
5. Classifying all the embeddings using the trained classifier.

Mapping Cropland in Togo

// 1. Load embeddings for region of interest
var roi = ee.FeatureCollection("FAO/GAUL/2015/level2").filter("ADM0_NAME=='Togo'");
Map.centerObject(roi, 7)
var embeddings = ee.Image("users/izvonkov/Togo/Presto_embeddings_v2025_06_19")

// 2. Load Togo points
var points = ee.FeatureCollection("users/izvonkov/Togo/points_2019")
var trainingPoints = points.filter(ee.Filter.eq("subset", "training"))

// 3. Create training dataset (training points + embeddings)
var trainingSet = embeddings.sampleRegions(trainingPoints, ["is_crop"], 10 )

// 4. Train a classifier (100 trees)
var model = ee.Classifier.smileRandomForest(100).setOutputMode('probability')
var trainedModel = model.train(trainingSet, 'is_crop', embeddings.bandNames());

// 5. Classify embeddings using trained model
var croplandPreds = embeddings.classify(trainedModel).clip(roi)
var croplandMap = croplandPreds.gte(0.7).rename("map_crop")

// 6. Display cropland maps
var classVis = {min: 0, max: 1.0, palette: ['yellow', 'green']}
Map.addLayer(croplandMap, classVis, 'Presto Embeddings Based Cropland');
            

The resulting cropland map is generated in less than 5 seconds:

Alternative Embeddings: Google DeepMind

Google DeepMind recently released their own pixel-level embeddings.
Here are some similarities and differences between the two:

ee.Image("users/izvonkov/Togo/Presto_embeddings_v2025_06_19")

ee.ImageCollection("GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL")
    .filterDate("2021-01-01", "2022-01-01")
    .mosaic()
    .clip(roi)

uint16

double

We create an alternative cropland map with Google DeepMind embeddings by simply substituting:

ee.Image("users/izvonkov/Togo/Presto_embeddings_v2025_06_19")

ee.ImageCollection("GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL")
    .filterDate("2021-01-01", "2022-01-01")
    .mosaic()
    .clip(roi)

Visual Assessment and Comparison

One way of comparing embeddings is visually assessing and comparing the generated cropland maps. We display the embedding-generated maps below, in addition to the best available cropland maps in Togo.

Green represents crops and yellow is everything else. The displayed maps are:

1. a cropland map derived from the WorldCover land cover map [3],
2. the GLAD cropland map [4],
3. the Presto embedding-based cropland map,
4. the DeepMind embedding-based cropland map [5].

1. WorldCover

2. GLAD

3. Presto
Embeddings

4. DeepMind
Embeddings

Observations from examining the maps from a country level:

• Presto and DeepMind embedding-based have a similar cropland amount and distribution
• Presto and DeepMind embedding-based show more crops than WorldCover
• The South of Togo is where all maps have differences

1 / 5

Satellite view in Google Earth Engine

2 / 5

WorldCover Cropland

3 / 5

GLAD Cropland

4 / 5

Presto Embeddings-based Cropland

5 / 5

DeepMind Embeddings-based Cropland

❮ ❯

• All maps with the exception of GLAD correctly identify built-up areas as non-crop (yellow)
• Embedding-based maps appear to classify more cropland (green) correctly
• The DeepMind embedding-based map over predicts crops in some shrub/tree areas
• Presto embedding-based map visually accurately identifies many of the tree areas as non-crop

Metrics Assessment and Comparison

To get a more comprehensive understanding of the map quality we conducted an accuracy assessment using the Togo test set from CropHarvest. It's important to note that Presto was benchmarked against CropHarvest.

Script for assessment: 6_Togo_cropland_embeddings_metrics

Concluding Thoughts

We show that using Presto embeddings for cropland mapping is simple, quick, and effective. Our results indicate the potential for using embeddings for more effective mapping in different regions and for different use cases.

[1] Lightweight, Pre-trained Transformers for Remote Sensing Timeseries. Gabriel Tseng, Ruben Cartuyvels, Ivan Zvonkov, Mirali Purohit, David Rolnick, Hannah Kerner, 2023. Neural Information Processing Systems (NeurIPS), Workshop on Tackling climate change with machine learning (link) ↵

[2] Rapid Response Crop Maps in Data Sparse Regions. Hannah Kerner, Gabriel Tseng, Inbal Becker-Reshef, Catherine Nakalembe, Brian Barker, Blake Munshell, Madhava Paliyam, Mehdi Hosseini, 2020. Knowledge Discovery and Data (KDD), Workshop on Humanitarian Mapping (link) ↵

[3] ESA WorldCover 10 m 2021 v200. Daniele Zanaga, Ruben Van De Kerchove, Dirk Daems, Wanda De Keersmaecker, Carsten Brockmann, Grit Kirches, Jan Wevers, Oliver Cartus, Maurizio Santoro, Steffen Fritz, Myroslava Lesiv, Martin Herold, N-E Tsendbazar, Panpan Xu, Fabrizio Ramoino, Olivier Arino, Zenodo (link) ↵

[4] Global maps of cropland extent and change show accelerated cropland expansion in the twenty-first century. Peter Potapov, Svetlana Turubanova, Matthew C Hansen, Alexandra Tyukavina, Viviana Zalles, Ahmad Khan, Xiao-Peng Song, Amy Pickens, Quan Shen, Jocelyn Cortez, Nature Food (link) ↵

[5] Better Maps Faster: Low-shot Classification and Regression with Embedding Fields. Chris Brown, Michal Kazmierski, Valerie Pasquarella, William J Rucklidge, Masha Samsikova, Chenhui Zhang, Evan Shelhamer, Sophia Alj, Emily Schechter, Sean Askay, Alexis Boukouvalas, AGU (link) ↵