A series of experiments exploring Google DeepMind's AlphaEarth satellite embedding dataset to understand patterns of environmental change through multi-dimensional embedding space.
- Embedding Distance Grid Interface – Interactive web application for exploring globally sampled satellite embeddings as grids in 1D, 2D, and 3D. Change number of samples, reduce dimensions with UMAP or t-SNE algorithm, view satellite image or false color.
- Embedding Topography Interactive Visualization – Interactive web application for exploring latent space. Globally sampled embeddings reduced to 3d with UMAP or t-SNE.
- False Color Embedding Globe – Interactive 3D globe visualization of satellite embeddings
- Earth Engine App – Interactive web app for real-time analysis
Generates a global dataset of classified satellite embeddings and reduces them to 3d, 2d, and 1d using tsne and umap.
- Labels global sampling by land classification using Copernicus land cover and geographical region
- Samples 1000–5000 locations globally, extract embedding vectors
- Applies UMAP or t-SNE dimensionality reduction to project embeddings to 3d, 2d, and 1d
- Writes the data to geojson in data/
- html + js read the data and produce interactive data viz using three.js showing animation from world to embedding space.
Inputs: Google Satellite Embedding V1 (2024), Copernicus Global Land Cover, UN shapefile
Outputs: GeoJSON files with embeddings coords, HTML point cloud visualization
Compares embedding spaces across years and generates large-scale similarity grids.
- global-similarity-grid.ipynb: Extends to global scale using pre-computed UMAP coordinates from dimension-reduction
- us-similarity-grid.ipynb: Creates US-wide grid of sample points, extracts embeddings for multiple years (2021–2024), computes Euclidean distances between year pairs
- gen-us-samples.ipynb: Generates random coordinate samples for the US grid
- Outputs satellite patches and metadata for each location
Inputs: Random or stratified coordinates, Google Satellite Embeddings (multi-year)
Outputs: csv/json, satellite patch image files, static grid images
Interactive web application for exploring dimensionality-reduced satellite embeddings in 1D, 2D, and 3D with live false color band selection.
- generate_web_data.ipynb: Preprocesses data from dimension-reduction experiments
- Uses Hungarian algorithm to arrange embeddings optimally in 1D, 2D and 3D grid layouts
- Downloads 62 false color band combinations per location (consecutive triplets: A00-A01-A02 through A61-A62-A63) and true color RGB satellite imagery
- Exports data tiles (
tiles/{index:04d}/{index}_A{b1}_A{b2}_A{b3}.png) and geojson files with grid coordinates
- Interactive Visualization (index.html, styles.css, app.js):
- Toggle between 1D / 2D / 3D arrangements, UMAP / t-SNE, true color satellite imagery / false color embedding bands
- Band selector: Vertical interactive slider to select 3 consecutive bands from 64 available (A00-A63) with live RGB updates
- Modal image viewer, hover popup metadata interaction, n samples slider, etc
Inputs: Classified embeddings from dimension-reduction/ , Google Satellite Embeddings, Sentinel-2/Landsat-8 imagery
Outputs: Tile directories with 62 band combinations + RGB per location, geojson, interactive HTML visualization
Investigates the relationship between geographic proximity and embedding similarity to identify anomalous location pairs.
- distance-similarity-correlation.ipynb: Comprehensive analysis of geographic distance vs embedding similarity patterns
- Samples 1000 locations from 5000 classified embeddings, computes pairwise distances and similarities for 5000 random pairs
- Identifies Far but Similar and Near but Different outlier pairs
- Creates scatter plots showing the negative correlation between distance and similarity
Inputs: 5000 classified embeddings from dimension-reduction/... Outputs: PNG visualizations (scatter/heatmap with/without anomalies), embedding dimension comparison plots, detailed statistical analysis of anomalous pairs
Generates web tiles and interactive globe for global embedding exploration.
- gen-tiles.ipynb: Exports global embedding data as GeoTIFFs (4 regional exports) from Earth Engine to Google Drive, then locally:
- Merges 4 GeoTIFFs into single global dataset
- Converts from 32-bit to 8-bit and EPSG:4326 → EPSG:3857
- Generates tile pyramid (zoom 0–8) in XYZ tile structure
- Creates interactive globe.html with Three.js or Cesium viewer
Inputs: Google Satellite Embedding V1 (2024), global geometry
Outputs: Tile pyramid (xyz structure for zoom 0–8), globe HTML viewer
Creates a cartographic projection centered on two geographic points using D3.js.
- Implements d3.js
geoTwoPointEquidistantprojection with two anchor points from near locations in embedding space (Alaska and Greenland) - Interactive HTML/SVG visualization
Inputs: land-50m.json (TopoJSON land geometry)
Outputs: Interactive HTML map with two-point equidistant projection, downloadable SVG
Generates visualizations for poster/publication materials. Also includes Adobe Illustrator poster file.
Creates drawing comparing four elemental typologies: forest, water, ice, desert.
Inputs: Element GeoJSON with locations
Outputs: Embedding vectors per element, satellite/drawing comparisons
Analyzes satellite embedding changes in NYC watershed between 2017 and 2024.
- Filters embeddings to 50km watershed region
- Compares 2017 vs 2024 embeddings via K-means clustering (6, 10, 20 clusters)
- Computes per-pixel Euclidean distance between years
- Identifies and visualizes the 3 embedding bands with greatest change
- Maps results interactively
Inputs: Google Satellite Embeddings (2017, 2024), NYC watershed geometry
Outputs: Geemap interactive layers (clustering, distance, band differences)
Analyzes Greenland glacier regions (Eqip Sermia) using embedding vectors to detect temporal changes.
- Compares embeddings from 2017 and 2024 for glacier regions
- Computes per-pixel Euclidean distances in embedding space
- Visualizes distance heatmaps to identify areas of significant environmental change
- Generates reprojected TIF outputs for further geospatial analysis
Inputs: Google Satellite Embeddings (2017, 2024), glacier region coordinates
Outputs: GeoTIFF files with RGB embeddings and Euclidean distance rasters
Explores spatial patterns in the Bay Area's satellite embedding space through k-nearest neighbor networks.
- Samples locations in San Francisco and the Bay Area using both random and grid-based methods
- Constructs k-NN networks in 64D embedding space using cosine similarity
- Compares embedding distances with geographic distances to identify environmental patterns
- Analyzes edge angles in the similarity network to discover that similar conditions cluster at specific orientations (125°), suggesting topographic influence
Inputs: Google Satellite Embedding V1 Annual dataset (2024), Bay Area/SF geometry
Outputs: Folium maps with network visualizations, angle distribution histograms
Analyzes embedding patterns for Palestine region.
Inputs: Google Satellite Embeddings, Palestine geometry
Outputs: Interactive maps with embedding visualizations
Exports all 64 embedding bands as RGB GeoTIFFs for a Wisconsin region.
- Iterates through all 64 embedding bands in groups of 3 (R,G,B)
- Downloads each triple as a separate GeoTIFF at 500m and 5000m resolution
- Stores as individual band-indexed TIFF files
Inputs: Google Satellite Embedding V1 (2024), Wisconsin region boundary
Outputs: 21 GeoTIFF files (~500m resolution) + 21 files (~5000m resolution) organized by band groups
pip install -r requirements.txt- Authenticate:
ee.Authenticate() - Initialize with your project:
ee.Initialize(project="your-project")
Each directory contains standalone Jupyter notebooks. Run them sequentially.
Uses Google Earth Engine's AlphaEarth satellite embedding dataset. See individual notebooks for detailed citations.