Skip to contents

GeoTessera

R interface to Tessera geofoundation model embeddings.

Overview

The GeoTessera package provides access to 128-channel dense representation maps at 10m resolution derived from Sentinel-1 and Sentinel-2 satellite imagery via the Tessera foundation model.

Installation 

# Install from GitHub (once published)
# remotes::install_github("lassa_sentinel/GeoTessera")

# Or install from local source
install.packages("path/to/GeoTessera", repos = NULL, type = "source")

Quick Start 

library(GeoTessera)

# Create a GeoTessera client
gt <- geotessera()

# Get tiles for a region (London)
tiles <- gt$get_tiles(
  bbox = c(-0.2, 51.4, 0.1, 51.6),
  year = 2024
)

# Export as GeoTIFFs
gt$export_embedding_geotiffs(
  tiles = tiles,
  output_dir = "london_tiles"
)

Key Features

  • Download Tessera tiles for geographic bounding boxes
  • Export as GeoTIFF, NumPy arrays, or Zarr archives
  • Sample embeddings at specific point locations
  • Visualize coverage globally or for specific regions
  • Country-level data access support
  • SHA256 hash verification for data integrity

Main Classes

GeoTessera

The main interface for downloading and exporting embeddings.

gt <- geotessera()

# Count available tiles
count <- gt$embeddings_count(bbox = c(-0.2, 51.4, 0.1, 51.6), year = 2024)

# Download a single tile
tile <- gt$download_tile(lon = 0.1, lat = 51.5, year = 2024)

# Fetch embedding data
result <- gt$fetch_embedding(lon = 0.1, lat = 51.5, year = 2024)
embedding <- result$embedding  # 3D array (height, width, 128 channels)

# Sample at specific points
points <- data.frame(lon = c(0.1, 0.2), lat = c(51.5, 51.6))
samples <- gt$sample_embeddings_at_points(points, year = 2024)

Registry

Manages tile metadata and downloads.

registry <- Registry$new()

# Get available years
years <- registry$get_available_years()

# Count tiles by year
counts <- registry$get_tile_counts_by_year()

# Load tiles for a region
tiles <- registry$load_tiles_for_region(
  bbox = c(-0.2, 51.4, 0.1, 51.6),
  year = 2024
)

Tile

Format-agnostic tile abstraction.

# Create from GeoTIFF
tile <- tile_from_geotiff("path/to/grid_0.15_51.55.tif")

# Load embedding data
embedding <- tile$load_embedding()

# Sample at a point
values <- tile$sample_at_point(lon = 0.15, lat = 51.55)

Visualization 

# Visualize global coverage
visualize_global_coverage(gt, output_path = "coverage.png", year = 2024)

# Create RGB mosaic from specific bands
tiff_files <- list.files("tiles/", pattern = "\\.tif$", full.names = TRUE)
create_rgb_mosaic(tiff_files, "mosaic_rgb.tif", bands = c(30, 60, 90))

# Create PCA visualization
create_pca_mosaic(tiff_files, "mosaic_pca.tif", n_components = 3)

Country Support 

# Get bounding box for a country
bbox <- get_country_bbox("United Kingdom")

# Find which country contains a point
country <- find_country_for_point(lon = -0.1, lat = 51.5)

Data Organization

Tessera data is organized hierarchically:

  1. Blocks (5° x 5° degrees): Registry organization units
  2. Tiles (0.1° x 0.1° degrees): Individual embedding files
  3. Pixels (10m resolution): Within each tile

Each tile contains 128 embedding channels at 1111 x 1111 pixel resolution.

Supported Formats

  • NPY: Native numpy format with separate scales files
  • GeoTIFF: Georeferenced raster format with compression
  • Zarr: Chunked array format (experimental)

Dependencies

Core dependencies: - R6: Object-oriented programming - sf: Spatial data handling - terra: Raster operations - arrow: Parquet file support - httr2: HTTP requests - cli: Progress bars and messages

Optional for visualization: - ggplot2: Plotting - rnaturalearth: Country boundaries

License

MIT License

Acknowledgments

This is an R port of the Python geotessera package.