Species Distribution Models on 'AlphaEarth' Satellite Embeddings

Fits species distribution models and maps habitat suitability at up to 10 m resolution from occurrence records alone, using the 'AlphaEarth' Foundations satellite embeddings (Brown et al. 2025) . The embeddings, 64 values per pixel per year from a geospatial foundation model, replace environmental layers, so none need to be sourced or aligned. Provides tools to format occurrence records, place pseudo-absences, train and evaluate an ensemble of machine learning models, and export habitat-suitability rasters. Sampling, model training and prediction all run on 'Google Earth Engine', which requires a free account for noncommercial use.

AlphaSDM

AlphaSDM fits species distribution models and maps habitat suitability at up to 10 m resolution, anywhere on Earth, from occurrence records alone. It models species on the embeddings of AlphaEarth, Google DeepMind's geospatial foundation model, instead of environmental layers you collect yourself, and runs every step on Google Earth Engine.

Saguaro around Tucson, Arizona: GBIF records and pseudo-absences (left), and the fitted habitat-suitability map at 30 m (right). The full example is in vignette("AlphaSDM").

Why AlphaSDM

• No environmental layers. There is nothing to find, download, reproject or align; the embeddings already describe every pixel.
• Fine resolution everywhere. 10 m pixels, every year from 2017, on land anywhere on Earth.
• Nothing to download but the map. Sampling, model fitting and prediction all run on Earth Engine, so a large study area costs your computer nothing.
• An ensemble with calibration built in. Support vector machine, random forest and boosted trees by default, scored with AUC, TSS and the Boyce index.
• Explicit modelling choices. Pseudo-absence placement follows Barbet-Massin et al. (2012), and AlphaSDM makes you choose the strategy rather than choosing it for you.

AlphaEarth embeddings

AlphaEarth Foundations is a Google DeepMind model that condenses optical, radar, lidar, climate and other data into 64 numbers per 10 m pixel per year. The annual embeddings are a public Earth Engine dataset, currently covering 2017 to 2025. Records are matched to the embeddings for the year they were made.

Installation

# install.packages("pak")
pak::pak("James-Longo/AlphaSDM")

Earth Engine setup

AlphaSDM runs on your own Earth Engine account, which is free for noncommercial use.

1. Register for Earth Engine. This gives you a Cloud project ID.
2. Connect once per machine. A browser window asks you to allow access, and the connection is remembered after that.

library(AlphaSDM)
setup_gee(project = "your-project-id")
gee_status()   # checks credentials, project and a live connection

On a machine without a browser, use setup_gee(auth_mode = "notebook") to paste a code instead. clear_gee_credentials() resets everything.

Example

Download saguaro records from GBIF, fit the default ensemble on 2022 records, test it on 2023 records, and map suitability:

library(AlphaSDM)

gbif_records <- function(year) {
  url <- paste0("https://api.gbif.org/v1/occurrence/search?",
                "scientificName=Carnegiea%20gigantea&year=", year,
                "&hasCoordinate=true&hasGeospatialIssue=false",
                "&coordinateUncertaintyInMeters=0,30",
                "&decimalLongitude=-111.4,-110.6&decimalLatitude=31.9,32.6&limit=300")
  do.call(rbind, lapply(c(0, 300), function(offset)
    jsonlite::fromJSON(paste0(url, "&offset=", offset))$results[
      , c("decimalLongitude", "decimalLatitude", "year")]))
}
coords <- c("decimalLongitude", "decimalLatitude")

# Fit on 2022 records with pseudo-absences
pres <- format_data(gbif_records(2022), coords = coords, year = "year")
occ  <- generate_pseudo_absences(pres, aoi = "bbox", strategy = "combined",
                                 n = nrow(pres))

# Test on 2023 records against random background
pres_2023 <- format_data(gbif_records(2023), coords = coords, year = "year")
test <- generate_pseudo_absences(pres_2023, aoi = "bbox", strategy = "random",
                                 n = 2000)
fit  <- evaluate_models(occ, predict_coords = test)
fit$metrics$ensemble

maps <- generate_map(occ, aoi = "bbox", scale = 30,
                     output_dir = "saguaro")

generate_map() writes one GeoTIFF per model plus the ensemble. Maps download straight from Earth Engine in tiles; a map Earth Engine will not compute that way goes through its batch system and Google Drive instead, which is slower.

Models

The default ensemble is c("svm", "rf", "gbt"). methods = also accepts "maxent", "glm", "cart", "knn", "mindist" and "similarity", all fitted on Earth Engine; see ?evaluate_models.

Getting help

Report bugs and request features in GitHub issues, or email [email protected]. AlphaSDM is in active development, so arguments and defaults may still change.

Citation

Run citation("AlphaSDM") in R, or use GitHub's "Cite this repository" button, which reads CITATION.cff.

License

MIT; see LICENSE.md. The AlphaEarth embeddings are provided by Google under the terms of the Earth Engine dataset.