Cropland mask dataset for the Canadian Prairies derived from Google Satellite Embedding imagery

Description

# Canadian Prairies Cropland Masks — 10 m (GeoTIFF) # Overview This repository/package accompanies a spatially explicit raster dataset of cropland mask for the Canadian Prairies (Alberta (AB), Saskatchewan (SK), and Manitoba (MB)) # Methods Summary (How the dataset was created) 1. Predictor generation: Multi-source predictors were derived from AlphaEarth embeddings. 2. Reference sampling: Stratified-random sample points were distributed across the Prairies to represent cropland and non-cropland conditions. 3. Model training: A Random Forest classifier was trained to separate cropland from non-cropland using AAFC crop mask as labels. 4. Wall-to-wall inference: The trained models were applied to generate annual binary cropland masks for 2017–2024 at 10 m. 5. Aggregation: Annual masks were summed to generate a cropland frequency layer. 6. Stable cropland definition: Pixels classified as cropland in more than two years were labeled as persistent cropland to produce a stable mask. -- ## Spatial and Technical Specifications - Geographic extent: Canadian Prairies — Alberta, Saskatchewan, Manitoba - Spatial resolution: 10 m - Raster format: GeoTIFF - Coordinate reference system (CRS): EPSG:4326) - NoData value: 0 - Pixel encoding: - Annual mask: 0 = non-cropland, 1 = cropland --- Caution / best practices: - Annual masks represent modeled predictions and may contain errors near edges of fields, heterogeneous land cover, or mixed pixels. - If comparing with other products, ensure consistent CRS, resolution, and resampling methods.

Steps to reproduce

## Methods Summary (How the dataset was created) 1. Predictor generation: Multi-source predictors were derived from AlphaEarth embeddings. 2. Reference sampling: Stratified-random sample points were distributed across the Prairies to represent cropland and non-cropland conditions. 3. Model training: A Random Forest classifier was trained to separate cropland from non-cropland. 4. Wall-to-wall inference: The trained models were applied to generate annual binary cropland masks for 2017–2024 at 10 m. 5. Aggregation: Annual masks were summed to generate a cropland frequency layer. 6. Stable cropland definition: Pixels classified as cropland in **more than two years were labeled as persistent cropland to produce a stable mask.

Institutions

• University of Saskatchewan

  Saskatchewan, Saskatoon

Categories

Funders

• Canadian Space Agency (CSA)
• Agriculture Development Fund (ADF)
• Western Grains Research Foundation (WGRF)

Licence