CROME
crome is a Python package for 10 m crop classification across England using Google's AlphaEarth Foundations satellite embeddings and DEFRA's Crop Map of England (CROME) vector references.
What it does
- Downloads AlphaEarth 64-band annual embeddings at 10 m resolution (2017--2025) via edown
- Downloads the national CROME GeoPackage reference from DEFRA's Data Services Platform
- Rasterizes CROME hexagon labels onto each AlphaEarth tile's native grid
- Trains a tile-local Random Forest classifier with cached, reusable training samples
- Predicts a 10 m crop map per tile
- Pools tile-level samples into regional or national models without re-reading rasters
All steps are accessible as individual CLI commands or as combined pipelines. For cluster execution, a Nextflow wrapper parallelizes tiles across Slurm on JASMIN.
Getting started
pip install .[ee]
crome download-run-baseline \
--year 2024 --aoi-label east-anglia \
--bbox -1 51 0 52 --output-root ./outputs
See the Installation guide for setup details and optional dependencies.
Documentation
| Page | Contents |
|---|---|
| Installation | Setup, dependencies, optional extras, environment variables |
| Architecture | Module map, data flow, design decisions |
| CLI reference | All 18 commands grouped by workflow stage with examples |
| Configuration | Label modes, overlap policies, output directory layout |
| Nextflow on JASMIN | Cluster execution with Slurm queue profiles |
| Migration status | What is implemented, what is deferred, remaining gaps |
Supported years
AlphaEarth annual embeddings cover 2017--2024 (stable) and 2025 (rolling by UTM zone). CROME references are available from DEFRA for 2016 onwards. Each download or pipeline run targets one year; multi-year workflows run one year at a time and pool the results.