TomOM04/miscanthus-alphaearth-uk-predictions

UK Miscanthus AlphaEarth field predictions

This repository contains four Random Forest binary classifiers trained to identify Miscanthus from AlphaEarth satellite embeddings, together with two UK-wide field-level prediction maps.

Model and prediction method

The positive class is Miscanthus, represented by the edited Miscanthus marker coordinates. Four separate negative classes were sampled from UKCEH crop-map fields: Grass, Cereal crops, Broadleaf/root-row crops, and Other crops. Each classifier is a 300-tree scikit-learn RandomForestClassifier using 64 AlphaEarth embedding bands. Class weights use balanced_subsample; the saved models are the final fits used for inference.

For the UKCEH map, each field selects the classifier matching its UKCEH crop group. For the FIBOA UKFields map, crop labels are unavailable, so all four classifiers are applied and the highest Miscanthus probability is retained. The output also stores all four FIBOA model probabilities and the winning model.

Probabilities are reported as percentages from 0 to 100. Uncertainty is the variation among the 300 individual tree probabilities, with an empirical 2.5--97.5% interval. The threshold-sweep CSVs report predicted field counts and area for every threshold from 1% through 100%.

Validation scores

Scores below are from the marker/UKCEH training run used to create the maps. Training accuracy is included for completeness but is optimistic; held-out test and cross-validation scores are more informative.

These are point/field validation scores for the binary classifiers, not an independent accuracy assessment of the two UK-wide maps. The FIBOA map uses a maximum across four classifiers and therefore should be interpreted as a candidate map requiring external validation.

Data sources

• —Google AlphaEarth annual 2024 embeddings, dequantized into 64-band GeoTIFF tiles covering the UK.
• —UKCEH/LCM 2024 crop map fields and crop labels, in British National Grid (EPSG:27700).
• —Edited Miscanthus marker coordinates used as positive training examples.
• —FIBOA UKFields field boundaries from Field_segmentation/data/labels/ukfields/ukfields.parquet, also in EPSG:27700.

Field-level predictions use the AlphaEarth value sampled at each polygon's representative point. Missing AlphaEarth coverage is recorded in the output and is not silently treated as a negative prediction.

Files

• —weights/*_final_model.pkl: four trained Random Forest weights.
• —model_manifest.json: model paths and validation scores.
• —maps/miscanthus_ukceh_boundaries.gpkg: UKCEH field-boundary predictions.
• —maps/miscanthus_ukfields_boundaries.gpkg: FIBOA UKFields predictions.
• —scores/*threshold_sweep.csv: predictions/area at thresholds 1--100%.
• —code/make_two_miscanthus_predictions.py: reproducible inference script.

The pickle files were trained with scikit-learn 1.7.2. Use the same version where possible when loading them; newer versions may emit an InconsistentVersionWarning.