Independent validation against LUCAS
LUCAS (Land Use/Cover Area frame Survey) is the EU's harmonised in-situ land-cover survey, run every three years across all member states by Eurostat (DG-ESTAT) with field surveyors. It is independent of every label source we trained on (CTY, RPG, BRP, INVEKOS, DUN, GERK). We compare model predictions at LUCAS point coordinates against the surveyor-recorded LC1 land-cover code, harmonised into the model's 13-class crop schema.
Overall agreement
LUCAS LC1 (harmonised) vs model predicted class.Per-class precision · recall · F1
One-vs-rest scores against LUCAS LC1. Classes with zero LUCAS support are hidden.Confusion matrix
Rows = LUCAS truth, columns = model prediction. Numbers are point counts; cell intensity = row-normalised share.How to read these numbers
What's measured. For every Romania LUCAS point in the year, we sample the model's input features (AEF 64-dim + TESSERA 128-dim + lon/lat) at that exact coordinate, run the production EU-pooled LightGBM classifier, and compare the predicted crop class to the harmonised LC1 ground truth recorded by the LUCAS surveyor.
Field-visit toggle. LUCAS 2022 mixes field-visit observations (obs_type 1–5, where the surveyor walked or drove to the point) with office photo-interpretation (obs_type 7). Field-visit points are higher quality. The toggle restricts to obs_type 1–5.
What this is not. LUCAS samples land-cover at a 1.5 m radius around a point; the model classifies a 10 m pixel. Edge effects, mixed-pixel parcels, and LC1 codes that don't map cleanly into our 13-class crop schema (e.g. fallow, mixed legume-fodder) will show up as disagreement that isn't really a model error.
LUCAS does not split orchards into apples / pears / cherries the way RPG does, and the B82 vineyards code is mapped 1:1; sub-class confusion inside orchards is invisible to this benchmark.