Overall Pearson r ≈ 0.82 reflects agreement across all crops pooled together — the model ranks crops in the right order almost everywhere (wheat and maize always dominate; potatoes and sugar beet are always minor). That between-crop ordering accounts for most of the variance in the pooled scatter, hence the inflated global r.

Per-crop r ≈ 0.30–0.72 measures a harder question: once you fix the crop, does the model track county-to-county variation correctly? Here the signal shrinks to the spread within one crop. Potatoes (r = 0.72) and sugar beet (r = 0.70) — the rarest crops with the clearest spatial clustering — track INSSE best. Sunflower (r = 0.32) and other cereals (r = 0.30) track poorly; their spatial distribution is noisier in both the model and the declared data.

MAE ≈ 4.8 pp overall (absolute percentage-point error) means the model's county share for a typical crop sits within ~5 pp of the INSSE figure. Maize has the biggest absolute error (13.6 pp — the model over- or under-estimates it by ~14 points in counties where maize is a major crop).

Model vs CTY (training reference): the CTY toggle swaps in per-county stats computed from the same cropland raster but pulled from CTY (the satellite-derived training labels). The same zonal-stats pipeline is run over DigiFarm field polygons. This answers a cleaner question: how much does the model track INSSE, and how much does CTY itself track INSSE? If CTY beats the model significantly, the model is losing signal relative to its own ground truth.

• Overall: model r = 0.824 · CTY r = 0.907 (+0.08) · model MAE 4.8 pp → CTY MAE 3.5 pp
• Biggest gains for CTY: sunflower (0.32 → 0.74), soybeans (0.53 → 0.85), maize (0.40 → 0.77)
• CTY years available: 2018-2023 (INSSE: 2017-2024). Model covers 2017-2025.

Caveat: the model and CTY both measure classified pixel share (includes non-cropland inside field polygons), while INSSE AGR108A is farmer-declared cultivated area. To make them comparable, all shares are renormalized to the subset of INSSE-reported crops.