Mislabel Rate by Country & Year
Cross-Year Consistency (France, 2018-2024)
Fields flagged as mislabeled in 2+ years are likely true declaration errors, not model mistakes.
Top Confused Crop Pairs
Most common cases where the model confidently predicts a different crop than declared.
| Declared | Predicted | Count | Avg Confidence |
|---|
Per-Class Mislabel Rates
Methodology
1. Satellite Embeddings
We use two complementary satellite embedding systems: AEF (64-dim, annual composites) and TESSERA (128-dim, foundation model). Combined into a 192-dim dual feature vector per field.
2. Spatial Cross-Validation
5-fold spatial block CV (0.5° grid) ensures no spatial leakage. Each field gets an out-of-fold prediction it was never trained on.
3. Confidence Thresholding
A field is flagged as mislabeled when the model predicts a different crop with >80% probability. This captures high-confidence disagreements.
4. Cross-Year Consistency
Fields flagged in 2+ years are classified as persistent mislabels (likely true errors). One-off flags are mostly model noise.