Model otwarty: tylko Sentinel-2
Cechy (2): s2_ndci_p90, s2_scum_frac. Sanepid tylko jako etykieta. Uczenie: [2023, 2024, 2025], test: [2026], protokół jak w train.py.
Cel: y_cyano_0
41901 wierszy testu, 590 z sinicami. 28 drzew, próg 0.04, ROC-AUC w CV 0.729.
| Wariant | ROC-AUC | PR-AUC | precyzja | czułość | F1 | top 10 | top 20 | top 50 |
|---|---|---|---|---|---|---|---|---|
| losowo | 0.496 | 0.014 | – | – | – | 3% | 6% | 11% |
| historia | 0.881 | 0.294 | – | – | – | 35% | 42% | 60% |
| open: S2 | 0.82 | 0.094 | 0.077 | 0.425 | 0.13 | 18% | 28% | 44% |
Cel: y_cyano_1
41305 wierszy testu, 587 z sinicami. 17 drzew, próg 0.02, ROC-AUC w CV 0.708.
| Wariant | ROC-AUC | PR-AUC | precyzja | czułość | F1 | top 10 | top 20 | top 50 |
|---|---|---|---|---|---|---|---|---|
| losowo | 0.501 | 0.015 | – | – | – | 3% | 6% | 12% |
| historia | 0.881 | 0.294 | – | – | – | 35% | 42% | 60% |
| open: S2 | 0.817 | 0.103 | 0.052 | 0.678 | 0.097 | 21% | 27% | 48% |
Cel: y_cyano_3d
41305 wierszy testu, 763 z sinicami. 31 drzew, próg 0.04, ROC-AUC w CV 0.679.
| Wariant | ROC-AUC | PR-AUC | precyzja | czułość | F1 | top 10 | top 20 | top 50 |
|---|---|---|---|---|---|---|---|---|
| losowo | 0.507 | 0.019 | – | – | – | 3% | 5% | 11% |
| historia | 0.874 | 0.277 | – | – | – | 29% | 38% | 57% |
| open: S2 | 0.777 | 0.09 | 0.079 | 0.515 | 0.137 | 16% | 23% | 41% |