B-MO IRON ORE TRAINING SUMMARY
Deposits used: 975   Background points: 1950   Spatial blocks: 122

Model ranking (spatial CV, all points), best first:
  0.858  XGBoost  [+terrain]   finds 60% of deposits at 10% false alarms
  0.853  Gradient Boosting  [+terrain]   finds 59% of deposits at 10% false alarms
  0.853  Random Forest  [+terrain]   finds 59% of deposits at 10% false alarms
  0.852  Extra Trees  [+terrain]   finds 57% of deposits at 10% false alarms
  0.849  LightGBM  [+terrain]   finds 61% of deposits at 10% false alarms
  0.848  Logistic Regression  [+terrain]   finds 60% of deposits at 10% false alarms
  0.845  SVM (RBF)  [+terrain]   finds 61% of deposits at 10% false alarms
  0.840  k-Nearest Neighbours  [+terrain]   finds 57% of deposits at 10% false alarms
  0.828  Extra Trees  [+thermal]   finds 52% of deposits at 10% false alarms
  0.823  Random Forest  [+thermal]   finds 52% of deposits at 10% false alarms

Statistical engine chosen: Statistical engine (likelihood variant), feature set '+terrain', AUC 0.782
Average engine score  -> known deposits: 55.4   background: 40.2
Where deposits and background land (out-of-fold, %):
    class  deposits_pct  background_pct
 Very low          10.8            50.0
      Low          29.0            25.0
 Moderate          18.6            15.0
     High          18.2             7.0
Very high          23.5             3.0

Top features: B8 (AUC 0.22), brightness (AUC 0.25), B11 (AUC 0.26), slope (AUC 0.73), B3 (AUC 0.29), B12 (AUC 0.30)
Hyperspectral: evaluated on the small covered subset - see model_comparison.csv