Classification: no promised win

Logistic regression predicts the central pixel’s class from all 36 standardized features or from PCA scores. Component count and C are chosen with three stratified training folds, each Pipeline refitting its mean, scale and PCA inside the fold; the test set is used only after selection.

Balanced accuracy averages recall over classes. A reference run selects 24 components for both PCA routes, with held-out balanced accuracy about 0.794 for all features, 0.806 for covariance PCA-24 and 0.798 for correlation PCA-24 — one benchmark, not a universal ranking.

In the confusion matrix, damp grey soil has much lower recall than cotton or red soil.