Across the Darab region of Iran, a quiet but consequential problem in soil science has found a partial remedy: the tendency of machine learning models to overlook rare soil types—the very ones most critical for land stewardship—has been partially corrected through a disciplined combination of statistical feature selection and synthetic data generation. A research team working with 140 soil profiles demonstrated that pairing VIF-based feature selection with SMOTE resampling and a Random Forest classifier could lift overall prediction accuracy by fifteen percent and bring a previously invisible