Accurate identification of potato leaf diseases presents considerable challenges due to the high visual similarity among disease types, varying lighting conditions, and complex backgrounds in real-world imagery. This study addresses these issues by constructing a comprehensive 176-dimensional handcrafted feature set, which incorporates global color statistics, texture descriptors, and local invariant features to capture diverse visual patterns. To improve classification performance and reduce dimensionality, three embedded feature selection techniques, LightGBM, Random Forest, and Logistic Regression with L1 regularization were systematically evaluated. Experiments on a real-world dataset demonstrate that LGBM-based feature selection achieved the highest accuracy of 79.10% using only 88 features, outperforming both the full feature set and the other selection methods. These findings underscore the importance of targeted feature selection in enhancing classification accuracy and reducing model complexity, particularly for agricultural image analysis in challenging visual conditions.