Kernel to computation: identifying optimal feature set for red rice classification

While existing research focuses extensively on white rice classification with readily available datasets, automated classification of red rice varieties remains largely unexplored with no publicly available datasets, creating a significant research gap in agricultural image processing applications. This research presents a study on red rice classification, a relatively unexplored area with no prior publicly available datasets or focused investigations on red rice variety identification. This study classifies three distinct red rice varieties—Uma, KCP-1, and Jyothi—primarily cultivated in Karnataka and Kerala, using image processing and machine learning techniques. Six ML models were evaluated with seven unique feature combinations derived from size, shape, and texture characteristics to identify the most discriminative feature set. Feature selection was performed using Recursive Feature Elimination and Backward Feature Elimination to enhance model efficiency. Hyperparameter tuning was applied to optimize classification performance, and k-fold cross-validation with statistical significance testing was used to assess generalization and validate model performance differences. The integration of size, shape, and texture features yielded the highest average accuracy across the models, with K-Nearest Neighbours achieving 98.67 % accuracy and Support Vector Machine reaching 97.34 % accuracy with the size and shape combination. The findings emphasize the importance of optimal feature selection and tuning in improving classification accuracy, contributing to the development of automated classification systems for red rice varieties.