Enhancing Hybrid Classification for Plant Diseases With Deep Feature Selection Based on Analytical Entropy and Statistical Method

The purpose of much research in the hybrid classification area is to reduce the number of deep features. However, many approaches overlook the relation between deep features and specific classes or diseases. This study aims to use information entropy to distinguish between valuable and less valuable features in the extracted deep feature set. The architecture of state-of-the-art models influences the characteristics of deep features, which are of three types: negative, zero, and positive values. Accurate predictions in typical machine learning models entail identifying valuable features, especially those with the most positive values. We perform one-way analysis of variance (ANOVA) and mutual information (MI) analysis to verify the correlation between the remaining deep features and specific classes or diseases. We discard features with weak relations based on predefined thresholds established by the involved f-statistic and MI mean values. This enhanced hybrid classification approach uses the entropy mean threshold, skewness visualization, the standard deviation, and the mean probability of valuable feature threshold to select valuable features and remove less valuable features, as well as one-way ANOVA and MI to evaluate relations and eliminate weakly related deep features. The results of all experiments show that the proposed approach exhibits higher accuracy than traditional hybrid classification and remains effective across all plant species. For tomatoes, ResNet-34 is the best model, achieving an excellent accuracy of 99.94% with 98 features. For sugarcane, ResNet-34 achieves a remarkably accuracy of 99.41% with 121 features. For corn, EfficientNet-B4 attains a significant accuracy of 99.77% with 123 features.