Advanced Hybrid ViT-ConvNeXt Framework for Robust Multi-Part Plant Disease Detection with Entropy-Driven Feature Selection

Abstract

Existing classification systems of hybrid plant disease detection use CNN-based features extraction (e.g., VGG-16, ResNet-34, EfficientNet-B4) with SVM classification but are subject to major limitations (redundant features are extracted, features and classes are not well correlated, small datasets are overfitted, and low-quality images are poorly classified). Moreover, they are mostly limited to leaf-only analysis and do not have the validation on field data. To eliminate these problems, this paper introduces a developed hybrid model combining Vision Transformers, ConvNeXt, and EfficientNetV2 to extract features better. It uses the entropy-, ANOVA-, and mutual information-based feature selection to eliminate dimensionality and class relevance. Preprocessing of images enhances the capabilities to resist degraded input, and multi-part analysis of the plants can be used by real-life data that goes beyond leaves. The framework has lower computational cost, less overfitting, greater robustness to noise and complex backgrounds, greater F1-scores on imbalanced data and can be applied to mobile/IoT-based real-time disease diagnostic.