Residual Vision Driven Ensemble Framework for Robust Mulberry Leaf Quality Classification

Abstract

Timely and accurate classification of mulberry leaf health is crucial in sericulture for optimizing crop yield and preventing disease proliferation. Despite the progress in deep convolutional networks, current solutions face challenges in generalization, computational cost, and robustness across diverse leaf classes. To alleviate these challenges, we propose a novel hybrid ensemble classification framework integrating Residual Vision Network-based feature extraction with Light GBM, Random Forest, and XGBoost classifiers, optimized through extensive hyperparameter tuning and stratified 5-fold cross-validation. The study leverages the Kaggle Mulberry Leaf dataset, encompassing three distinct classes, with class-balanced data split into training and test subsets at a ratio of 80:20. The results reveal that the proposed ensemble strategy not only outperforms standalone deep networks such as DenseNet121, VGG19, and MobileNetV3 but also exhibits superior discriminative performance, achieving an F1-score of 98.17%, AUC of 99.38%, and specificity of 98.36%. Ablation analysis further substantiates the contribution of the Residual Vision module, yielding performance gains of up to 3.12% in F1-score over non-residual variants. This work effectively bridges the limitations of current deep models by combining residual-enhanced feature extraction with lightweight gradient-boosted decision ensembles, offering a scalable and accurate approach for agricultural vision systems. The proposed framework holds substantial promise for deployment in real-time leaf monitoring applications with low computational overhead and high interpretability.