Multi-Variant ResNet and HOG Descriptor-Based Ensemble for Citrus Disease Recognition

Abstract

Early and accurate detection of citrus plant diseases is vital for improving crop yield and supporting sustainable agricultural practices. This paper proposes a robust hybrid deep learning framework that integrates three ResNet architectures—ResNet34, ResNet50, and ResNet101—for effective disease detection and classification. The system begins by acquiring a citrus leaf image dataset and performing preprocessing techniques such as resizing, normalization, noise reduction, and data augmentation. Color space transformations and morphological operations are applied to isolate diseased regions, while shape-based features are extracted using Histogram of Oriented Gradients (HOG). Leveraging transfer learning, the pre-trained ResNet models are fine-tuned to classify specific citrus diseases. An ensemble learning approach—employing soft voting, majority voting, and stacked generalization—combines the strengths of all three models, enhancing classification accuracy and generalization capability. The model's performance is comprehensively evaluated using accuracy, precision, recall, specificity, sensitivity, and F1-score. The results demonstrate that the proposed hybrid system provides a scalable, accurate, and efficient solution for early citrus disease detection, enabling smarter agricultural monitoring and optimized yield management.