Plant Disease Classification Using Deep Learning Driven by Lesion-Focused Next-Generation Segmentation

Abstract

Plant disease detection using leaf images has become a critical component of
precision agriculture, enabling timely intervention and improved crop management. This study
presents a deep learning-based framework for accurate and robust plant disease classification. A
novel dataset, EruCauliflower DB, was developed, comprising high-resolution images of
cauliflower leaves affected by Alternaria Leaf Spot (114 images) and Black Rot (99 images).
The proposed framework consists of three stages. First, a novel lesion-focused segmentation
technique, BorB, is introduced to effectively isolate diseased regions by combining Lab and
RGB color space features through a logical OR operation. Second, data augmentation
techniques, including geometric transformations, are applied to enhance data diversity and
improve model generalization. Finally, four deep learning models VGG16, ResNet50,
EfficientNetB3, and MobileNetV3 Large—are employed for classification. Experimental results
demonstrate that the proposed system achieves 100% classification accuracy on the Eru
Cauliflower DB dataset across all models. The framework is further validated on the Mango Leaf
BD dataset, also achieving 100% accuracy, and on 15 classes from the Plant Village dataset,
reaching 99.78% accuracy. These results highlight the effectiveness, robustness, and
generalization capability of the proposed approach, making it suitable for real-world agricultural
applications.