A Hybrid CNN and Segmentation-Based Pruned Deep Learning Approach for Precision Plant Disease Detection

The identification of plant diseases has become increasingly challenging due to the interference of complex backgrounds in images, which often hinders the accuracy of classification models. Recent studies have employed various Deep Learning (DL) techniques to overcome this issue, utilizing both publicly available and custom datasets. However, achieving high accuracy while managing background complexity remains a significant hurdle. This paper aims to address this challenge by introducing a two-step DL approach for plant disease classification. The approach begins with an enhanced Convolutional Neural Network (CNN), developed through a comparative analysis of several CNN architectures, including customized and cascaded versions of prominent DL models, achieving an accuracy of 93.3%. To further enhance accuracy, segmentation techniques such as DeepLabV3+, UNet, Iterative UNet, and UNet with Atrous Spatial Pyramid Pooling (ASPP) are integrated before applying customized CNN architectures. These segmentation methods effectively isolate diseased portions of leaf images, improving classification performance. The proposed methodology introduces model pruning to optimize performance and computational efficiency by removing redundant parameters and less significant features. The UNet with ASPP architecture, in combination with pruning strategies, significantly reduces time complexity and feature redundancy, leading to an impressive accuracy of 99.8%. This approach outperforms other existing models in terms of accuracy and efficiency. The model is trained on the Plant Village dataset, which includes 10 different diseases across plant species such as tomato, corn, and potato, offering a comprehensive solution for plant disease identification.