A hybrid deep learning paradigm integrating segmentation architectures for precise plant disease identification and classification

Abstract

Addressing the challenging issue of complex background interference in plant disease identification, recent research employs diverse deep learning (DL) methodologies on both publicly available and customized datasets. This study introduces a two-step DL approach for plant disease classification. Initially, an enhanced convolutional neural network (CNN) is developed through a comparative analysis of prominent CNN architectures, including customized and cascaded versions of select DL models, achieving an accuracy of 93.3%. To further enhance accuracy, segmentation architectures such as DeepLabV3+, UNet, Iterative UNet, and UNet with Atrous Spatial Pyramid Pooling (ASPP) are integrated before customized CNN architectures. These segmentation algorithms effectively isolate the diseased portions of leaf images. Notably, the UNet with ASPP architecture demonstrates reduced time complexity, minimizing the number of features to be trained, and significantly improves accuracy to 99.8%, outperforming other predefined architectures. The models are trained on a plant village dataset, detecting 10 different diseases across various plant species, including tomato, corn, and potato.