Xception Deep Kronecker Network for Severity Plant Disease Classification Using Hyperspectral Leaf Image

Abstract

Plant diseases have always been a problem because they can significantly decrease both the quality and quantity of crops. Diseases, pests, and weeds present major challenges in crop cultivation, leading to substantial crop damage and posing significant risks to the economy and food security. Plant diseases pose a significant threat to the quality and yield of agricultural products. Prompt and reliable detection and identification of these diseases are crucial for ensuring sustainable agriculture and food security. Preventing ailments and providing guidance to farmers is crucial to enhancing the yield on a large scale. Manual feature extraction is the most expensive approach used in earlier plant disease detection methods. Additionally, many of the real-time applications face issues with cost complexity, misclassification, and overfitting. Hence, an effective model called Xception-Deep Kronecker Network (Xception-DKN) is proposed for severity disease classification utilising hyperspectral leaf image. Initially, the hyperspectral leaf image is pre-processed. Then, the selection of the band phase is performed utilising Fractional Water Wheel Plant Dingo Optimizer (FWWPDO), that is the incorporation of Dingo Optimizer (DOX), Fractional Calculus (FC), and Water Wheel Plant Algorithm (WWPA). Outputs from the selection bands are forwarded into the leaf segmentation phase that is carried out using Black Hole Entropic Fuzzy Clustering (BHEFC). Next, using a majority voting approach, a fusion of bands is performed. Then, fused band output as well as individual leaf segmentation outcome is exposed into the Feature Extraction (FE) stage for extracting the features, including Weber Local Descriptors (WLDs) and Local Binary Patterns (LBPs). Thereafter, disease recognition is executed on leaves by utilising a Deep Conval Neural Network (deep CNN) for normal and abnormal cases. Nevertheless, Deep CNN hyperparameters are fine-tuned utilising FWWPDO, which is developed by integrating the Water Wheel Plant Dingo Optimizer (WWPDO) and Fractional Concept (FC). Thereafter, severity level classification is performed using the proposed Xception-DKN into low, moderate and severe cases. Xception-DKN is the combined form of Xception and Deep Kronecker Network (DKN), where the layers are adjusted by Taylor concepts. The Xception-DKN has achieved the highest accuracy of 92.204%, true positive rate (TPR) of 94.011%, and true negative rate (TNR) of 91.210%.