Design of an Ensemble Segmentation, Feature Processing & Classification model for identification of Cotton Fungal diseases

Abstract

Cotton fungal diseases include rust, alternaria leaf spot, fusarium wilt, grew mildew, and root rots. Identification of these diseases requires design of efficient fungi segmentation, feature representation & classification models. Existing methods that perform these tasks, are highly complex, and require disease-specific segmentation techniques, which limits their scalability levels. Moreover, low-complexity models are generally observed to showcase low accuracy levels, which restricts their applicability for real-time use cases. To overcome these issues, proposed design focused on a novel ensemble segmentation, feature processing & classification model for identification of cotton fungi diseases. The proposed model initially uses a combination of Fuzzy C Means (FCM), Enhanced FCM, KFCM, and saliency maps in order to extract Regions of Interest (RoIs). These RoIs are post-processed via a light-weight colour-feature based disease category identification layer, which assists in selecting the segmented image sets. These image sets are processed via an ensemble feature representation layer, which combines Colour Maps, Edge Maps, Gabor Maps and Convolutional feature sets. Due to evaluation of multiple feature sets, the model is able to improve classification performance for multiple disease types. Extracted features are classified via use of an ensemble classification model that combines Naïve Bayes (NB), Support Vector Machines (SVMs), Logistic Regression (LR), and Multilayer Perceptron (MLP) based classifiers. Due to this combination of segmentation, feature representation & classification models, the proposed Model is capable of improving classification accuracy by 5.9%, precision by 4.5%, recall by 3.8%, and delay by 8.5% when compared with state-of-the-art models, which makes it useful for real-time disease detection of crops.