Comparative Evaluation of Hybrid Deep Learning Models for Multiclass Capsicum Leaf Disease Detection

Abstract

In modern agriculture early and accurate disease diagnosis is vital for increasing crop productivity and minimizing economic losses in plant leaf diseases. Capsicum (Capsicum annuum L.) is a high-value vegetable crop that is very vulnerable to several fungal diseases of leaves and timely disease diagnosis is important to achieve successful crop management. This study proposes an image processing system for recognizing the multi-classes capsicum leaf diseases by using hybrid deep learning methods. The dataset consisted of 7,735 images of leaves from the Mendeley Pepper Bell Disease dataset and images of leaves taken from the field in Ahilyanagar district of Maharashtra, India. There are six types of leaf diseases: Cercospora Leaf Spot, Bacterial Spot, Leaf Curl, Nutrition Deficiency, Powdery Mildew, & Healthy leaves. Before the development of the models, all of the images were preprocessed by resizing, normalization, and augmentation of the data. The models of the Convolutional Neural Network (CNN), CNN-Support Vector Machine (CNN-SVM), and CNN-Long Short-Term memory (CNN-LSTM) were designed and tested under the same training, validation, and testing conditions. The accuracy, recall, specificity, precision, F1 score & multiclass confusion matrix analysis were used to evaluate the results of the model. The CNN-LSTM model performed between 98.9% classification accuracy, which is higher compared to CNN-SVM (96.7%) and conventional CNN (95.3%) models. The results prove that hybrid deep learning architectures can offer better feature learning and classification accuracy to detect capsicum leaf diseases, which can be applied for the development of intelligent decision support systems in precision agriculture.