An Improved Particle Swarm Optimization integrated Dilation Convolutional Neural Network for Automatic Soybean Leaf Disease Identification

Abstract

Leaf disease is a prominent and destructive ailment that affects plants. Timely identification and early detection are crucial for enhancing the future probability of leaf diseases that impact plants. The investigation of soybean plant leaf disease detection has gained importance owing to its major impact on soybean growth, leading to decreased productivity and quality. The traditional method of identifying soybean leaf diseases mostly relies on agricultural specialists, resulting in a significant amount of time being utilized. Deep Learning (DL) models are promising techniques to identify soybean leaf disease detection. However, various ongoing investigations are going on to achieve an effective model with efficient practical application. To address this problem, this study proposes the use of a hybrid, smart and intelligent model based on dilation Convolutional Neural Network (CNN) to identify diseases of soybean leaves. Selecting and designing the ideal model structure is still a difficult task, even though DL networks demonstrate remarkable efficacy. The accuracy of plant disease detection based on leaf analysis may be improved by fine-tuning the values of the hyper-parameter of dilation CNN. The proposed framework has been trained using a dataset of 1620 soybean leaf images that have been divided into six different diseased groups. The Velocity Pausing Particle Swarm Optimization (VP_PSO), a well-studied metaheuristic technique, is employed to optimize the hyper-parameters of the dilation CNN. This optimization aims to improve the effectiveness of the dilation CNN in accurately recognizing diseases present in soybean plant leaves. The suggested hybrid model performs better than other standard hybrid models such as classical CNN, VGG16, MobileNetV2, ResNet101, dilation CNN and PSO_Dilation_CNN. As per the experimental research, the suggested VP_PSO _Dilation CNN model has a detection accuracy of 95.32%.