An automated hybrid deep learning framework for paddy leaf disease identification and classification.
Journal:
Scientific reports
Published Date:
Jul 24, 2025
Abstract
In India, agriculture remains the primary source of livelihood for many people. Pathogen attacks in crops and plants significantly diminish both the yield and quality of production, leading to financial losses. As a result, identifying diseases in crops is highly important. As the population grows, the demand for rice also rises. Therefore, disease management is vital in rice cultivation, and rapid identification of rice diseases is critical for timely pesticide application and effective control. Consequently, there is a need to boost agricultural productivity by adopting new technologies. Deep learning is a popular area of research in various fields. This research aims to design and propose a new automated model using a deep learning model for the disease identification and categorization of paddy leaves. The system follows a structured workflow comprising several stages: image acquisition, pre-processing, feature extraction, feature selection, and classification. Images of paddy leaves were obtained from the paddy doctor dataset hosted on Kaggle. The data is pre-processed by choosing the RoIs, labelling, enhancement, and segmentation using adaptive thresholding and grouped using K-means clustering. The MobileNetV3 model, a pre-trained transfer learning approach, extracted colour, shape, and texture features. The vital features are selected using the hybrid Genghis Khan Shark Optimization (GKSO) with Simulated Annealing (SA) algorithm. The chosen features are subsequently fed into the CatBoost for disease classification. The deep learning techniques introduced for disease identification and classification have been compared with various conventional classifiers, and the system's performance has been validated using metrics such as accuracy, sensitivity, and F1-score. Performance investigations prove that the technique efficiently yields a higher accuracy of 98.52%, outperforming state-of-the-art techniques.