JoWUA

Abstract

Agricultural productivity, crop quality, and food security worldwide can be highly impacted by plant diseases. Early detection and accurate diagnosis of disease in crops are vital for minimizing losses and sustaining precision farming practices. Regrettably, all modern disease diagnosis approaches based on deep learning techniques have concentrated on image classification, neglecting the time dependency of the disease process during different stages of the crops’ development cycle. Furthermore, traditional CNN-LSTM models have been associated with increased computational complexities and high memory costs. This research suggests a CNN-GRU hybrid model for early disease detection using sequential analysis of crop images. The suggested technique involves integrating the CNN and GRU, which helps the network develop the capability to learn spatiotemporal information on the crop plant disease. Datasets of sequential crop images that represent the progression stages of the diseases were collected from Plant Village and augmented crop images. The CNN portion of the model is responsible for extracting spatiotemporal characteristics of the diseases, including lesions, discolored parts, and texture changes. It is evident that the suggested Hybrid CNN-GRU method has higher accuracy (94.1%), precision (94.0%), recall (93.9%), and F1-score (94.0%) compared to CNN and CNN-LSTM models. The validation of efficacy and robustness of the suggested approach has been confirmed through standard deviation, paired t-test, and 5-fold cross-validation. Furthermore, the method demonstrated high scalability and strong tolerance to variations in light intensity, noise, and images from the fields of crop plants.