Efficient Deep Learning Technique for Rice Leaf Disease Detection: A Light-Weight Model Approach

Rice is one of the major contributing crops to the total economy of Bangladesh. It is the primary source of carbohydrates, especially for subcontinental people. Diseases like Blast, Bacterial Leaf Blight, Tungro, and Brown Spot can significantly decrease the production of rice and degrade the quality of the produced crop. These diseases impose an intolerable threat to the global food chain, availability, and security system which causes serious yield loss. Therefore, the early detection of rice leaf diseases for effective production control management is crucial to minimize loss. This study proposes a novel resource-efficient, low-complexity lightweight model for distinguishing these diseases from healthy data. The proposed pipeline mitigates the data imbalance issue by employing the SMOTE over-sampling technique. The performance of the proposed Light-Weight Dense Model is evaluated by comparing it with benchmark deep learning models including Resnet50, DenseNet201, VGG19, and MobileNet. It surpasses these traditional models significantly with 99.68% accuracy and secures an AUC score of 1.00. In addition, this study thoroughly assesses the performance of the proposed model across various image shapes, consistently demonstrating noteworthy results. Furthermore, stress testing is carried out to evaluate the adaptability and reliability of the proposed model on various datasets. The results emphasize the effectiveness of the model, outperforming other deep learning models with a 100% accuracy rate (AUC = 1.00). The findings demonstrate the development of a generalized pipeline for the early detection of rice leaf diseases using a diverse training dataset of varying data sizes.