Rice Disease Detection Using an Optimized Modified Lightweight Convolutional Neural Network

Abstract

Rice is an important staple food crop in the world, but the yield is dramatically lowered by fungal diseases including brown spot, leaf blast and neck blast. Thus, disease detection algorithms that are not only computationally efficient but also accurate enough are crucial for the real-time application of precision agriculture systems. In this paper, an Optimized Convolutional Neural Network from Modified Lightweight Weight Convolutional Neural Network (MLWCNN) is proposed to multi-class rice disease classification for the purpose of achieving high accuracy with low computational complexity. The proposed MLWCNN is tested on a Kaggle public data set, which comprises the rice leaf disease dataset as well as sub-images of healthy and diseased leaves with 3-Fold and 5-Fold cross-validation for robustness and unbiased performance evaluation. Experimental results show that they converge consistently well on objective and generalize well under various validation settings. With the strictest evaluation setup (i.e., 5-Fold cross-validation and a training stage of 30 epochs) the novel model demonstrated a validation accuracy just under around 94%. A comparison with state-of-the-art deep learning networks such as wide VGG16, InceptionV3, ResNet50, ResNet152 SqueezeNet and MobileNetV2 on typical datasets proves that the proposed MLWCNN achieves competitive classification accuracy with a higher speedup. Analyzing the class‐wise performance using confusion matrix, balanced prediction characteristics throughout all disease categories are observed. Moreover, computational cost analysis demonstrates that the proposed model needs much less floating-point operations (0.000095 GFLOPs) than deep models, which is ideal for deployment in resource-limited mobile, edge and embedded agricultural applications. The experimental results demonstrate that the MLWCNN achieved high accuracy, thereby providing a practical and scalable solution for automated rice disease detection in precision agriculture.

Code and data availability

The paper uses a public Kaggle rice leaf disease dataset (5,256 images, 4 classes) for its MLWCNN experiments, but no authors' code, trained model, or dataset URL is provided. The only URLs appearing in the text belong to cited references, not to this paper's own assets, and none match the allowed URL set, so no paper-

No evidence-backed public reproduction asset is currently recorded.