Classification of rice plant diseases using efficient DenseNet121

Abstract

Abstract Agriculture and global food security are critically dependent on accurate and timely identification of plant diseases and pests. Traditional approaches to disease identification rely heavily on visual inspection and expert knowledge, which frequently lack the accuracy, speed, and scalability needed to address growing agricultural challenges. Early and precise disease detection enables proactive interventions that can prevent widespread crop damage and reduce excessive pesticide use, thereby supporting sustainable agricultural practices. Artificial intelligence, particularly deep learning methods, has emerged as a transformative solution for automated plant disease diagnosis. Convolutional neural networks (CNNs) have demonstrated remarkable capabilities in image classification tasks, evolving from individual architectures to sophisticated ensembles and transferring learning models. However, existing CNN-based research on rice disease identification has typically focused on a limited number of disease classes, restricting their practical applicability in real-world agricultural settings. This study addresses these limitations by implementing DenseNet121, an advanced CNN architecture known for its efficient feature reuse and gradient flow, for comprehensive rice disease classification. We utilized a dataset comprising seven of the most common rice diseases, significantly expanding the scope beyond previous studies. The model employs transfer learning with pre-trained ImageNet weights and is optimized using the Adam optimizer with carefully tuned hyperparameters. The experimental evaluation on an independent test set demonstrates that our proposed model achieves an overall accuracy of 97.9%, with individual disease classification accuracy ranging from 94% to 99.67%. The model exhibits balanced performance across multiple metrics, including precision (96.2%), recall (97.97%), and F1-score (97%), confirming its robustness and generalizability. These results establish DenseNet121 as a highly effective framework for automated rice disease diagnosis, offering a practical tool for enhancing agricultural productivity and food security.

Plant phenotyping relevance

イネ葉の画像から病害状態を分類する深層学習手法が研究の中心であり、独立テストセットによる性能評価も実施しているため、植物病害フェノタイピング手法として収載する。

Code and data availability

The paper's rice disease classification experiments use the public Kaggle Paddy Disease Classification dataset (8030 images, 7 disease classes), explicitly cited and linked by the authors in the Data Availability Statement. No author code or trained model checkpoints are disclosed.