Improving rice disease diagnosis with a deep learning approach using a CNN trained from scratch

Abstract

Rice is a crucial staple food that sustains millions of people globally; however, its productivity is continuously threatened by various diseases. In recent years, artificial intelligence (AI) and computer vision have emerged as powerful tools for improving the diagnosis and classification of crop diseases, thereby reducing reliance on manual inspection. Despite their success, transfer learning (TL) models based on convolutional neural networks (CNNs) have faced challenges when dealing with limited datasets or closely related image classes. This study aimed to develop and evaluate a CNN model trained from scratch for the classification of rice diseases and to compare its performance with two popular transfer learning models, VGG16 and InceptionV3. A CNN architecture comprising 22 layers was designed and trained using a dataset containing images of eight distinct rice disease classes. The model’s performance was compared to that of the transfer learning models under identical experimental conditions, using accuracy and F1-score as key performance metrics. The proposed CNN model demonstrated superior classification performance, achieving an accuracy of 95%, which significantly outperformed InceptionV3 (73%) and VGG16 (71%). Additionally, the CNN model recorded higher F1-scores across all classes (ranging from 0.91 to 0.99) compared to the TL models (ranging from 0.59 to 0.87). The results confirmed that a CNN model trained from scratch could outperform traditional transfer learning models in rice disease classification, particularly when sufficient data was available. These findings highlighted the potential of custom-built CNN architectures in enhancing disease diagnosis accuracy and suggested further improvements with expanded datasets and additional training.