RiceLeafClassifier-v1.0: A Quantized Deep Learning Model for Automated Rice Leaf Disease Detection and Edge Deployment

作     者：Oluwaseun O. Martins Christiaan C. Oosthuizen Dawood A. Desai 

作者机构：Department of Mechanical and Mechatronics Engineering Faculty of Engineering and the Built Environment Tshwane University of Technology Pretoria South Africa Department of Mechatronics Engineering Faculty of Engineering Federal University Oye-Ekiti Ikole Ekiti State Nigeria 

出 版 物：《Engineering Reports》 (Eng. Rep.)

年 卷 期：2025年第7卷第6期

基　　金：Tshwane University of Technology  TUT 

主　　题：convolutional neural networks deep learning edge AI rice leaf disease detection RiceLeafClassifier-v1.0 smart agriculture 

摘      要：Rice diseases critically threaten global food security, necessitating rapid, accurate detection methods. This study presents RiceLeafClassifier-v1.0, a lightweight quantized convolutional neural network (CNN) that classifies five rice leaf conditions: blast, bacterial blight, brown spot, healthy, and red stripe, with high accuracy and real-time performance. To improve generalization, the model was trained on 2807 images, 1144 field-collected and 1663 public. Training enhancements included data augmentation, dropout, dynamic learning rate scheduling, and early stopping. Unlike previous transfer learning approaches, RiceLeafClassifier-v1.0 was built from scratch to retain fine visual features while remaining efficient. Quantization reduced model size from 78.03 to 6.51 MB, enabling deployment on edge devices like the Raspberry Pi 4. Statistical validation ( p 0.05) confirmed that RiceLeafClassifier-v1.0 outperforms VGG-16, VGG-19, and ResNet-50, achieving a classification accuracy of 92% compared to 49% (VGG-16), 48% (VGG-19), and 44% (ResNet-50). Post-training quantization further improved accuracy from 92% to 94% ( p = 0.0165) while reducing memory usage by 68% (from 82.14 to 26.24 MB, p 0.0001). Additionally, inference time per image was significantly lower at 2.28 ± 0.35 s for the quantized model compared to 0.01 ± 0.01 s for the standard model ( p 0.0001), demonstrating substantial gains in efficiency. Despite some limitations, including dataset bias and sensitivity to extreme conditions, the model shows very strong and highly promising potential for real-time disease monitoring in precision agriculture. Future work will expand the dataset, adopt advanced optimization techniques, and integrate IoT systems to support smallholder farmers in reducing crop losses and boosting food security.