Classification of rice varieties using RGB images integrated with image-derived geometric traits based on deep learning models
Year: 2026
Type: article
Abstract: Deep learning (DL) offers a promising approach for accurate, scalable plant classification. However, most DL studies on rice depend on seed images. In this study, we assessed leaf-based classification of nine rice varieties using red, green, and blue (RGB) images and geometric traits derived from these images at the vegetative stage, creating an image-guided feature integration model. The classification strategies included traits-based, RGB image-based, and combined RGB image and traits-based methods. For geometric trait measurements, six traits were calculated with the OpenCV module in Python. Classification was conducted at six growth stages: 13, 21, 23, 29, 33, and 35 days after sowing (DAS). Traits-based classification was performed using random forest and support vector machine models. For RGB image-only classification and RGB images combined with trait classification, four convolutional neural networks (CNNs): EfficientNetV2, InceptionV3, ConvNeXt-Tiny, and ResNet-50 were evaluated along with an Ensemble model of ResNet and EfficientNet. Traits-based classification achieved less than 0.5 accuracy at all stages. For RGB image-based evaluation, EfficientNetV2, ResNet-50, and ConvNeXt-Tiny gave an accuracy of 0.76 at 29 DAS, with the lowest being 0.43 for InceptionV3 at 13 and 23 DAS, and the highest 0.80 for Ensemble at 29 DAS. Similarly, for the trait integration method, the Ensemble model achieved the highest accuracy, 0.82, at 29 DAS. The highest-misclassified varieties (Haepum misclassified as Chindeul 41 times) showed mean scaled trait values similar to those of true Chindeul, indicating that morphological similarity was associated with higher misclassification, whereas divergence was associated with improved classification. The gradient-weighted class activation mapping (Grad-CAM) visualization for the top-performing classification model showed that the heatmap concentrated on the leaf cluster and canopy structure, suggesting that the plants’ geometric traits also played a key role in distinguishing the varieties. These findings demonstrate that CNNs provide a feasible approach for varietal classification. Future work will extend this framework to field conditions and integrate remote sensing and hyperspectral imaging to enhance large-scale precision phenotyping.
Source: Frontiers in Plant Science
Institutions: Kyungpook National University, University of Kinshasa
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7234167477
