Enhanced rice leaf disease classification via contour-driven segmentation and optimized deep transfer learning architectures.

Abstract

Pakistan is the fourth-largest rice producer and the fifth-largest exporter worldwide. Timely disease detection remains challenging due to the scale of cultivation and reliance on manual monitoring. Developing reliable, ongoing computerized systems for plant health management is essential for efficient disease control. A deep learning approach is used as the core method to identify diseases in rice leaves. This methodology employs a range of advanced deep learning architectures to achieve top-tier feature extraction and classification. The publicly available rice leaf disease dataset on Zenodo supports research reproducibility and data transparency. We systematically process a balanced dataset of 1914 image samples using Python with TensorFlow and a GPU to enable high-speed computation for large-scale image processing. This study conducts a systematic comparative evaluation of five deep transfer learning architectures (InceptionV3, DenseNet201, ResNet152V2, EfficientNetV2L and MobileNetV2) trained independently. The base backbone models are then integrated with guided GrabCut segmentation with contour-detection method for interpretable disease localization. In this work, the methods of segmentation by GrabCut and contour detection are introduced to make the results of the study easier to interpret and explain the disease areas, but the final classification outcomes are obtained only on the basis of the underlying deep transfer learning models. As a result, infected leaf areas can be identified more effectively, allowing for better understanding and explainable of the disease.To enhance interpretability, GrabCut segmentation and contour detection are applied as post-hoc visualization techniques to highlight diseased regions corresponding to CNN predictions. These techniques do not influence the classification training process. All five models InceptionV3, DenseNet201,ResNet152V2,EfficientNetV2L and MobileNetV2 demonstrated their effectiveness in detecting rice diseases during training, validation, and testing phases, with models trained over 30 epochs. The training methods and accuracy rates of the models were compared during validation and final testing. InceptionV3 demonstrated the most moderate performance of 98.80% training, 98.44% validation, and 98.43% test accuracy, which means that it has strong generalization and consistent learning behavior. The performance of very high-density networks such as DenseNet201 (98.72% train, 98.43% val, 98.43% test), ResNet152V2 (99.02% train, 99.22% val, 97.39% test), EfficientNetV2L model accuracies (39.01% train, 48.70% val, 44.50% test) also showed competitive results, which validated the effectiveness of deep transfer learning in the classification of rice leaf disease, while MobileNetV2 model accuracies (98.09% train, 98.18% val, 96.87% test) indicate that a lightweight model can still achieve reliable classification performance with lower computational complexity. In general, the comparative analysis defines InceptionV3 as the most stable and efficient model in the framework proposed. These results illustrate InceptionV3 superior generalization ability, supported by explainable methods for improved feature localization, confirming the viability of transfer learning for accurate and practical rice disease detection using GrabCut segmentation and contour detection technique. The complete implementation code and data used for the research experimentation is publicly available at https://github.com/ummershakeel03/Rice-Leaf-Diseases-Classification for reproducibility and reuse.

Plant phenotyping relevance

イネ葉の病徴領域を画像から分類・局在化する深層学習ワークフローが研究の中心であり、GrabCut・輪郭検出と複数モデルの比較評価を含むため、植物病害状態の画像ベース表現型計測として採用。

Code and data availability

The paper explicitly states that the complete implementation code and the rice leaf disease image dataset (1914 samples) used in this study are publicly available: code on the authors' GitHub repository and the dataset on Zenodo (DOI 10.5281/zenodo.15817084). Both are paper-specific, public, and actionable.