Double Transfer Learning-Based Capsule Network for Multi-Crop Plant Disease Classification Using Heterogeneous Leaf Image Datasets

Abstract

Abstract Crop disease is a major worldwide problem in agricultural production and food security, adversely affecting yield and quality for a variety of plant species. To overcome these drawbacks, this research provides a Double Transfer Learning-based Capsule Network (DTL-CapsNet) approach for automated plant disease classification with multiple crops. Based on the image pre-processing, segmentation, double transfer learning, and Capsule Networks technologies, the proposed framework extracts discriminative features of the diseases and maintains spatial relations between the leaves symptoms effectively. This double transfer learning approach involves extracting general visual features from pre-trained deep learning models and then fine-tuning these features to classify plant diseases. Capsule Networks then leverage the visual similarity of disease patterns to make the recognition more robust, while simultaneously adding hierarchical part–whole relationships in leaf structures, thereby improving the feature representation. Experiments were performed on a heterogeneous data set consisting of 21,927 leaf images belonging to 17 different healthy and diseased classes of apple, chilli, cotton, corn and potato crops. The experimental results proposed DTL-CapsNet framework is more accurate compared to the traditional CNN-based models and conventional transfer learning models. The proposed method of double transfer learning and Capsule Networks offers an efficient and scalable approach for intelligent plant disease diagnosis, offering significant potential in precision agriculture and real-time crop monitoring systems.

Plant phenotyping relevance

葉画像から植物病害状態を分類する画像・計算手法が研究の中心であり、提案手法の開発と既存モデルとの比較検証が行われているため。

Code and data availability

The paper uses publicly available leaf image datasets (Kaggle-PlantVillage; Mendeley Cotton Leaf Disease Dataset) for its multi-crop disease classification experiments, but provides no authors' public URL for these inputs, no analysis code, and no trained model release. Paper-specific additional data are offered only '

No evidence-backed public reproduction asset is currently recorded.