Details

Autor(en) / Beteiligte

• Bathe, Kavita
• Patil, Nita
• Patil, Sanjay
• Bathe, Devanand
• Kumar, Kuldeep

Titel

ConvDepthTransEnsembleNet: An Improved Deep Learning Approach for Rice Crop Leaf Disease Classification

Ist Teil von

• SN computer science, 2024-04, Vol.5 (4), p.436

Ort / Verlag

Singapore: Springer Nature Singapore

Erscheinungsjahr

2024

Link zum Volltext

Quelle

SpringerLink (Online service)

Beschreibungen/Notizen

• Agricultural sector, in association with its allied sectors, plays pivotal role in the progress of a country. In spite of notable contribution, Agricultural sector suffers from several challenges. Crop loss due to diseases is considered as one of the major challenges. Rice is one of the major cereal crops in India. Its production is highly impacted due to crop diseases. Early and accurate detection of rice crop leaf diseases is essential for aforementioned issue. Over the years, conventional methods of crop leaf disease detection are used but with its own limitations. Recent technological advancement in computer vision, deep learning has created new pathways in agricultural sector. Deep learning models require huge data which is quite a great challenge. In such case, building up of a robust deep learning model could work with limited and unbalanced dataset, with good generalization performances. In this study, weighted deep ensemble learning approach is used for performance improvement of rice crop leaf disease classification task. The ensemble method ConvDepthTransEnsembleNet is proposed. To show the effectiveness of proposed work, experiments are conducted on diverse datasets. ConvDepthTransEnsembleNet is a lightweight model which has achieved the accuracy of 96.88% on limited and unbalanced dataset. The experimental results show that proposed model outperforms the individual classifiers based on conventional methods and transfer learning approach in terms of significant reduction in parameters and improved upon generalization performance. The proposed model is highly useful for implementing deep learning models with resources constrained devices.