Plant leaf infected spot segmentation using robust encoder-decoder cascaded deep learning model.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate leaf disease diagnosis is crucial for agricultural productivity and cost reduction.
• Inaccurate segmentation of diseased leaf spots can lead to misclassification of plant diseases.
• Overlapping disease features and dimensions pose challenges for precise segmentation.

Purpose of the Study:

• To propose a novel Deep Encoder-Decoder Cascaded Network (DEDCNet) for precise leaf image segmentation.
• To accurately segment diseased leaf spots and differentiate between similar plant diseases.
• To improve the overall accuracy of leaf disease classification and diagnosis.

Main Methods:

• Developed a DEDCNet model comprising an Infected Spot Recognition Network (ISRN) and an Infected Spot Segmentation Network (ISSN).
• ISRN integrates cascaded Convolutional Neural Networks (CNNs) with Feature Pyramid Pooling for infected spot identification.
• ISSN utilizes an encoder-decoder architecture with multi-scale dilated convolutions for precise segmentation.
• Employed pre-learned CNNs for texture feature extraction and Support Vector Machine (SVM) for disease classification.

Main Results:

• Achieved 94.89% accuracy on the Betel Leaf Image dataset with high precision, recall, and F-score.
• Demonstrated low under-segmentation (6.2%) and over-segmentation (2.8%) rates on the Betel Leaf dataset.
• Attained 96.5% accuracy on the PlantVillage dataset, outperforming existing models.
• Achieved Dice coefficients of 0.9822 and 0.9834 on respective datasets in under 0.1 seconds.

Conclusions:

• The DEDCNet model significantly improves leaf disease segmentation and classification accuracy.
• The proposed method offers greater efficiency compared to existing models for plant disease analysis.
• Accurate segmentation is vital for reliable disease diagnosis, contributing to better agricultural outcomes.