RiPa-Net: Recognition of Rice Paddy Diseases with Duo-Layers of CNNs Fostered by Feature Transformation and Selection

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Rice paddy diseases cause significant crop yield and quality losses.
• Accurate and timely disease identification is crucial for effective management.
• Deep learning (DL) offers potential for automated disease recognition but faces generalization challenges due to subtle inter-class variations.

Purpose of the Study:

• To develop an automated system for accurate identification and categorization of nine rice paddy diseases and healthy crops.
• To address the generalization limitations of existing DL models in paddy disease recognition.
• To improve the efficiency and accuracy of paddy disease detection.

Main Methods:

• Proposed a novel pipeline, RiPa-Net, utilizing three lightweight Convolutional Neural Networks (CNNs).
• Extracted deep features from two distinct layers of each CNN.
• Integrated spectral-temporal information using Dual-Tree Complex Wavelet Transform (DTCWT) on first-layer features.
• Reduced feature dimensionality using Principal Component Analysis (PCA) and Discrete Cosine Transform (DCT).
• Combined spatial features from the second layer with fused time-frequency features from the first layer.
• Implemented a feature selection process to retain impactful features and reduce complexity.

Main Results:

• The RiPa-Net pipeline achieved a high recognition accuracy of 97.5% using a cubic Support Vector Machine (SVM) with 300 selected features.
• Combining deep features from multiple layers of lightweight CNNs enhanced recognition accuracy.
• Incorporating spatial-spectral-temporal information improved model performance.
• The proposed method demonstrated competitive performance compared to existing research.

Conclusions:

• The RiPa-Net pipeline effectively identifies and categorizes rice paddy diseases with high accuracy.
• Multi-layer feature fusion and the inclusion of spectral-temporal information are key to improving DL model generalization in this domain.
• The developed system offers a promising solution for automated paddy disease management.