Curvelet Decomposition-Based Tri-Branch Coupling Network for Hyperspectral Unsound Maize Seeds Identification

Area of Science:

• Agricultural Science
• Computer Vision
• Data Science

Background:

• Accurate and non-destructive classification of maize kernels is crucial for seed screening and quality control.
• Current hyperspectral imaging methods using Mamba architectures have limitations in time-frequency analysis and multimodal fusion.
• Traditional methods often require extensive spectral preprocessing, potentially introducing errors and reducing model robustness.

Purpose of the Study:

• To develop a novel cross-modal classification framework for maize kernel analysis.
• To enhance feature extraction and deep fusion by integrating hyperspectral data and visible-light images.
• To overcome limitations in existing methods regarding feature fusion and spectral preprocessing.

Main Methods:

• Proposed a cross-modal classification framework named CD-TriMamba.
• Designed an innovative feature extraction module with Spectral Curvelet Convolution (SCC) for hyperspectral data and Curvelet-Decomposed Convolution (CDC) for spatial modeling.
• Implemented a feature rearrangement mechanism and a ConvNeXt-guided tri-branch cross-fusion structure (TriMamba) for deep feature integration.

Main Results:

• The CD-TriMamba model achieved outstanding performance in maize kernel seed classification.
• Attained an accuracy (Acc) of 99.2% and a Kappa value of 99.1%.
• Demonstrated the effectiveness of cross-modal feature fusion for enhanced classification.

Conclusions:

• The proposed CD-TriMamba framework effectively integrates hyperspectral and visible-light data for maize kernel classification.
• Cross-modal feature fusion significantly improves classification accuracy and robustness.
• The model shows strong potential for practical applications in seed screening and quality evaluation.