Classification of maize seed hyperspectral images based on variable-depth convolutional kernels.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate seed classification is crucial for germplasm utilization and breeding efficiency.
• Manual classification is labor-intensive and error-prone.
• Convolutional Neural Networks (CNNs) show promise but struggle to integrate spectral and textural data effectively.

Purpose of the Study:

• To develop a novel CNN architecture for enhanced hyperspectral seed classification.
• To simultaneously extract spectral and textural features from hyperspectral images.
• To improve the accuracy and robustness of corn seed classification.

Main Methods:

• Proposed a Variable-Depth Convolutional Neural Network (VD-CNN) architecture.
• Employed adaptive kernel depth modulation for spectral feature extraction.
• Utilized hierarchical convolutional operations for textural pattern capture.
• Trained and evaluated 12 models with varying kernel depths on corn and rice seed datasets.

Main Results:

• The VD-CNN achieved optimal performance with a kernel depth of 15, reaching 98.65% training and 96.97% test accuracy for corn seeds.
• Demonstrated superior generalization on a public rice seed dataset, outperforming benchmarks by 3.14%.
• Validated the model's robustness across different crop species and imaging conditions.

Conclusions:

• The VD-CNN effectively integrates spectral and textural information for superior hyperspectral seed classification.
• The proposed architecture offers a significant advancement for seed classification and other agricultural hyperspectral imaging applications.
• Highlights the potential of adaptive deep learning models in precision agriculture.