Research on terahertz image analysis of thin-shell seeds based on semantic segmentation

Area of Science:

• Agricultural Science
• Biotechnology
• Imaging Technology

Background:

• Assessing crop seed phenotypic traits is crucial for agricultural innovation and germplasm enhancement.
• Traditional methods struggle with rapid, non-destructive analysis of thin-shelled seeds due to their tough outer layers.
• Internal structures and quality attributes of seeds are difficult to assess without damaging them.

Purpose of the Study:

• To explore the potential of combining terahertz (THz) time-domain spectroscopy and imaging with semantic segmentation models for rapid, non-destructive examination of crop seed traits.
• To evaluate the performance of deep learning models (SegNet and DeepLab V3+) for automatic tissue segmentation in watermelon seeds.
• To provide precise phenotypic trait analyses for seeds with thin shells.

Main Methods:

• Acquisition and reconstruction of THz spectral images from 120 watermelon seeds using a transmission imaging modality and correlation coefficient method.
• Application of deep learning-based SegNet and DeepLab V3+ models for automatic segmentation of seed tissues (outer layer and inner kernel).
• Comparative analysis of SegNet and DeepLab V3+ performance in terms of speed and accuracy.

Main Results:

• DeepLab V3+ significantly outperformed SegNet in speed and accuracy for semantic segmentation of watermelon seed tissues.
• DeepLab V3+ achieved 96.69% pixel accuracy and 91.3% intersection over union for outer layer segmentation, with high accuracy for inner kernels.
• The models effectively distinguished between the seed coat and kernel, enabling precise phenotypic trait analysis.

Conclusions:

• Deep learning, particularly DeepLab V3+, offers a proficient method for rapid, non-destructive phenotypic trait analysis of thin-shelled crop seeds.
• THz imaging combined with semantic segmentation advances agricultural research and practices by enabling detailed seed examination.
• This approach holds significant potential for breeding innovations and germplasm enhancement in agriculture.