Soybean seed pest damage detection method based on spatial frequency domain imaging combined with RL-SVM

Area of Science:

• Agricultural Science
• Optical Engineering
• Machine Learning

Background:

• Soybean seed quality is compromised by pests like Riptortus pedestris.
• Current manual inspection methods fail to detect sub-surface pest damage.
• Efficient detection of pest-damaged soybean seeds is crucial for quality control.

Purpose of the Study:

• To develop an automated method for detecting sub-surface pest damage in soybean seeds.
• To combine spatial frequency domain imaging with machine learning for pest damage assessment.
• To improve the accuracy and efficiency of soybean seed quality screening.

Main Methods:

• Acquired soybean optical data using single integration sphere technique and germination vigor index.
• Extracted characteristic wavelengths using successive projections and competitive adaptive reweighted sampling algorithms.
• Utilized spatial frequency domain imaging to obtain sub-surface images and inverted optical coefficients.
• Developed RL-MLR, RL-GRNN, and RL-SVM models for pest damage prediction.

Main Results:

• Spatial frequency domain imaging showed low errors (<15% for absorption coefficient, <10% for reduced scattering coefficient).
• Reinforcement learning enhanced model Macro-Recall metrics by 10%-15%.
• The RL-SVM model achieved a high Macro-Recall of 0.9635 for classifying pest damage levels.

Conclusions:

• Spatial frequency domain imaging is effective for assessing sub-surface optical properties of soybean seeds.
• Reinforcement learning significantly improves the performance of pest damage detection models.
• The RL-SVM model offers a highly accurate and efficient solution for identifying stink bug damage in soybean seeds.