Machine Learning Analysis of Hyperspectral Images of Damaged Wheat Kernels

Area of Science:

• Agricultural Science
• Plant Pathology
• Remote Sensing Technology

Background:

• Fusarium head blight (FHB), caused by *Fusarium graminearum*, significantly impacts small grain yield and quality.
• Deoxynivalenol (DON) is a harmful mycotoxin produced by *F. graminearum*, posing risks to food safety.
• Accurate and rapid detection of FHB and DON contamination is crucial for the grain industry.

Purpose of the Study:

• To investigate the efficacy of hyperspectral imaging (HSI) for classifying FHB-induced damage in wheat kernels.
• To correlate HSI data with deoxynivalenol (DON) mycotoxin concentrations in wheat.
• To identify optimal machine learning algorithms for HSI-based FHB damage assessment.

Main Methods:

• Wheat kernel samples were analyzed using hyperspectral imaging (HSI).
• Gas Chromatography-Mass Spectrometry (GC-MS) was employed to quantify deoxynivalenol (DON) content for sample classification.
• Machine learning algorithms, including G-Boost and Mask R-CNN, were utilized for image analysis and classification.
• Instance segmentation (Mask R-CNN) was used to isolate wheat kernels for region of interest (ROI) analysis.

Main Results:

• The G-Boost algorithm achieved 97% accuracy in classifying wheat kernels based on FHB severity.
• Mask R-CNN demonstrated high performance in segmenting wheat kernels, with a mean Average Precision (mAP) of 0.97.
• Combined HSI analysis and machine learning correlated imaging data with DON concentration, yielding an R² of 0.75.

Conclusions:

• Hyperspectral imaging (HSI) is a promising non-destructive technique for detecting and quantifying Fusarium head blight (FHB) damage in wheat kernels.
• Machine learning, particularly G-Boost and Mask R-CNN, significantly enhances the accuracy of HSI-based disease assessment.
• This approach shows potential for real-time monitoring of DON contamination in commercial grain handling facilities, improving food safety and quality control.