VIS-NIR-SWIR Hyperspectral Imaging and Advanced Machine and Deep Learning Algorithms for a Controlled Benchmark of...

Area of Science:

• Agricultural Science
• Spectroscopy
• Biotechnology

Background:

• Conventional seed identification methods are destructive, labor-intensive, and difficult to scale.
• Accurate seed accession identification is crucial for germplasm conservation, traceability, and breeding programs.

Purpose of the Study:

• To evaluate the efficacy of visible-near-infrared-shortwave infrared (VIS-NIR-SWIR) hyperspectral imaging (HSI) for classifying grain-legume seed accessions.
• To compare the performance of different spectral analysis methods, including principal component analysis (PCA), wavelength selection algorithms, and machine learning models.

Main Methods:

• Utilized VIS-NIR-SWIR HSI (449.54-2399.17 nm, 563 bands) to capture spectral data from 3200 seeds across 32 grain-legume accessions.
• Employed principal component analysis (PCA) for spectral variance analysis and ReliefF for feature selection.
• Applied linear discriminant analysis (LDA), support vector machines (SVM), and deep learning models (MLP_Wide) for seed classification.

Main Results:

• PCA effectively captured spectral variance, with the first three components explaining 97.42%.
• Linear discriminant analysis achieved 96.35% balanced accuracy using full-spectrum data.
• Deep learning models, particularly MLP_Wide, reached 84.90% test accuracy, demonstrating the potential of advanced algorithms.

Conclusions:

• VIS-NIR-SWIR HSI is a viable, non-destructive technique for accurate legume seed accession identification.
• Visible-domain spectral contrasts are key drivers for discrimination in compact representations.
• Full spectral context is important for distinguishing confusable accessions and guiding future sensor development.