A feature decoupling and attention-based method for small-sample NIR spectral classification: application to...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Near-infrared (NIR) spectroscopy is crucial for identifying the geographical origin of botanicals like Astragalus membranaceus.
• Challenges in NIR spectroscopy include limited training samples and spectral collinearity, which hinder classification accuracy.
• Effective methods are needed to enhance the performance of NIR-based origin identification, especially in data-scarce scenarios.

Purpose of the Study:

• To develop a small-sample NIR classification method for geographical origin identification of Astragalus membranaceus.
• To leverage masked autoregressive flow (MAF) for spectral data transformation and the Simple, Parameter-Free Attention Module (SimAM) for feature enhancement.
• To improve class separability using a prototype-constrained loss function.

Main Methods:

• Implemented MAF to reduce spectral redundancy and inter-variable dependence via latent-space transformation.
• Integrated SimAM to enhance discriminative latent features without adding trainable parameters.
• Incorporated a prototype-constrained loss function to boost class separability.

Main Results:

• The proposed MAF-SimAM method achieved an average accuracy of 98.50% ± 0.33% on an independent test set with 600 samples from five geographical origins.
• In a few-shot learning scenario (20 training samples/class), MAF-SimAM reached an accuracy of 94.00% ± 1.71%.
• The method outperformed traditional techniques like partial least squares discriminant analysis, autoencoder-multilayer perceptron, and 1D convolutional neural networks.

Conclusions:

• The developed MAF-SimAM framework offers accurate and stable NIR spectral classification for geographical origin identification, particularly under limited-sample conditions.
• This approach effectively addresses challenges posed by small sample sizes and spectral collinearity in NIR analysis.
• The findings demonstrate the potential of advanced deep learning techniques in botanical origin authentication.