A CNN-based self-supervised learning framework for small-sample near-infrared spectroscopy classification.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Near-infrared (NIR) spectroscopy offers non-destructive, rapid analysis but faces challenges with complex data preprocessing and feature selection.
• Deep learning methods for spectral analysis typically require large labeled datasets, limiting their application in scenarios with limited samples.

Purpose of the Study:

• To develop a self-supervised learning (SSL) framework using convolutional neural networks (CNNs) to improve spectral analysis performance, particularly for small sample sizes.
• To address the limitations of traditional spectral analysis and data-hungry deep learning models in scenarios with insufficient labeled data.

Main Methods:

• Proposed a two-stage SSL framework involving pre-training on pseudo-labeled data to learn intrinsic spectral features, followed by fine-tuning on a small labeled dataset.
• Utilized convolutional neural networks (CNNs) as the core architecture for feature extraction and model training.

Main Results:

• Achieved high classification accuracy (99.12%) on a custom dataset of three tea varieties.
• Demonstrated superior performance over traditional machine learning methods on three public datasets, with accuracies reaching up to 99.89%.
• Showcased the significant contribution of the pre-training stage, leading to accuracy improvements of up to 10.41%.

Conclusions:

• The proposed SSL-CNN framework effectively enhances spectral analysis for small sample sizes, overcoming limitations of existing methods.
• This approach provides a robust and viable solution for accurate spectral data analysis in data-scarce environments.
• Highlights the potential of self-supervised learning for advancing applications in spectroscopy and related fields.