Comparative Analysis of Machine Learning and Deep Learning Algorithms for Assessing Agricultural Product Quality...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Near-infrared spectroscopy (NIRS) analysis success relies on precise calibration models.
• Shallow learning (SL) algorithms struggle with spectral data complexity and noise, limiting NIRS applications.
• Deep learning (DL) offers potential for improved feature extraction from limited spectral samples.

Purpose of the Study:

• To evaluate the robustness and effectiveness of NIRS calibration models.
• To compare the performance of SL, consensus learning (CL), and DL methods.
• To propose and validate a novel G-CACNN model for NIRS discriminant analysis.

Main Methods:

• Utilized discriminant analysis on wheat kernels and Yali pears datasets.
• Compared partial least squares discriminant analysis (PLS-DA) with DL and CL models.
• Developed a Gramian angular difference field and coordinate attention convolutional neural network (G-CACNN) model.

Main Results:

• DL and CL models showed less sensitivity to spectral preprocessing than SL.
• The proposed G-CACNN model achieved high accuracy (98.48% and 99.39%) in discriminant tasks.
• G-CACNN demonstrated superior robustness and noise resistance compared to other models.

Conclusions:

• Deep learning significantly enhances NIRS analysis accuracy and robustness.
• The G-CACNN model provides a powerful and noise-resistant approach for NIRS applications.
• This study advances NIRS calibration model development for broader applicability.