Quantitative analysis of genetically modified maize based on terahertz spectroscopy and DeepSpectra models

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Widespread cultivation of genetically modified (GM) maize necessitates reliable detection methods.
• Accurate quantification of GM components is critical for regulatory and commercial purposes.

Purpose of the Study:

• To develop and evaluate end-to-end DeepSpectra models for classifying GA21 genetically modified maize.
• To assess the performance of different spectral preprocessing techniques and comparative models.

Main Methods:

• Acquisition of spectral data using terahertz time-domain spectroscopy.
• Application of outlier removal (isolation forest) and various preprocessing techniques (Savitzky-Golay smoothing, standard normal variate, baseline correction, FD, and second derivative).
• Development of three DeepSpectra models (V1-V3) and comparison with Support Vector Machine (SVM), Random Forest (RF), and 1D Convolutional Neural Network (CNN).

Main Results:

• The DeepSpectraV2 model, combined with first derivative (FD) preprocessing, achieved the highest classification accuracy of 96.56%.
• DeepSpectraV2 significantly outperformed comparative models (SPA-GS-SVM, SPA-GS-RF, 1D CNN) by 1.52% to 15.52%.
• The study demonstrates the effectiveness of DeepSpectra models for GM maize detection across various GM concentrations.

Conclusions:

• DeepSpectra models offer a novel and effective approach for rapid, non-destructive detection of GM crops.
• Optimized spectral preprocessing, particularly FD, enhances classification accuracy in GM component analysis.
• This methodology provides a robust solution for identifying GM maize with high precision.