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

With the widespread cultivation of genetically modified (GM) maize, accurate detection of GM components has become critical. This study developed three end-to-end DeepSpectra models (V1-V3) to classify GA21 maize at six GM concentrations (blank, levels 1-5). Spectral data were acquired using a terahertz time-domain spectroscopy system. Outliers were removed using an isolation forest, and the spectral data were preprocessed with Savitzky-Golay smoothing, standard normal variate transformation, baseline correction, first derivative (FD), and second derivative. Three comparison models were constructed, a support vector machine (SPA-GS-SVM) and a random forest (SPA-GS-RF) based on successive projection algorithms (SPA) and grid search (GS), as well as a one-dimensional convolutional neural network. Experimental results showed that the DeepSpectraV2 model with FD preprocessing achieved the best classification accuracy of 96.56%, outperforming the best comparative models by 1.52 to 15.52% across other preprocessing methods. This study presents a novel approach for the rapid non-destructive detection of GM crops.