Quantitative analysis method for near-infrared spectroscopy based on DeepSpectra model

Traditional near-infrared (NIR) spectroscopy quantitative analysis suffers from fragmented preprocessing and modeling, relying on manually designed algorithms with limited generalization. This paper proposes an analysis method based on the deep spectral model DeepSpectra (DS), integrating preprocessing and predictive modeling. DS automatically performs noise suppression, baseline correction, and feature extraction via convolutional layers, batch normalization, and attention mechanisms, eliminating manual intervention. Experiments on 450 milk samples show that DS preprocessing achieves R² values of 0.978, 0.910, and 0.863 for fat, protein, and lactose prediction, outperforming traditional methods like SNV and MSC. The DS end-to-end model exhibits the best performance in modeling, with a fat prediction R² of 0.981 and RMSE of 0.218, significantly surpassing combined models such as DS-PLS and DS-SVM. This method improves quantitative accuracy, simplifies workflows, and provides a reference for intelligent NIR spectroscopy applications.