Monte Carlo band selection with residual network for noise-robust near-infrared spectral quantitative analysis.

A noise-robust Monte Carlo continuous band selection and light residual network (MCC-LRNet) framework is proposed for robust quantitative analysis of industrial near-infrared (NIR) spectra. The framework integrates three innovations, including a parameterized preprocessing module for noise and outlier removal; a Monte Carlo Continuous (MCC) band selection algorithm that extracts informative contiguous wavelength intervals; and a custom Residual Network (LRNet) for regression. Experiments on four industrial datasets (two tobacco and two beer, covering 18 indices) show that compared with PLS, SVM, and ResNet18, the complete MCC-LRNet framework achieves the lowest RMSE on 10 of the 18 indices, with RMSE ranging from 0.065 (Nitrogen) to 0.750 (Total sugar) and R2 values ranging from 0.882 (K) to 0.988 (Moisture). The proposed framework can serve as a competitive baseline for spectral quantitative analysis.