Near-infrared (NIR) spectroscopy combined with multivariate calibration methods is widely used for the quantitative analysis of complex samples. However, the high-dimensional redundancy of spectra may compromise model predictive accuracy, making it necessary to select variables before modeling. The beluga whale optimization (BWO) algorithm is known for its fast convergence speed, high accuracy and few parameters. The present study employed the discretized BWO (DBWO) algorithm in conjunction with partial least squares (PLS) for spectral quantitative analysis of complex samples. After the optimal number of iterations and transfer function were determined, the PLS models were established based on the randomization test (RT), uninformative variable elimination (UVE) and Monte Carlo uninformative variable elimination (MC-UVE). The predictive performance of DBWO-PLS was compared with full-spectrum PLS, RT-PLS, UVE-PLS and MC-UVE-PLS using wheat, tablet and cocoa bean samples. The results show that all four variable selection methods enhanced model prediction accuracy, with the DBWO-PLS model notably achieving superior performance.