Based on the feature of the (NIR) spectra, this paper analyses the method of wavelength selection using the partial least squaresc (PLS) regression coefficients and points out the existing problems, then proposes a new method for selecting wavelengths. It normalizes the PLS regression coefficients into the probability of the selected corresponding wavelengths, then a Monte-Carlo simulation based on the aforementioned probability is calculated. Some PLS models are constructed and evaluated using different random wavelengths combinations. The model with minimum predictive error is retained and the corresponding wavelength combinations are selected. This procedure can be iterated using the previous selected wavelengths to select fewer and fewer wavelengths. This method is tested on 3 NIR datasets and compared with the PLS-based uninformative variable elimination (UVE-PLS) and genetic algorithm (GA). Experimental results show that this method could select fewer wavelengths without sacrificing the complexity and predictive ability of the PLS model and could effectively improve the accuracy and stability of the wavelength selection.