Water content of rock is an important index to affect the physical, chemical and mechanical properties of rock. In geotechnical engineering, tunnel engineering and other fields, water content is the key factor to induce disaster and disease. Compared with the traditional method, the determination of rock water content by using the feature of NIR spectrum (NIRS) has obvious advantages of nondestructive and quantitative analysis, and the difficulty and key is the feature selection of NIR spectrum. In order to solve this problem, laboratory experiments were carried out to study the feature selection of near infrared spectra of rock under different water content. The Filter method of feature selection, using the inherent characteristics of the sample data, evaluates the importance of the feature, enhances the correlation between the feature and the class, and reduces the correlation between the features, so it has the advantages of low complexity, being intuitionistic, high efficiency and strong universality and accords with the characteristics of the data studied in this paper. Therefore, this paper selects the Filter type dependency metric for feature selection. In the laboratory experiment, 11 kinds of sandstone samples with different moisture content were prepared, and 44 NIR spectra were collected respectively at 4 test points on the front, behind, left and right sides. Then, the first derivative method was used to preprocess the spectrum. Based on this, the spectral characteristics were analyzed at 1 400 and 1 930 nm, and six initial characteristic variables (the peak area, peak height, width of half height, width of left shoulder, width of right shoulder, the ratio of the width of the left shoulder to the width of the right shoulder )were extracted respectively. Considering the different dimensions and variation range of the six initial characteristic variables, the original data were normalized to eliminate the influence of different dimensions and variation ranges. And then, according to the principle of independent variable selection, redundant variables with strong linear correlation between independent variables were removed. Then, used the statistical correlation coefficient in the dependency metric as the measure of correlation degree, and the correlation among the initial characteristic variables and the correlation between the initial characteristic variables and water content were analyzed. The optimal characteristic variables at two strongly correlated spectral segments were obtained. Finally, multiple regression models were constructed at the strong correlation spectral segments, and the models were tested and analyzed. The results showed that: (1) the characteristics of the near-infrared spectral absorption peaks around the wavelengths of 1 400 and 1 930 nm are significantly correlated with the rock water content; (2) the peak height, right half width and left half width at the wavelength of 1 400 nm have linear correlation with the water content obviously, and the peak height and right half width at the wavelength of 1 930 nm also have linear correlation with the water content obviously; (3) the multiple linear regression model can accurately express the correlation between the water content and the near-infrared spectrum, and the model can be used to predict the water content of water-bearing rock based on the characteristics of near-infrared spectrum. It provides basic modeling data for dynamic monitoring and evaluation of rock water content by using near infrared spectrum analysis technology.