The strength of sedimentary rocks is often affected by water, and the influence degree is different with different water content. Therefore, the measurement of the water content of the rock has important value for the subsequent evaluation of its physical and mechanical properties. In measuring rock water content, traditional methods are often time-consuming and laborious, destroying the integrity of the structure. At present, near-infrared spectroscopy is a very potential method, with real-time and nondestructive advantages. In this paper, sandstone's spectral characteristics and the water content prediction are studied based on near infrared spectroscopy. Firstly, near-infrared spectrum curves of sandstone samples with different saturations were obtained through laboratory tests. Secondly, the first derivative of the original spectral curve is preprocessed to eliminate the influence of noise, environment, and other factors. Thirdly, the spectral characteristic variables of R1 (1 400 nm) and R2 (1 900 nm) were extracted and normalized to eliminate the influence of dimension and domain value. Fourthly, the extracted spectral characteristic variables are analyzed and screened based on the maximum information coefficient; Finally, the self-built BP neural network classifier is used to predict the water content of sandstone. The conclusions are as follows: (1) The near-infrared absorption curve of water-bearing sandstone has obvious absorption peaks near 1 400 and 1 900 nm, the absorption peak is near 1400 nm, the band is relatively broad, the absorption peak is near 1 900 nm, and the band is relatively sharp. As the water content increases, the absorption intensity of each absorption peak is also increasing, which has an obvious correlation and can be used as the main spectrum band for sandstone water content analysis and prediction. (2) According to the calculated maximum information coefficient value, the peak height near 1 400 nm has the strongest correlation with water content, and the peak height near 1900 nm has the strongest correlation with water content. Peak area and peak height near 1 400 nm, peak area, peak height, half-height width, and right shoulder width near 1 900 nm are 6 characteristic variables. The maximal information coefficient value is greater than 0.9, which can be used as characteristic variable for BP neural network to predict sandstone water content. (3) Using the maximum information coefficient to screen out the characteristic variables of the two absorption peaks at 1 400 and 1 900 nm for BP neural network modeling, the accuracy of the training set of the sandstone water content prediction model established by it was 90.3%, and the accuracy of the test set was 83.9%. The method based on near-infrared spectroscopy analysis technology to predict the water content of sand and gravel is feasible.