Water content is a critical factor affecting the preservation of paper cultural relics. To establish a rapid, non-destructive method for detecting the moisture content of paper artifacts, this study focuses on four-foot single-layer Xuan paper made of cotton. We utilized near-infrared (NIR) spectrometry combined with chemometrics for non-destructive moisture detection. Seven different humidifying salts were placed in a sealed environment box to create humidity conditions ranging from 37% to 97% relative humidity (RH). The Xuan paper samples were equilibrated in this controlled environment for seven days. The water content of the samples was measured to range between 6.35% and 15.55% using the drying method. NIR spectra were collected over the range of 900 to 1 700 nm. The raw spectral data were divided into 168 training sets and 42 validation sets using the spectral-distance joint method (SPXY) at a ratio of 4∶1 for a total of 210 samples. The data were preprocessed using Standard Normal Variate (SNV), Baseline Correction (BC), and normalization, both individually and in combination. Feature bands were selected using Successive Projections Algorithm (SPA) and Competitive Adaptive Reweighted Sampling (CARS). Subsequently, linear partial least squares regression (PLSR) models were established for the full spectrum and selected feature bands, as well as a nonlinear double-layer backpropagation neural network (DL-BPNN) model. The results indicated that the best prediction model for the full spectrum was SNV-PLSR, with a root mean square error (RMSEP) of 0.644 5 and a coefficient of determination ($R_{\text{p}}^2$) of 0.928 3. For the feature bands, the original spectrum-CARS-PLSR model performed best, with an RMSEP of 0.570 7 and an $R_{\text{p}}^2$ of 0.943 8. Among the DL BPNN models, the WT-Normalize-CARS-DL-BPNN model yielded the best results, with an $R_{\text{p}}^2$ of 0.942 4 and an RMSEP of 0.577 6. Comprehensively comparing the prediction effects of the three models, the original spectrum-CARS-PLSR model exhibits the best prediction ability, indicating that the CARS feature extraction method effectively retains important features while eliminating redundant information. This study confirms the feasibility of using NIR spectroscopy for non-destructive moisture content detection in Xuan paper, establishes therelationship between NIR spectra and moisture content, and provides a reliable technical means for measuring the moisture content of paper cultural relics in China.