Wood density is an important physical property, which determines the mechanical properties of wood. In recent years, as NIR has the advantages of simple, convenient and fast operation, it has already been used in terms of wood density prediction. However, in practical application, the sample sets shortage, spectral characteristics selection and non-linear fitting inaccuracy still not been solved definitely, and the accuracy of wood density prediction model needs to be further improved. Among all the wood density parameters, the absolute dry density of wood is relatively stable, and the measurement results are relatively accurate. In this paper, the prediction of absolute dry density of oak is studied. By collecting spectroscopy information under different moisture content, a non-linear prediction model of absolute drying density suitable for arbitrary moisture content is constructed. The near-infrared optical fiber spectrometer of INSION Company in Germany was selected, and the spectral information of oak samples with different moisture content was collected by SPEC view 7.1 software. Then, the calibration set and prediction set were divided according to 2∶1 using SPXY sample partition method, and multivariate scattering correction, second derivative spectroscopy and S-G smoothing method were used to reduce the influence of scattered light and high-frequency noise; After that, continuous projection algorithm SPA was used to extract effective wavelength information; finally, a BP network (IPSO-BPNN) was used to establish the correlation between near-infrared spectra and oak absolute dry density under different moisture content, which was optimized by a non-linear weighted particle swarm optimization algorithm here. The density and spectral information of 100 samples of oak wood was obtained under absolute drying condition, and the spectral information was collected corresponding to different moisture content. The experimental results show that SPXY guarantees the uniform distribution of calibration samples and improves the generalization ability of the model; Using a combination of MSC, second derivative and S-G convolution can smoothly suppressthe high-frequency noise signal in the original spectrum and make the peak value more prominent;16 characteristic wavelengths were selected by SPA from 117 spectral data. Generally, IPSO-BPNN model has a higher correlation coefficient than SPA-PLS, BP and PSO-BP, own smaller root mean square error. The correlation coefficient of the absolute dry density of oak is 0.938, and the root mean square error is 0.012 9.