The selection of modeling method is one of the main factors influencing the quantitative accuracy with visible-near infrared (Vis-NIR) spectroscopy. We compare the performance of three calibrations methods, i.e., principal component regression (PCR), partial least squares regression (PLSR), and back propagation neural network (BPNN) based on Vis-NIR reflectance spectra of soil total nitrogen (TN) quantitative forecast results. Covered in the 470~1000 nm wavelength range, spectroscopy of 48 soil samples selected from 12 profiles are air-dried, screened and mushed, then processed by the first order derivative and Savizky-Golay smoothing methods. Leave-one-out cross validation is also adopted to determine the optimal factor numbers. The results indicate that PCR and PLSR linear models are able to meet general prediction and with little difference, where coefficients of determination (R2) are 0.74 and 0.8, respectively, and residual predictive deviation (RPDs) are 2.23 and 2.22. The two nonlinear models built by BPNN in combination with PCR and PLSR, respectively, are superior to the linear models of PCR and PLSR in the precision of prediction. BPNN, principal components (PCs) whose input is the PCs resulted from the PCR, while the BPNN latent variables (LVs) whose input is the first 4 LV results obtained from PLSR has the best performance (R2=0.9, RPD is3.11). It is recommended to adopt BPNN-LV model to rapidly predict the vertical spatial and temporal distribution of TN with Vis-NIR spectroscopy.