Building universal deduction models for predicting the soil total nitrogen (TN) content by using data mining of large soil spectral libraries is one of the most important applications of hyperspectral analysis. In this study, 1661 soil samples representing 17 soil types from 13 provinces of China (e.g., Tibet, Xinjiang, Heilongjiang and Hainan) are employed for modeling the soil TN content using global partial least squares regression (PLSR), locally weighted regression (LWR) and fuzzy K-means clustering combined with PLSR (FKMC-PLSR). Another 104 paddy soil samples collected from Zhejiang Province are used to validate the established models. Results showed that when predicting soil TN from a large dataset, global PLSR underestimates high values of TN, which generates an overall low prediction accuracy. By contrast, LWR and FKMC-PLSR perform better than global PLSR. It is suggested that the results can provide useful information for establishing robust and universal models for soil TN prediction using large soil spectral libraries.