Seasonal surface air temperature(SAT) anomalies over China from 1951 to 2003 is investigated by applying a neural-network-based nonlinear principal component analysis(NLPCA) method. The results of the first NLPCA mode show that the seasonal SAT anomalies have some nonlinear characters;its nonlinearity is stronger in spring(MAM) and summer(JJA), and weaker in autumn(SON) and winter(DJF). The SAT approximation by 1-D NLPCA is closer to the observations than that by 1-D PCA.