Abstract: Soil salinization is a major problem of land degradation in arid and semi-arid agricultural area. It is crucial to detect the salinity of saline soils accurately and dynamicly in time in order to prevent soil salinization and achieve sustainable development in agriculture. Taking Changling County western Songnen Plain, as the example, this paper constructed remote sensing predictive model of saline soils using hyperspectral data. The salinity was measured by electric conduction method, and hyperspectral data was collected using ASD spectrometer. Derivative transformation of spectral reflectance was used to select best spectral bands which can represent the salinity of saline soils, e.g. 550, 720, 760, 820 and 940 nm. The best performance was achieved in the 5-6-1 architecture (R2 = 0.895, RMSE = 0.089) in 72 different architectures in the three- and four-layer networks. Compared with traditional regression model (R2 = 0.81, RMSE = 0.25), the method combining hyperspectral data with artificial neural network can improve the predictive accuracy of saline soil, showing that artificial neural network is prominently advanced in establishing the relationships between spectral reflectance and soil parameters.