Abstract: Abstract: The expanding trend of soil salinization has become more and more severe especially in arid and semi-arid areas in the northwest China. Therefore it is specially important to dynamically monitor the soil salinization in the arid and semi-arid areas scientifically, accurately and rapidly. Microwave remote sensing technique has become a promising method to detect and monitor the soil salinity due to its many advantages. The aim of this study was to investigate the capability of C-band RADARSAT-2 SAR (synthetic aperture radar) data in soil salinity estimation over agricultural fields. In this study, Jiefangzha zone of Hetao irrigation district, Inner Mongolia, China was selected as the study area. Based on the back-scatter coefficient value and soil salt content, this paper used 3 kinds of methods including the multiple linear regression (MLR), geographically weighted regression (GWR) and back propagation artificial neural network (BP ANN) to establish the quantitative inversion models of soil salt content. Soil salinity information was extracted from the RADARSAT-2 SAR data, which had a kind of fine four-polarization SLC (single look complex) format and were bought in 2013, and covered an area of 25 km × 25 km with 8 m ground resolution. Taking the spatial unevenness distribution of the saline soil into account, 69 sampling points were designed in the study area, and field digging depth of soil was 10 cm. Hand-held GPS (global position system) receiver was used to record the coordinates of sampling points, and the soil total soluble salt content was measured in the indoor. Mainly use the SAR Scape module of ENVI software to perform the radar image processing, including radiometric calibration, geometric correction, slant range turning and filtering. The four-polarization back-scatter coefficient values corresponding to the sampling points were extracted based on the previous results by the spatial analysis module of ArcGIS. Total salt content was took as dependent variable, four-polarization back-scatter coefficient values and soil moisture as independent variables, and the MLR and GWR salinity prediction models were established. Use the MLR module of SPSS and the GWR module of GWR4 software to deal with data respectively. The correlation of GWR model was higher than the MLR, but these 2 statistical models were not significant, and the inversion results were difficult to reflect the distribution of severe salinity and saline-alkali soil. Therefore the emphasis was focused on the building of the 8:140:1 structure of three-layer BP ANN model. Training set contained the data of 49 sampling points in the test area, the input layer was made up of sampling point coordinates, soil moisture, four-polarization back-scattering coefficients and their combined value, and the number of net neurons was 8; the output layer was a neuron corresponding to the total salt content of sample point. After a lot of tentatively computation, the optimal number of neurons in the hidden layer was selected as 140. The hidden layer of the three-layer BP ANN model used the hyperbolic tangent sigmoid activation function, the output layer utilized the linear activation function, the traingda function was selected to perform network learning and training, and the fitting target error was set as 0.1 mg/g. Then the BP model was applied to the RADARSAT-2 remote sensing images to achieve the high-precision quantitative inversion of soil salinity when the BP model inversion accuracy met the requirements of salinity prediction. Whole calculation process was implemented by the MATLAB neural network toolbox. It was found that the correlation of the MLR and GWR models was weak and their standard error (SE) was 0.55 and 0.47 mg/g respectively, but the SE of internal and external inspection using the ANN (BP) model was only 0.24 and 0.33 mg/g respectively, better than the other 2 models, whose salinity area accounted for about 65.4%, and was basically consistent with ground validation. The artificial intelligence inversion model of soil salt considered the moisture's influence, was directly based on the polarization radar scattering coefficients and their composition, and had no complex dielectric constant model. So the ANN (BP) model can reduce the smoothing effect compared with the 2 traditional models and improve the accuracy and reliability of model predictions, which meets the needs of soil salinity monitoring to a certain extent, and can promote and develop the application of microwave remote sensing in the soil salinity monitoring.