It is of great significance to obtain soil salt information timely and accurately for guiding rational irrigation, ensuring normal growth and development of crops, and realizing high yield. Sunflowers of four kinds of croplands with different salinizations in Shahaoqu District of Hetao Irrigation Area were set as the study object, remote sensing data were obtained by using multi-spectral camera and thermal infrared imager, meanwhile, the soil salt data at different soil depths in the region were collected. The soil salinity inversion models were constructed for sunflower field in different growth stages and soil depths with four regression methods, including partial least squares regression（PLSR）, support vector machine (SVM), back propagation neural network (BPNN) and extreme learning machine（ELM）, which were based on canopy temperature, and spectral index was screened by grey correlation method. The result showed that the effect of salt inversion model constructed based on the data of sunflower budding stage was better than that of flowering stage on the whole，the model constructed with the preferred salt index and spectral index as the variable group was better than that of vegetation index variable group and the soil depth with good salinity inversion was 0~20cm and 20~40cm. The comparison showed that the effect of machine learning salt inversion model was better than partial least squares regression model, BPNN salt model constructed with spectral index as variable group had the best inversion effect at the depth of 0~20cm soil in sunflower germination stage, in which the modeling R2 and validation R2 were 0.773 and 0.718，and the RMSE and CC of validation reached 0.062% and 0.813，respectively. The research result provided a reference for the application of UAV remote sensing in sunflower field soil salinity monitoring and related research.