The rapid estimation of soil moisture content (SMC) is of great significance to precision agriculture in arid and semi-arid areas. Using Organ River-Kuqa River delta oasis as research area, we adopt wavelet transform to realize 1-8 layer wavelet decomposition for reflectance spectrum. The maximum number of decomposition layers is determined by correlation analysis, nine routine mathematical transformation methods are used for conducting characteristic spectrum of each layer from original reflectance to maximum number of decomposition layers, and the correlation analysis between reflectance of soil and SMC is carried out. Waveband with maximum correlation coefficient is taken as sensitive waveband filtrated from all kinds of transformation of characteristic spectrum of each layer. Optimum waveband combination is filtrated by grey relational analysis (GRA). SMC prediction model is developed and analyzed by partial least squares regression. The results show that, with the increase of the number of decomposed layers, the correlation between soil reflectance and SMC increases and then decreases, and L6 is the most significant band at 0.01 level. In general, the characteristic spectrum of L6 can maximally preserve the spectral details while denoising, so the maximum decomposition order of the wavelet is 6 order decomposition; In general, it is shown that the combination of wavelet transform and differential transform can deepen the spectral potential information and improve the correlation between reflectance of soil and SMC. Comparing the predictive effects of SMC estimating models, the model based on L-GRA is much better than others, and it has better performance in predicting SMC in the study area (root mean square error of calibration is 0.026, determination coefficient is 0.710, root mean square error of prediction is 0.030, determination coefficient is 0.965,and residual predictive deviation is 2.800). It is shown that the combination of wavelet transform and GRA makes it possible to lose the spectral details as little as possible and remove the noise more completely when the model is established, at the same time, it can effectively remove the non-information variables.