Abstract: Abstract: The rapid estimation of soil moisture content (SMC) is of great significance to precision agriculture in arid areas. Hyperspectral remote sensing technology has been widely used in the estimation of SMC due to that it's non-destructive and rapid, and has high spectral resolution characteristics. Meanwhile, there are a lot of factors, such as massive spectral data, and surface conditions, which might affect the spectra, increasing the difficulty in extracting the effective information, and reducing the prediction accuracy of SMC. Noise reduction must be considered in developing hyperspectral estimation models, but how to reduce noise while retaining as much useful information as possible needs investigation. As advanced spectral mining methods, competitive adaptive reweighted sampling (CARS) was used to solve this problem in this study. In the present study, a total of 39 soil samples at 0-20 cm depth were collected from the delta oasis in Xinjiang. The samples were brought back to the laboratory to be dried naturally, ground and passed through a screen with 2 mm hole, and then filled into the black boxes with 12 cm diameter and 1.8 cm depth, which were leveled at the rim with a spatula. Reflectance of soil samples was measured using ASD (analytical spectral devices) Fieldspec 3 Spectrometer in a dark room. We used the following steps to process soil reflectance: First, discrete wavelet transformation (DWT) was used to decompose the original spectra in 8 levels using db4 wavelet basis with MATLAB programming language. In order to select the maximum level of DWT, correlation coefficients between the SMC and the spectra of each level were computed. Secondly, the CARS was used to filter the redundant variables, the wavelength variables with better correlation with SMC were screened out and the characteristic wavelength variables of each decomposition level were superimposed as the optimal variable set. Thirdly, partial least squares regression (PLSR) was employed to build the hyperspectral estimation models of SMC. And then, root mean square error of calibration set (RMSEC), determination coefficient of calibration set (R2 c), root mean square error of prediction set (RMSEP), determination coefficient of predicting set (R2 p) and relative prediction deviation (RPD) were used for accuracy assessment. The results showed that: 1) With the increase of the number of decomposed layers, the correlation between soil reflectance and SMC showed a trend of increasing first and then decreasing, and L6 was the most significant band at 0.01 level. In general, the characteristic spectrum of L6 was denoised, and at the same time, the spectral detail was preserved to the maximum extent, so the maximum decomposition order of the wavelet was 6-order decomposition. 2) The characteristic wavelength variable of the characteristic spectrum was selected by coupling wavelet transform and CARS algorithm. However, if only the CARS selection result of the feature spectrum was taken into account, it was easy to ignore the water features of other characteristic spectra. Therefore, in this study, by adding the characteristic wavelength variables of each layer as the optimal set of variables, it contained 131 wavelength variables near the absorption band (450, 1400, 1900, 2200 nm). 3) Compared with the full-band PLSR model, the accuracy of PLSR model of CARS preferred variables for each decomposition level was high, and the PLSR model of the optimal variable set had the highest accuracy and a better performance in predicting SMC in the study area (RMSEC=0.021, R2 c=0.721, RMSEP=0.028, R2 p=0.924, RPD=2.607). It is shown that the combination of wavelet transform and CARS algorithm makes it possible to remove the noise as much as possible and to remove the noise completely when the model is established, and at the same time, it can effectively remove the non-information variable and provide a new idea of the screening of the SMC spectral variable in this region.