Abstract: Abstract: Hyperspectral technology is a popular method of predicting soil moisture content nowadays, however, soil spectra include quantities of invalid redundant information, which is a serious bottleneck problem that could lead higher complexity and lower accuracy of prediction model. In this study, 96 fluvo-aquic soil samples were collected at 0-20 cm depth in fields in Qianjiang city, Hubei province, China, and then the samples were pretreated by air-drying, grinding and sieving in a laboratory. Samples with different soil moisture content (SMC, mass fraction of 0-40%) were prepared. For each sample, hyperspectral reflectance was measured by an ASD Field Spec3 instrument. Outliers with abnormal data were removed by Monte Carlo cross validation method. After that, the raw spectral reflectance was processed by Savitzky-Golay smoothing method and continuum removal method. Then, spectra of samples were divided into 2 subsets by Kennard-Stone algorithm. One subset was a calibration set with 47 samples and the other subset was a prediction set with 30 samples. The wavelength variables sensitive (SWV) to SMC were selected from the full-spectrum by competitive adaptive reweighted sampling (CARS) method, and they were considered as an optimal variable set. The multivariate calibrations were performed with partial least squares regression by using the full-spectrum (F-PLSR) and the optimal variables (CARS-PLSR), respectively. The prediction accuracy was assessed by comparing determination coefficients (R2), root mean squared error (RMSE) and relative percent deviation (RPD). Results showed that the SMC greatly affected soil spectral reflectance. The soil spectral reflectance reduced with the SMC increase, and 4 soil moisture absorption peaks were obvious around 450, 1 400, 1 900 and 2 200 nm. The peaks at 1 400 and 1 900 nm had the obvious redshift phenomenon. The SWV of samples with different moisture were obtained by CARS, and then an optimal variables set was generated including wavelengths of 443-449, 1 408-1 456, 1 916-1 943 and 2 209-2 225 nm. Using the CARS-PLSR model calibrated by the optimal variables, the predicting accuracy was improved compared to the F-PLSR model calibrated by the full-spectrum. The predicting accuracy of CARS-PLSR (R2, RMSE and RPD were 0.983, 0.0144 and 8.36, respectively) was higher than F-PLSR model (R2, RMSE and RPD were 0.976, 0.0166 and 7.21, respectively). Meanwhile, compared to F-PLSR model, the CARS-PLSR model reduced the variable numbers from 2 001 to 101. In brief, the CARS-PLSR model not only enhanced the forecasting capability but also reduced the model complexity. Thus, this approach could facilitate the development of soil moisture sensor in field in the near future.