In order to predict the content of iron oxide in soil accurately and quickly, 135 soil samples are collected from the southern edge of Lufeng Dinosaur Valley, and soil spectral data and iron oxide content are measured in the laboratory. The original spectrum is smoothed by the Savitzky-Golay filter, and then conventional spectral transform and continuous wavelet transform are performed. The correlation coefficient (CC) method is used to analyze the correlation between the transform spectrum and iron oxide content. The wavelengths that pass the 0.01 significance test in each scale are selected as the coarse wavelengths, and the wavelengths selected by the competitive adaptive reweighted sampling (CARS) are further used as the characteristic wavelengths. Finally, the support vector regression (SVR) optimized by the genetic algorithm is used for modeling. The results reveal that continuous wavelet transform can improve the correlation between the soil spectral reflectance and iron oxide content. The number of independent variables for modeling can be effectively reduced by the CC-CARS wavelength selection method. The model constructed by the fourth-scale continuous wavelet decomposition (L₄-CC-CARS-SVR) has the best effect. The coefficient of determination R² and root-mean-square error E_RMSE of its calibration set is 0.760 and 5.236 g·kg^-1,respectively. The R², E_RMSE and performance to interquartile range ratio R_PIQ of its validation set is 0.663, 7.798 g·kg^-1 and 2.598, respectively, which indicates that the model has good stability and predictive ability.