Soil hyperspectral technique could estimate soil organic carbon content efficiently. Continuous wavelet transform had unique advantages in noise removal and effective information extraction of hyperspectral data. However, the spectral data after continuous wavelet transform was decomposed into multiple scales, and the information of a single decomposition scale could not represent the information of different decomposition scales. Making full use of the wavelet coefficients of multiple decomposition scales becomes a difficult problem for hyperspectral estimation of soil organic carbon content. Lake Bosten was the largest inland freshwater lake in China, and the lakeside oasis, as an important interlacing zone between land and water, had a unique spatial and temporal structure and played an important role in maintaining and restoring the health of the lake ecosystem. The study area was the lakeside oasis of Bosten Lake. 138 surface soil samples were collected in September 2020 at a depth of 0~20 cm, 3 outlier samples were excluded to obtain 135 useful samples, soil sample spectra were collected outdoors, and soil organic carbon content was determined by potassium dichromate-external heating method. The continuous wavelet transform was then performed with Gaussian4 as the wavelet basis function to convert the soil hyper spectrum into wavelet coefficients at 10 decomposition scales, and the correlation coefficient method, Stability Competitive Adaptive Reweighted Sampling, Competitive Adaptive Reweighted Sampling, Successive Projections Algorithm, Genetic Algorithm, five special wave filtering methods to further reduce noise and eliminate redundant information, calculate the root mean square of wavelet coefficients as wavelet energy feature scale by scale, and form a wavelet energy feature vector (Energy Feature Revector) with 10 scales of wavelet energy features, and build a BP neural network model (BP neural network model) based on the wavelet energy feature vector. The result showed that wavelet continuous transform could effectively improve the correlation between spectral reflectance and soil organic carbon content, with poor correlation at the 1~3 decomposition scale and good correlation at the 4~10 decomposition scale, with an average increase of 43.66% in the correlation coefficient and an average increase of 67.93% in the maximum value of the correlation coefficient. The feature band screening CC algorithm was mainly distributed in 400~1 500 nm visible and NIR short wavelength; sCARS and CARS algorithms were concentrated in 1 500~2 500 nm NIR long wavelength; SPA algorithm was concentrated in 760~2 500 nm NIR band; GA algorithm was uniformly distributed in 400~2 500 nm. The hyperspectral wavelet energy feature could better estimate the organic carbon content of the surface soil of the lakeshore oasis, and the mean R2 values of the training and validation sets of the six models were 0.73 and 0.74, respectively; the mean RMSE values were 7.64 and 7.28, respectively; and the mean RPD value was 1.95. The model accuracy showed that CC-EFV-BPNN＞sCARS-EFV-BPNN＞Full-spectrum-EFV-BPNN＞CARS-EFV-BPNN＞GA-EFV-BPNN＞SPA-EFV-BPNN. The continuous wavelet transform combined with the feature variable screening method to extract the wavelet energy feature vector effectively reduces the spectral data dimensionality and hyperspectral wavelet energy feature model complexity, an important reference value for rapidly estimating surface soil organic carbon content.