Selecting the characteristic wavelength in spectra for modeling can reduce the interference by redundant wavelengths and improve modeling accuracy. The spectral data of 104 soil samples collected are preprocessed by a wavelet threshold de-noising method. The wavelengths are selected for modeling by 9 wavelength selection methods including interval partial least squares, uninformative variable elimination, successive projection algorithm and swarm intelligence algorithm. The results show that the wavelet threshold de-noising method can reduce the noise in spectra effectively. Using wavelength selection methods to select the wavelengths for modeling not only can reduce modeling variables, but also can improve the prediction accuracy of the model. Particularly, since the discrete particle swarm optimization algorithm uses 26 wavelengths for modeling, its prediction determination coefficient reaches 0.81 and its relative standard prediction error is 2.31. The selection of spectral wavelengths not only can reduce the complexity of the model, but also can effectively predict the organic matter content in soil.