One hundred thirty for soil samples of Qitai in Xinjiang were selected, and the first derivative spectrum of the soil sample logarithmic reflectance was decomposed to many layers by using 4 wavelet functions respectively, and PLSR was used to establish the prediction models respectively, and precision values were tested. The results show that: 1-3 layers low-frequency coefficients of wavelet decomposition were better, while the rest were worse. In 6 layers of all function decomposition, the highest accuracy of inversion models constructed by low-frequency coefficients were all ca2, while with increasing the decomposition layers, the precision and significance decreased significantly. In the same scale, there was little accuracy difference between inversion models constructed by 4 wavelet functions low-frequency coefficients, while Bior1.3 was optimal. The best inversion model was ca2 that built by Bior 1.3, with R2 and RMSE being 0.977 and 7.51 mg x kg(-1) respectively, reaching to significant level. Upon testing, it can be used to estimate the alkaline hydrolysis nitrogen content quickly and accurately.