The soil organic matter (SOM) content is an important index for evaluating soil fertility. Weigan-Kuqa region in Xinjiang is selected as the study area, based on the laboratory-derived SOM content and reflectance data, the pretreatment of Savitzky-Golay (S-G) smoothing and first order derivative (FD) are carried out. In order to further reduce the influence of sensitive band selection on modeling accuracy, we introduce the harmonic analysis (HA) algorithm to conduct the harmonic decomposition of all wavelengths. Seven principal components are obtained using dimensional reduction treatment of the principal component analysis (PCA). Subsequently, the SOM contents of soil samples are quantified by means of three methods: back propagation (BP) neural network, Genetic Algorithm (GA)-BP, and multiple linear regression (MLR). The accuracy of these methods is compared here. The results show that the correlation coefficient between SOM content and HA pretreated spectral data is improved effectively compared with those of FD data. The estimate accuracy of the non-linear model, BP neural network, is better than that of the linear model, MLR. In terms of non-linear models, the estimate accuracy of the GA-BP model is the best, with the optimal determining coefficient of 0.92, root mean square error of prediction set of 3.92×10 -3, and the relative analysis error of 1.93. This study validates the effectiveness of the HA algorithm for the depth mining of spectral data, and the BP neural network model optimized by GA can improve estimate accuracy of SOM content, which can further provide scientific reference for the quantitative estimation of multiple soil properties.