Existing prediction models of soil organic matter content (SOC) are restricted by some factors, such as sampling scale, soil type and spectral parameters of samples. Therefore, it is necessary to make a comparative analysis on larger scales to build a quantitative model with better feasibility and greater accuracy. A total of 225 soil samples were collected in an extensive region of the upper reaches of Heihe river basin. SOC and spectral reflectance were being measured. All the samples were divided into 2 subsets-a modeling subset (180 samples) and a validation subset (45 samples). Six indices were obtained through transformation of soil spectral reflectance (R), continuum-removal (CR), reciprocal (REC), logarithm of reciprocal (LR), first-order differential (FDR) and Kubelka-Munck transformation coefficient (K-M). To build the mathematical model of SOC with 12 spectral indices, two methods, i.e., stepwise linear regression and partial least-square regression were used based on the modeling subset, respectively; the validation subset is used for model evaluation. The results indicated that: (1) Regardless of different modeling methods, model between SOC and LR index was always the best among the 6 reflectance-related indices. LR was the best index for predicting SOC; (2) For the model based on the LR index, the accuracy of model using partial least-square regression method was better than that using stepwise linear regression method; (3)225 samples were compared to verify the former available published SOC model. Both the predicted and measured values passed the mean value t-test, and the Pearson correlation coefficient reached 0.826. It shows that local prediction model can be applied to the research of predicting SOC in the larger scale.