[Objective] The objective of this study is to explore effective strategies for predicting soil organic carbon (SOC) in local areas with high precision based on the spectral similarity indexes in the global spectral library. It goes specifically as follows: (1) to compare different spectral similarity indexes obtained separately with three different similarity matching algorithms (Euclidean Distance, Mahalanobis Distance and Spectral Angle Mapper) in effect on prediction accuracy; (2) to compare calibration sets different in size in effect on prediction accuracy; and (3) to compare different modeling strategies (PLSR modeling and the assignment strategy) in effect on prediction accuracy. [Method] From the global spectra library a total of 245 China samples were selected to form a prediction set and the remaining 3 537 samples a reference set. From each spectrum in the prediction set, different numbers of similar spectra (5, 10, 20, 30, 40, 50, 100, 150, 200, 250, 300, 400 and 500) were picked out separately with the three different similarity matching algorithms for comparison between the spectra selected by the different similarity algorithms. Based on the reference sets of different sizes selected by the different algorithms, PLSR models were built to predict SOC contents, and effects of the similarity matching algorithms and size of the modeling set on prediction accuracy were evaluated using R², RMSE and RPD. Then a similarity matching algorithm with the highest prediction accuracy was selected and on such a basis, comparison was performed between the different modeling strategies in effect on prediction accuracy: Ⅰ, Prediction with PLSR modeling; Ⅱ, Prediction with direct assignment. [Result] Compared with the overall model, the three similarity algorithms greatly improved prediction accuracy. Among the three, the SAM model was a bit higher than the other two in prediction accuracy (R² =0.75, RPD=1.73). The low prediction accuracy might be attributed to the wide distribution of the soil samples in the global soil spectral library that caused marked variation. Size of the modeling sets did have a great impact on modeling accuracy, and the optimal size for the three similarity algorithms varied in the range of 400~500 (0.71<R² <0.75, 1.56R² >0.6, RPD>1.4). However, when the modeling set was big in size (>50), the PLSR modeling strategy was higher in prediction accuracy (R² >0.6, RPD>1.4). [Conclusion] Compared with the global model, the models based on the three spectral similarity indices all significantly improve SOC prediction accuracy. In general, the spectral angle algorithm is slightly better than Euclidean distance and Mahalanobis distance; type of the similarity algorithm, size of the modeling set and method of the modeling all have a great impact on precision of the SOC prediction.