In recent years, proximal hyperspectral technology provides a new approach in timely, effectively and nondestructive way to detect soil organic matter (SOM). However, the hyperspectral dataset contains too many wavelengths which could lead to the collinearity, redundancy and noise to models. The Normalized Difference Spectral Index (NDSI) derived from soil spectral reflectance could enhance the relationship between spectral features and SOM, and also could eliminate the irrelevant wavelengths. In this paper, 56 topsoil samples at 0～20 cm depth were collected as research objects from Gong’an County in Jianghan Plain, the spectral reflectance was measured using the ASD FieldSpec3 spectrum analyzer, and the SOM was determined using potassium dichromate external heating method in the laboratory. In the next stage, the raw spectral reflectance (Raw) was prepared for three spectral transformations, i.e. inverse-log reflectance (LR), first order differential reflectance (FDR) and continuum removal reflectance (CR). 2-D correlograms of the determination coefficients (R2) were constructed using all two-band combinations of 4 spectral transformations in NDSI against SOM in the range of 400～2 400 nm. Then, the determination coefficients (R2) of the 4 spectral transformations for 1-D determination coefficients and 2-D determination coefficients by F significant test were got (ｐ<0.001), which could be used to extract sensitive bands and spectral index. At last, partial least squares regression (PLSR) method were used to build quantitative inversion model of SOM based on sensitive bands and spectral index for this study area, respectively. Feasibility of 2-D spectral index for building model was this study aimed to explore. The results showed that, the 2-D determination coefficients were better than 1-D determination coefficients, especially the determination coefficients of LR was improved by about 0.26. Compared to the sensitive bands derived from 1-D determination coefficients, on the whole, the sensitive spectral index derived from 2-D determination coefficients using PLSR method could obtain more robust prediction accuracies. The prediction accuracy of NDSILR-PLSR was the best, and its values of R2, RPD for the predicted model were 0.82, 2.46, which could estimate SOM comprehensively and stably. In the future, this method could be applied to air- or space-borne images with a lower spectral resolution (e.g. ASTER, Landsat TM), and the results could also provide great potential in the field of sensor design for portable proximal sensing researching.