Soil organic matter (SOM) is not only an important indicator of soil fertility but also an important source and sink in the global carbon cycle. Therefore, it is essential to acquire the information of SOM for soil management. The visible and near-infrared (VIS-NIR) reflectance spectroscopy technique, known as a novel, rapid, accurate, environment-friendly and efficient approach when compared with conventional laboratory analyses, is a promising one to acquisition of soil properties. Construction of a calibration set is key to use of VIS-NIR quantitative analysis in building up a quality prediction model. Conventionally, selection of samples for the calibration set is based on soil physical and chemical properties or soil spectral information, like the concentration gradient method(C) and Kennard - Stone (KS) method, which are able to select samples that may be representative of physical and chemical properties or spectra, but not of geographical space and multivariate information. Impacts of the shortages on prediction accuracy of the model have rarely been reported. The aim of this paper is to explore how sample selection methods affect accuracy of the VIS-NIR reversion model in estimation of SOM, using soil samples collected from lands under different types of land use in the riparian areas of the Jianghan Plain. A total of 270 soil samples were collected, air dried and ground to pass a 2 mm sieve, for analysis of VIS-NIR spectra using a FieldSpec3 spectrometer. The spectral curves were preprocessed with log10, Savitzky-Golay (SG), multiplicative scatter correction (MSC) and mean center (MC). A total of four categories of ten sample selection methods based on multivariate soil information were proposed for constructing calibration sets. The first category, including the concentration gradient method and the method adopted several properties (P-KS), depends on soil physical and chemical properties; the second category, including the KS method and the Reduce on Neighbor Samples (RNNS) method, is based on spectral information; the third category, including the C-KS and C-RNNS methods, combines soil physic-chemical properties with spectral information; and the forth category uses land use type hierarchy in combination with all the aforementioned methods. The P-KS method takes into comprehensive account parameters, like SOM, Fe, N, P and bulk density (BD), that may be quite high in weight of impacts on soil spectra and uses KS algorithm to select soil samples representative of a variety of physical-chemical properties for construction of the calibration set. The C-KS and C-RNNS methods divide SOM concentration into six levels, from each of which two-thirds of the samples were selected using the KS and RNNS methods to form the calibration sets. The methods based on land use type hierarchy divide the entire sample set into three categories, namely dry land, paddy field and the others. For each category, soil samples representative of SOM distribution or soil spectra were selected in combination with the concentration gradient method, KS, RNNS and C-KS, separately to form calibration sets, which were then merged into a calibration set representative of land use type. On such a basis, a partial least squares regressions (PLSR) model was established, showing that in the first and second categories, the models with calibration sets formed with the C, KS and RNNS methods, representative of SOM distribution or soil spectra singularly, were not so good in prediction accuracy; and those with the P-KS method were much better, with determination coefficient for prediction (R_p²)being 0.55, root mean squared error of prediction (RMSE_p) being 7.54 and ratio of performance to standard deviation (RPD) being 1.47. The models with calibration sets formed with the C-KS method, representative of both physical and chemical properties and spectra, were good in accuracy with R_p² being 0.64, RMSE_p being 7.13 and RPD being 1.66. The inclusion of land use type in forming calibration sets, greatly improved the models using the C, RNNS and C-KS methods in prediction accuracy, bring R_p² up to 0.70, 0.59 and 0.68, RMSE_p to 6.34, 6.47 and 6.58, and RPD to 1.84, 1.84 and 1.51, respectively. It is therefore, quite obvious that the use of calibration sets formed with soil samples representative of multi-layers of soil information can improve the models in prediction accuracy. The L-C method has turned out to be the best method for sample selection in construction of calibration sets for VIR-NIR models for prediction of soil organic matter contents in the riparian areas of the Jianghan Plain.