Soil organic matter (SOM) is an essential indicator for the fertility assessment of farmland. and An efficient and stable prediction model is in need to rapidly estimate SOM in larger scale. Spectroscopic technology has been proved as a powerful tool to access SOM in the last decade. The aims of this paper were: to compare different selection method of calibration set for modeling SOM in paddy soil by using visible-near infrared (VNIR), mid-infrared (MIR) and VNIR-MIR spectra and to assess the prediction ability of the results. All spectra were transformed from reflectance to absorbance, and preprocessed by Savitzky-Golay smoothing algorithm. The prediction models of SOM were built by using partial least squares regression (PLSR) coupled with three selection methods of calibration set in VNIR, MIR and VNIR-MIR regions. The selection method of calibration Rank-KS performed better than Rank method and KS method, meanwhile the models in MIR region showed stronger prediction ability than VNIR and VNIR-MIR regions. The best prediction model was obtained with the MIR model combined with Rank-KS, and the root mean square error of prediction (RMSEP) and ratio of performance to deviation (RPD) were 3.25 g·kg-1 and 4.24. According to variable in the projection (VIP) score, important bands for SOM prediction in paddy soil were identified in VNIR and MIR region. Our results show that MIR spectroscopy could make quantitative prediction of SOM in paddy soil and Rank-KS is an effective method for selection of calibration sets, so as to provide some scientific basis for fertility assessment of farmland and rational fertilization.