Near-infrared reflectance spectroscopy (NIRS) was used as a rapid and nondestructive method to determine the protein content and gossypol content in cotton kernel powder samples, using 49 upland cotton (Gossypium hirsutum L.) germplasms and 188 recombinant inbred lines (RILs). The cottonseed samples harvested from the upland cotton germplasms and RILs grown in different cotton growing regions in different years were analyzed chemically for protein and gossypol contents, as well as scanned in the reflectance mode of a scanning monochromator. Using ISI software for scanning and data analysis, protein and gossypol calibration equations were obtained with a standard normal variate + detrending scatter correction and a 2, 4, 4, 1 math treatment and modified partial least square (MPLS) as the regression method. The protein content calibration results revealed that the multiple correlation coefficients (RSQ) and statistic 1--variance ratio (1-VR) for the determination of protein content in cottonseed kernels were 0.933 and 0.929, respectively, and its standard error of calibration (SEC) and standard error of cross validation (SECV) were 0.623 and 0.638, respectively. As the calibration equations were judged by the calibration RSQ (or 1-VR) and SEC (or SECV), the results indicated that NIRS is comparable to chemical methods in both accuracy and prediction and is reliable in the determination of protein content in cottonseed kernels. However, the RSQ, SEC, 1-VR and SECV for gossypol content determination of NIRS were 0.836, 0.811, 0.074 and 0.079, respectively. Although it was weaker than that of protein content, the NIRS method is still good enough for the determination and prediction of the gossypol content in cottonseed kernels. Therefore, NIRS models were successfully developed for protein content and gossypol content analysis of cotton kernel powder sample in the present study and they could be introduced into the cotton germplasm evaluation and breeding program for the cottonseed quality improvement.