Soluble solid content (SSC), including sugar, acid, fibrin and mineral components, is a comprehensive index for evaluating the fruit maturity and quality, which can affect the taste, flavor and shelf life. Non-destructive and rapid detection of SSC in watermelon is very important for determining the maturity and monitoring the internal quality during storage and transportation, and is helpful to improve production efficiency and market competitiveness of watermelon. For the rapid and non-destructive near infrared (NIR)-based detection of the watermelon SSC, many researchers have used near infrared diffuse transmission method, which requires high light energy and high power transmission, and high power transmission will affect the internal quality. In contrast, the number of researches on near infrared diffuse reflectance method are relatively smaller. It has the advantages of low light energy and low cost, which is in favor of miniaturization and portability of the instruments, and will avoid the fruit quality changes caused by high power transmission. In this study, the greenhouse watermelon was used as the research object, and the near infrared reflectance spectra were collected in the watermelon stem, navel and equator at near 976, 1 186 and 1 453 nm by using JDSU portable near infrared spectrometer. The models between watermelon SSC and near infrared reflectance spectroscopy were established by using partial least square regression (PLSR). Firstly, the sample collection of different parts in the watermelon was divided based on the joint x-y distances (SPXY) method, with SSC as y variables and spectral as x variables. The samples distances were calculated by using x and y variables, and the watermelon samples were divided into 51 calibration sets and 15 prediction sets. The SSC of the calibration sets has a wide distribution range, which covers that of the prediction sets, and can increase the diversity and representativeness of samples and help to build a stable and reliable prediction model. Secondly, the prediction accuracy of quantitative models between the near infrared reflectance spectroscopy and SSC in different detection positions was investigated, and higher correlation and better prediction performance was found in the equator position with prediction correlation coefficient of 0.629 and root mean standard error of prediction of 0.49%. The accuracy of the models between SSC and near infrared spectra information in different watermelon positions was related with the spectrum collection ways and the differences in growing area, variety and maturity. Therefore, the determination of the detection position in the watermelon should be based on the actual situation in the model-building process. Finally, in order to improve the prediction accuracy of the models built for the watermelon equator, the spectra should be pre-processed with the model built for the watermelon equator, and then normalize the results, based on which we can obtain the best prediction model of PLSR. The prediction correlation coefficient was 0.864 and the root mean standard error of prediction was 0.33%, showing higher correlation and improved prediction accuracy. In conclusion, the results indicated that the SSC of the greenhouse watermelon can be accurately predicted based on detecting the equator position by near infrared reflectance spectroscopy. Therefore, it has the potential for improving the rapid and non-destructive testing technology and developing small and portable equipment to detect watermelon SSC by near infrared spectroscopy.