For the real-time compositional endophenotype detection of individual wheat kernel, a quantitative and non-destructive device based on near infrared diffuse reflectance was designed and developed. The hardware consisted of a novel stereoscopic light source unit, spectrum acquisition unit and control unit, also the corresponding detection software were presented. NIR miniature LED lamps were placed in six rows multiple eight columns along an aluminum cylindrical tube to provide a constant centripetal light and the 48 lamps were connected in parallel by 16 copper conductor to inform a physically stereoscopic light structure. A pair of infrared radiation sensor was adopted on the top of the cylindrical tube to trigger the NIR spectrometer for spectra collection as a seed fell through the borosilicate glass tubing inside the light source. The spectrometer of Ocean Optics was connected to a PC with its standard interface and pin definitions, which was used to collect spectrum from the top and bottom of the light source in real-time through a bifurcate structure fiber. All the soft functions were designed in C++ language of Visual Studio platform. The realtime diffuse reflectance spectrum of each seed was transferred to absorbance via PC and predicted its real protein content according to a model embedded in the program. In order to set up a reliable prediction model, totally 300 individual wheat samples were collected to acquire absorbance spectra in the range of 900~1700nm, and pretreated with standard normal variate correction (SNV) algorithm. Two calibration models were established based on full spectra (FS) and feature wavelengths optimized by successive projections algorithm (SPA) respectively. Data showed that the calibration model based on SPA had a relatively lower determination coefficient (R2) of 0.8446 in contrast to the R2 value of 0.9604 based on FS, but the validation model based on SPA had relatively equal prediction accuracy with the model based on FS. For the nine feature wavelengths selected by SPA eliminated the collinearity relationship in spectral data but preserved characteristic of protein on spectrum band with a concise model equation, it was chosen as a superior model and developed in software to predict individual seed protein content online. To verify system design and performance, a series of experiment was conducted for wavelength repeatability, absorbance repeatability and protein predictive repeatability. The results indicated that the compositional detection device based on stereoscopic light source was able to realize fast, nondestructive and real-time detection of protein content for individual seed, also had certain applications potential on other compositional endophenotype detection for wheat and other crop seeds.