Hyperspectral nondestructive testing technology is widely used in quantitative nondestructive testing of fruit. In this paper, the spatial characteristic spectra of jujube, grape and pear are taken as the research objectives, and the influencing factors and inversion methods of spatial characteristic spectra are explored, which provides a new idea for improving the accuracy of outdoor fruit nondestructive testing. The spectral library of three kinds of fruits was extracted, and the spatial characteristic spectra were calculated. The characteristic wavelengths were selected by Mahalanobis distance, concentration residual and competitive adaptive weight sampling algorithm. Model characteristic spectra of three kinds of fruits after pretreatment with quality parameters and positional parameters respectively.The modeling results are as follows: In the sugar model, the R of jujube, grape and pear were 0.853 3, 0.822 7 and 0.913 3 respectively; In the water model, the R were 0.741 3, 0.784 7 and 0.891 3 respectively; In the detection angle model, the R were 0.985 6, 0.992 7 and 0.974 7 respectively; In the azimuth angle model, the R were 0.941 8, 0.910 5 and 0.936 9 respectively; In the phase angle model, the R were 0.960 9, 0.957 0 and 0.956 3 respectively. In summary, the correlation of different fruit positional models was significantly higher than quality models. Therefore, the positional factor is the main reason affecting the characteristic spectrum. Therefore, the roujean model and waltall model are used to invert the characteristic spatial spectrum of different directions. The inversion results are as follows: roujean model is used when retrieving the spatial characteristic spectra of three kinds of fruits (in the order of jujube, grape and pear), R² is 0.934 4, 0.928 1 and 0.830 6 respectively; R is 0.990 2, 0.983 9 and 0.969 1 respectively; RMSEP is 0.030 9, 0.048 7 and 0.062 7 respectively; the average model error is 7.27%, 11.02% and 8.61% respectively. The results showed that R² was 0.943 3, 0.859 7, 0.839 0; R was 0.991 8, 0.971 8, 0.970 2; RMSEP was 0.036 6, 0.066 1, 0.068 7; the average model error was 6.19%, 15.40%, 7.84%. It can be seen that roujean model can well describe the characteristic spatial spectrum of jujube and grape, and also can better describe the characteristic spatial spectrum of pear; waltall model can well describe the characteristic spatial spectrum of jujube, and also can better describe the characteristic spatial spectrum of grape and pear. In conclusion, roujean model can be used to invert the characteristic spatial spectrum of grape and pear, and waltall model can be used to invert the characteristic spatial spectrum of jujube to improve the accuracy of outdoor fruit nondestructive testing.