The unmixing algorithms of mixed snow pixels and the fractional snow cover products are an important research direction for snow remote sensing. In the present study, we first designed the mixed snow pixels of different snow fraction/proportion in Northern Xinjiang, China as ground truth. Then, a SVC HR-1024 ground-based spectral radiometer was used for measuring the spectral property of this designed pixel for different snow fractions and different underlying surfaces. Finally, using the measured spectral data, the four mixed-pixel decomposition models were validated and evaluated for their performance in terms of accuracy and computational efficiency. The results showed that the reflectivity does not decline linearly with the reduction of snow ratio in the pixel, and that the unmixing accuracy inversely depends on the scales of the observation. Further, the comparison of the above mentioned unmixing algotihms showed that the linear regression method has the worst accuracy, especially when the snow proportion is less than 50%; the accuracy of sparse regression algorithm and non-negative matrix factorization were slightly higher than the full constrained linear mixed-pixel decomposition; however, full constrained linear mixed-pixel decomposition method had higher computational efficiency than the other two methods; the sparse regression algorithm has lowest computational efficiency. With unmixing remote sensing images, due to the large data volumes, we must consider the algorithms’ computational efficiency. This study would promote quantitative researches on snow mixed pixel decomposition, and provide a theoretical basis for accurately extracting the snow coverage of interest area using remote sensing images.