Fractional Snow Cover Mapping Using MODIS Data In Subpixel Scale

With the development of remote sensing (RS) technology, the RS data has become an indispensable material in snow cover dynamic monitoring, which plays an important role in the snow monitoring and disaster analysis. The MODIS data has been widely used in the study of Fractional Snow Cover (FSC) for its higher spatial resolution and temporal resolution, which is convenient for the snow quantitative analysis. Tibetan Plateau (TP) is not only an important snowfall region in China, but also one of the important regions influencing the global climate. It is rather difficult to retrieve snow cover fraction in TP bacause of the complex terrain there. Snowmelt provides precious fresh water resource for people, and the distribution of snow cover is cosely related with global climate change. Therefore, FSC mapping of TP area using MODIS data has great significance.Considering the issues of the MODIS standard fractional snow cover product (MOD10A1), a segment model (NDSITR) between NDSI and FSC was built in TP based on the principle of NDSI threshold method, and the extraction of endmembers algorithm was optimized, four kinds of spectral unmixing algorithms including Fully Constrained Least-Squares (FCLS), Fully Constrained Scaled Gradient Method (FCSGM), Sparse regression unmixing (SPARSE) and Polynomial unmixing (POLY) were used to unmixing, then a new FSC product with more accuracy in TP was produced. Using Landsat 30-m observations as "ground truth", the accuracis of these six kinds of FSC products including MOD10A1 were validated and analyzed. Furthermore, the errors of these FSC products combined with the terrain factors were analysed. The results showed that:(1) The NDSITR can retrieve the snow cover information in TP more comprehensively and accurately than MOD10A1, especially, in the edge of snow cover, the transition region and the fragmentary distribution area of snow. While the NDSITR overestimates the snow cover area much more than MOD10A1, their average negative error is-0.225 and -0.201. The Root Mean Square Error (RMSE) of the NDSITR and MOD10A1 is 0.146 and 0.209, the correlation coefficient is 0.824 and 0.77, respectively, which means the segment model NDSITR is more accurate and stable than that of MOD10A1.(2) Taking Landsat FSC as snow cover truth to verify and analyze these six kinds of FSC products. The results showed that the SPARSE gets the highest accuracy and the lowest error, secondly the FCLS and POLY, then the NDSITR and FCSGM, while the MOD10A1 has the lowest accuracy. The RMSE of these products are in order of 0.14,0.147,0.15,0.162,0.166, 0.191, and the correlation coefficient is 0.784,0.761,0.76,0.751,0.708 and 0.59, respectively. The average error of the NDSITR is always a negative value in this study, which means the NDSITR tends to overestimate snow cover.(3) Using spectral unmixing algorithms to map a long time series of snow cover is not a easy work at present, which is time consuming and inefficient. Thus, a time series spectral library is first established by choosing the expected spectra information using prior knowledge. It is a new idea in snow cover mapping and can obtain endmembers accurately, we also can establish a spectral library based on the prior knowledge of one year to produce snow cover products for other years, the retrieved FSC has better accuracy than that of MOD10A1, and the correlation coefficient with the Landsat truth reaches 0.8, which is the foundation for the quantitative analysis of remote sensing.(4) In complex topography condition, the elevation significantly influenced the accuracy of FSC than other terrian factors. It suggests that the elevation is the main factor which leads to the error, then is aspect and slope. Generally speaking, the spectral unmixing algorithms are all good at retriving snow cover, in which the SPARSE gets the minimum error and the FCSGM gets the maximum. Meanwhile, the NDSITR and MOD10A1 also have their own retriving advantages in special condition, the NDSITR has the best ability than other algorithms at high altitude areas above 5700m, and the MOD10A1 also has good retrival results in low altitude areas.