Research On MODIS Cloud Fraction Estimation Method Via Spectral Unmixing

Cloud plays an important role for controlling the atmospheric environment and climate change,and cloud fraction is one of the main influence factors. MODIS can be used to long-term monitoring for land, atmosphere, ocean and so on, its multiband data can provide many information that reflect the characteristics of cloud,but due to the sensor design requirements of large scope, high spectral resolution information acquisition,it causes a lower spatial resolution of MODIS remote sensing image,and the remote sensing cloud images that we got often exist mixed pixels. In addition,the spectrum of cloud pixels are highly related to the fraction,height,and thickness of cloud,which caused a huge within-class variance,the traditional unmixing methods can’t got a very good result. Therefore,this paper takes the above questions into consideration,starting with the characteristics of cloud,with the help of MODIS remote sensing image and the CALIPSO data that can provide cloud fraction information, based on the spectral unmixing technology,extending the study of MODIS cloud fraction estimation via spectral unmixing technology. Including the following aspects in detail:First of all, this paper mainly studied the characteristics of cloud in the MODIS remote sensing image and the spectral feature extraction of cloud. Through the analysis of the characteristics of cloud and the bands in MODIS that cloud changes prominent,we extract three groups of features,which can effectively distinguish the cloud and other background to complete the feature extraction of cloud.Then, this paper translated the problem of cloud fraction estimation into the spectral unmixing problem,studied the linear spectral unmixing methods based on the nonnegative matrix factorization and its improved algorithms,improved the precision of the cloud fraction. This paper used MODIS all bands and cloud feature bands these two groups of data made some relative experiments for these linear spectral unmixing methods. In the experiment, we used principal component analysis method to determine the number of endmembers,the vertex component analysis and full constrained least square methods to finish the initialization of endmembers matrix and abundance fractions matrix. Finally, according to the percentage of cloud pixels in the abundance fractions matrix from the unmixing solution, completed the cloud fraction estimation,and also verified the advantages of cloud feature bands to unmixing.At last,according to the characteristics of cloud,we found that the influence of fraction,height,and thickness of cloud would cause a huge within-class variance,the traditional linear mixed methods didn’t apply to it. Therefore,this paper based on the nonnegative matrix factorization,introduced the kernel method and combined with the spectral structure information of data,achieved spectral unmixing based on the kernel graph regularized nonnegative matrix factorization. Further,we used the multiple kernel to substitute the single kernel learning, made full use of the advantages of different single kernel,effectively improved the precision of cloud fraction estimation. This paper also studied the spectral unmixing methods that used of cloud features,for further improved the precision of cloud fraction estimation,and through the kernel alignment technology at the same time,analyzed the effect of different cloud features to the cloud fraction estimation.