Spectral unmixing of MODIS data based on improved endmember purification model: application to forest type identification

Abstract

Because of high spectral and temporal resolutions, large coverage and low cost, MODIS (Moderate Resolution Imaging Spectroradiometer) data has been widely used to extract information of forest types at regional, national and global scales. However, its coarse spatial resolution often leads to mixed pixels and impedes increasing classification accuracy of forest types. Spectral unmixing can, to some extent, increase the accuracy of classification. But, how to accurately extract pure endmembers for a study area is a great challenge. The selection of linear or non-linear spectral unmixing algorithm is another challenge. In this study, a method to extract endmembers - different land cover and vegetation types from MODIS images was developed. In this method, the time series of MODIS derived vegetation index was first obtained and the phenological variation of forest types were analyzed. Moreover, decision tree classification was then conducted and the obtained results were then used to enhance the extraction of endmembers. With the endmembers, linear spectral unmixing of MODIS images with and without constraints, and nonlinear spectral unmixing were finally carried out and the classification results were compared. In addition, for comparison, the classification was also made using a widely used classifier - maximum likelihood. This study was conducted in Hunan of China, where typical vegetation types included coniferous forests, deciduous forests, bamboo, and shrubs. Moreover, water, built up area, and agricultural lands were involved. The classification accuracy of the land cover types using MODIS images was assessed using the data from a total of 1179 forest inventory plots and the area data of the land cover classes from forest inventory across Hunan, and the classification results using Landsat Thematic Mapper™ images for Zhuzhou City of Hunan, respectively. The results showed that the overall accuracies for three kinds of validation data were 85.8%, 87.4% and 85.9% for linear spectral unmixing without constraints, 85.1%, 88.4% and 84.7% for linear spectral unmixing with constraints, 64.2%, 67.5% and 64.7% for nonlinear spectral unmixing, and 72.7%, 79.7% and 73.8% for maximum likelihood classifier. These implied that linear spectral unmixing regardless of with and without constraint led to much higher accuracy than the maximum likelihood classification and non-linear spectral unmixing.