Green Cover Extraction Based On Mixed Pixel Decomposition Using Logit Model

Now urban forest construction, which can improve the ecological carrying capacity of the city, has become one of the most pressing problems for development in a modern urban. Understanding the distribution of the green space of the whole city is important for making a plan to grow the forest, In order to understand green space information, pre-investigation is generally required. But in the green survey with remote sensing images, low resolution images are not applied widely because of the existence of mixed pixels. Especially in southeast coastal cities, there are more developed cities and suburbs with a higher fragmental level. The influence of mixed pixel is more serious in these cities. So, it is necessary to solve mixed pixel decomposition and make remote sensing application reach sub-pixel level from pixel level and enter pixel inside in order to improve the accuracy of green coverage rate with remote sensing images.This study selected four districts in Yuhang City, in which it is the established district, as the study area. A Landsat Thematic Mapper (TM) image of May 2,2008 was selected to process, In order to improve classification accuracy, multinational logit model has been used to handle the mixed pixel problems. According to the administrative boundary, the study area were divided into built up area and suburb and logit models were used to decompose them separately. Then, the accuracies of green land extraction by logit model with zoning or not and non-constrained linear mixing model were compared, the conclusions are as follows:(1) To obtain a good result, it is essential to comprehensively use various methods in the endmembers' extraction. In this paper, five endmembers (named high reflective geophysics, low reflective geophysics, bare soil, green land and farmland) were refined by a minimum noise fraction (MNF) transform, pixel purity index(PPI), n-Dimensional Visualizer and artificial visual interpretation.(2) To achieve a good classification effect, collinearity was weakened by putting the remote sensing data after MNF transform as the input sample to estimate regression coefficients. This paper compared the pre-classification results with different samples that original image and the data after MNF transform, and obtained that MNF transform data as sample can get relatively good results.(3) By comparing the results of mixed pixel decomposition with nonlinear Logit model and non-constrained linear mixing model, it was found the decomposition of Logit model was more actual. And comparing the mixed pixel decomposition used Logit model with zoning or not, the method with zoning got a better result. Visible, mixed pixel decomposition used nonlinear Logit model with zoning, which to a certain extent solving mixed pixel problems, could improve the accuracy of green coverage extraction.