| The Yanhe Basin, as the experimental area of this study, has a complicated terrain surface and is located in the central of Loess Plateau that is well known all over the world for its deep loess deposits and serious soil erosion. To extract the information of land cover in sub-pixel level from Remote Sensing Data, firstly the paper used decision tree model to build two masks, and to every mask, selected different endmember models for the spectral unmixing of three Landsat Thematic Mapper imageries by using our IDL program. Based on the result of Linear Spectral Mixture Model (LSMM), sub-pixel classification of land cover was established. Then the fractional vegetation cover, the vegetation cover and the management factor were estimated. Finally, we assessed the accuracy of LSMM, the impact of heterogeneous surface and different methodologies. The main results are presented below.(1) Comprehensive application of SMACC and MNF rotation in our work suggested that the reasonable number of endmembers is five. The MNF transform was applied in this study, and the pixel purity index (PPI) was used to find the most spectrally pure in the image. Final endmembers were then selected with a reference from IKONOS images. Seven distinct endmembers were identified, i.e. forest, grass, farmland, bare soil, residence and road, dark lake/shadow and river. The average root mean squared error (RMSE) was 0.0129. The total accuracy of the sub-pixel classification was 75.56%, and the KAPPA Coefficient was 0.6354. Meanwhile, there is good linear relationship between the abundance of LSMM and the land cover of IKONOS. The correlation coefficient between the abundance of forest and TC2, NDVI respectively were 0.8529 and 0.9028, while the coefficient between the abundance of farmland and TC1, NDSI respectively were 0.7625 and 0.7441. Furthermore, the coefficient between the fraction of water and NDWI, NDMI respectively were 0.6192 and 0.6885. It is suggested that extracting the sub-pixel information by applying LSMM is feasible and promising.(2) This paper calculated vegetation fractions by the abundance of forest directly. This result had a significant linear relationship with NDVI. In the dense vegetation area especially, the rise of the vegetation fractions along with the increase of NDVI was faster than that in the moderate vegetation cover district. There was a strong linear relationship between the results of this study and the one that was estimated via the minimum and maximum of NDVI. It correlation coefficient reached 0.9025. In Yanhe Basin, the fraction of vegetation of LSMM was slightly smaller than that from NDVI, and the mean difference between them was 0.1484. This was mainly impacted by its significant relation with the choice of grass endmember. Experiment results indicated that the vegetation fraction can be derived by using LSMM with a promising accuracy and the LSMM can be able to achieve an improvement in the dense forest area.(3) This study estimated vegetation cover and management factor (C factor) based on the fractional abundance of all endmembers (LSMM-C). And the results (NDVI-C), calculated by using the vegetation fraction of NDVI and the experiential value of relevant type of land cover, were applied to comparisons. The experiment results showed that there was significant linear relationship between the two sides, whose correlation coefficient was as high as 0.9314. It is suggested that the end of LSMM was little higher than the other. The difference came from the methods that were applied for vegetation fraction and application of abundances of sub-pixel types of land cover. LSMM-C was greater than NDVI-C due to the difference of Vegetation factors and the difference were 0.0706. While, applications of fraction of sub-pixel leaded that LSMM-C were 0.0457 smaller than NDVI-C. All of the results show that the more heterogeneous at one pixel, the more differences between the two sides. Each type of land use has its own response to heterogeneity in the pixel. |
PDF Full Text Request