The land cover mapping with airborne hyperspectral remote sensing imagery in Yanhe river valley

The hyperspectral remote sensing can detect the more detailed objects on earth surface for its hundreds of wave bands. This make it can be used for very detailed land cover mapping. The study area Yanhe river valley locates on the north west of China, its size of area is 7725 km2 and length 286.9km. The most objects on the area are vegetations including crops, trees, grasses, and water. For test the real accuracy of mapping with airborne hyperspectral remote sensing imagery, parts of the aerial surveying area were investigated and mapped manually during the photographing time. Because of the soil erosion and drought, the land covers are usually mixed especially the crops mixes with grasses, the tree seedling mixes with grasses and withered seedling, some fruit trees mix with crops etc. This is the big challenge of precise land cover mapping with hyperspectral remote sensing imagery. The present work focuses on the detailed vegetation classification problems especially the easy mixed classified objects. From the field investigation, some areas are not covered by single vegetation class but mixtures of multi-species of vegetation. This phenomenon may cause mixed pixel and misclassified problem or even unmixed pixel and right classified but the confused vision in map for the mixed colors in a continuous site of land cover. The general method to solve the mixed pixel problem is the spectral unmixing. But there are some questions in classical unmixing model, one of them is the equation of unmixing model is usually ill-conditioned because of the heavy correlation among hundreds of bands. The research proposed a feature subspace based spectral unmixing model. The unmixing was based on a feature subspace of full data space, the aim is to remove the amount of correlate information and get the right solution of components of endmemeber spectra. The subspace was constructed by principal component analysis(PCA) directions and the easy mixed classes oriented projection pursuit (- - PP) directions. For the unmixing model, the research discussed the results of unconstrained and constrained solutions. Based on the unmixing results, the land cover map was achieved and the accuracy was verified.