Wetland Fine Classification Based On Sparse Constrained Least Squares Spectral Unmixing Algorithm

Wetland is one of the important ecosystems on the Earth,which has great ecological value and economic value.However,with the rapid development of human economy,wetland in China has been seriously damaged,which directly affects the growth and distribution of wetland vegetation.In recent years,people have realized the value of wetland and paid more and more attention to the protection and research of wetland vegetation.In the past,the classification of wetland vegetation is usually conducted based on multispectral remote sensing images,and the number of bands is limited,which makes it difficult to classify the wetland vegetation.The hyperspectral remote sensing image has higher spectral resolution,so it can obtain the information which is difficult to distinguish from the traditional low spectral resolution image data and improve the classification precision of the wetland vegetation.However,with the influence of sensor spatial resolution and the complexity of ground objects,mixed pixel is common,and the hyperspectral image with lower spatial resolution is especially serious,which not only affects the recognition and classification precision of ground objects,but also becomes the main obstacle to the quantitative development of remote sensing science.Taking Zhalong nature reserve as research area and using HJ-1A HSI hyperspectral image data,the performance of least squares spectral unmixing algorithm based on sparse constraint(SUFCLS)and traditional least squares linear spectral unmixing algorithm based on fully constraint(FCLS)are compared.The classification accuracy and errors of the two algorithms are evaluated and analyzed,and the following research results are obtained:(1)Because the traditional linear spectral unmixing algorithm uses the same group endmembers for each pixel in one scene image,it does not take into account whether the unmixing endmember is in the pixel field,especially in the complex scene.In this paper,the least squares spectral unmixing algorithm based on sparse constraint is proposed.The algorithm adaptively selects the highest percentage endmembers combination in the spectral library,and applied the selected endmembers in least squares linear spectral unmixing algorithm based on fully constraint to realize the abundance inversion.This method overcomes the shortcoming of the traditional linear spectral unmixing algorithm in the process of the endmembers selection.(2)For international important wetland Zhalong protected area,this paper constructs a classification system suitable for the land use in this area,and defined 10 land cover types,including reed swamp,carex swamp,sheep meadow,weed meadow,clear water surface,saline-alkali soil,paddy field,dry land,residential area and road.By means of field sampling,the spectral database of 10 land cover types are constructed,which is used for the least squares linear spectral unmixing based on sparse constraint and the accuracy comparison.(3)In this paper,the least squares spectral unmixing algorithm based on sparse constraint and the least squares linear spectral unmixing algorithm based on fully constraint are applied to the HJ-1A HSI hyperspectral image data of zhalong wetland respectively.The RMSE index was utilized to evaluate the classification accuracies.The results showed that the correlation coefficient between least squares spectral unmixing algorithm based on sparse constraint and the wetland vegetation communities(reed swamp,carex swamp,sheep meadow and weed meadow)visually interpreted from high-resolution imagery was higher than the least squares spectral unmixing algorithm based on fully constraint.The RMSEs are also improved obviously,which indicated that the SUFCLS has a potential advantage in the wetland information extraction.