Research On Nonlinear Mixed Pixel Decomposition Algorithms For Hyperspectral Remote Sensing Image

Hyperspectral remote sensing images have very high spectral resolution.Each pixel has dozens to hundreds of narrow bands in the visible to short wave infrared electromagnetic spectral range with relatively continuous spectral profiles.Hyperspectral images fuse spectral information characterizing objects and image information reflecting spatial geometric relationships,and have the ability to distinguish the spectral nuances of objects that panchromatic and multispectral remote sensing cannot provide.Hyperspectral remote sensing plays a very important role in the fields of agriculture,minerals,vegetation,aviation,medicine,ocean,atmosphere and environment.However,due to the spatial resolution limitation of hyperspectral remote sensor and the complexity of natural objects,the spectral information provided by a single pixel is superimposed by multiple pure objects at different scales,i.e.,mixed pixel,which brings a very great obstacle to the application of hyperspectral remote sensing images.Therefore,the primary problem of hyperspectral remote sensing image processing is mixed pixel decomposition.Mixed pixel decomposition usually includes two steps: endmember extraction and abundance estimation.In this paper,we conduct research on nonlinear spectral mixing factors in hyperspectral remote sensing images as follows.(1)Orthogonal projection endmember extraction algorithm based on distance measure is proposed.The algorithm combines the distance measurement method with the projection method to extract the object endmembers in the image one by one in a sequential manner.The algorithm first calculates the mean value of all sample data,takes the sample point with the closest distance to the mean point as the center point,finds the point with the largest distance to the center point as the initial endmember,then uses the extracted endmembers to construct the orthogonal projection operator,and projects all sample data into the orthogonal subspace to obtain the projection data,recalculates the center point of the projection data and the distance between all projection data and the center point,and takes the point with the largest distance measurement to the center point as the initial endmember.The sample data point with the largest distance measurement from the centroid is used as the next endmember,and so on,until a preset number of endmembers are extracted.The algorithm uses several distance measures such as Euclidean distance,Mahalanobis distance,Chebyshev distance,Manhattan distance,etc.The experimental results show that the endmember extraction results using Manhattan distance are better.In addition,the effect of using the sample mean as the centroid on the endmember extraction results is also compared and verified,and the experimental results show that the endmember extraction results are better using the centroid selection method in the paper.(2)Endmember extraction algorithm based on manifold learning is proposed.The algorithm uses the manifold structure to model the hyperspectral data to characterize the nonlinear factors in the spectral data,introduces the geodesic distance into the endmember extraction process,and extracts the object endmembers in the image one by one in a sequential manner using the orthogonal subspace projection technique.The algorithm selects the pixel with the farthest geodesic distance from the origin of the high-dimensional space as the initial endmember,and then projects the original spectral data onto the orthogonal subspace constructed from the extracted endmembers to eliminate the corresponding object endmember components in the mixed pixels.For the projected data,the geodesic distance is continued to be calculated and the pixel with the farthest geodesic distance from the origin is selected as the second endmember.The above steps are repeated until a preset number of endmembers are extracted.The K-nearest neighbor algorithm is used in the algorithm to calculate the geodesic distance,and a combination of ATGP and unconstrained least squares abundance estimation is used to approximate the data reduction of the original spectral data before the endmember extraction in order to reduce the computation.The experimental results of simulated and real image data show that the endmember spectra extracted by the algorithm can characterize the nonlinear factors in the spectral mixing,and the endmember extraction results are better than the traditional automatic target generation endmember extraction methods.(3)Based on the convex simplex geometric model and the proposed nonlinear endmember extraction method,nonlinear abundance estimation is performed using geodesic distances.The geodesic distance matrix of the original data is constructed,and the simplex volume composed of the endmembers is derived from the geometric model using the known endmember positions as a priori knowledge,and the new simplex volume is derived by replacing an endmember of the endmember simplex with a pixel,and its ratio to the volume of the endmember simplex is the abundance value of the replaced endmember in that pixel,and so on,find the abundance of all endmembers in each pixel.