Research And Implementation Of Parallelized Hyperspectral Image Endmember Extraction Algorithm

Hyperspectral remote sensing images usually contain information on hundreds of bands with high spectral resolution.Due to the complexity of the features and the low spatial resolution of the remote sensor,there are a large number of mixed pixels in the hyperspectral image.The existence of mixed pixels can cause certain errors and effects on the traditional pixel-level remote sensing classification accuracy and target detection.At the same time,can not truly reflect the ground cover.Mixed pixel decomposition is an important solution to the above problems.The key step of the approach is endmember extraction.In addition,because the amount of hyperspectral image data is too much,and MapReduce technology is a programming model for parallel processing of large data volume.Therefore,this paper studies the extract endmembers in hyperspectral images based on unsupervised classification and convex geometry theory and gives a parallelization implementation scheme,the main research contents and results are as follows:(1)This paper focuses on the endmember extraction of hyperspectral image.The algorthim combines unsupervised classification with endmember extraction.Firstly,the hyperspectral data is unsupervised and classified by the binary K-means clustering algorithm for the spectral information of different features on the image.According to its material properties,the pixels with similar properties are clustered in the same class.The pixels in the same class have similar spectral features,but the spectral information of pixels in the same category is quite different.Secondly,use the binary K-means clustering algorithm to classify hyperspectral images,which avoids the random selection of the initial clustering center result in different classification results,there affecting the accuracy of endmember extraction.At the same time,according to the theory that the spectral shannon entropy the pixel with smaller spectral shannon entropy in each class are calculated and selected to be added to the candidate end set and the initial end.According to the theory of convex geometry to calculate the volume of the simplex volume,the pixel at the vertices of the largest simplex line is the endmember.The above method can quickly and effective extract endmembers in the hyperspectral image.(2)The amount of hyperspectral data is large,and the traditional binary K-means clustering algorithm takes a lot of time and space when performing unsupervised classification of hyperspectral data,this paper proposes parallelization of the binary K-means clustering algorithm to solve the problem.In addition,when performing the calculation of the volume of the simplex,due to the large number of matrix multiplication and determinant operations,these operations will occupy space and the running time of the algorithm,and reduce the operating efficiency of the endmember extraction algorithm.Based on the above problems,this paper proposes an algorithm for parallel processing of endmember extraction of hyperspectral image,so that the binary K-means clustering algorithm,matrix multiplication and determinant calculations parallelization under the MapReduce framework,and gives the parallelization scheme.The experimental results show that the parallelization of the endmember extraction algorithm will speed up the endmember extraction algorithm,reduce the space utilization rate,and have scalability.