Synchronous, asynchronous and grouping asynchronous parallel implementation for N-FINDR algorithms in hyperspectral remote sensing image

Abstract

N-FINDR is a widely used endmember extraction algorithm in hyperspectral imagery. Nevertheless, its computational complexity is high. Plaza's parallel implementation of N-FINDR, namely, P-FINDR, demonstrates an excellent way to improve the computing performance of N-FINDR by incorporating with parallel computing technique. In this paper, three parallel implementation patterns, i.e., synchronous, asynchronous and grouping asynchronous pattern, are presented. Two versions of N-FINDR, i.e. iterative N-FINDR and successive N-FINDR, are considered to be implemented in parallel by these patterns respectively. Thus obtains six parallel N-FINDR algorithms. In experiment, both solution quality and parallel performance of these algorithms are compared and suitable patterns for parallel implementation of N-FINDR are obtained. IndexTerms-hyperspectral remote sensing, endmember, N-FINDR, parallel computing INTRODUCTIONSpectral unmixing in hyperspectral remote sensing image has been widely researched in the last decade [1]. Endmember extraction is the most important and challenging task in spectral unmixing. Recently, many endmember extraction algorithms (EEAs) have been developed. N-FINDR [2] algorithm is one of the most commonly used EEA and different implementation had been explored [3][4][5][6].Nevertheless, endmember extraction in a large volume hyperspectral remote sensing image is a time-consuming task which cannot meet the time requirement of many applications. The utilization of the high performance computing (HPC) technique into hyperspectral remote sensing analysis throws a new light on the hyperspectral image processing algorithm research and becomes more and more widespread today [7]. Recently, many state-of-the-art parallel EEAs had been developed. A parallel version of N-FINDR was developed in [8], which demonstrated high computing performance and greatly improve the efficiency of N-FINDR algorithm.

Fast Implementation of Maximum Simplex Volume-Based Endmember Extraction in Original Hyperspectral Data Space

“…Thirdly, sequential implementation, instead of finding endmembers all together simultaneously, was adopted to short cut the searching processing [17], [18], [21]. Fourthly, parallel implementation skill was also used to speed up the algorithm [23], [24]. Additionally, some other references applied some preprocessing before EE to speed up the algorithm: [18] narrowed down the search region, [20] divided all pixels into N classes, and [22] eliminated most of sample vectors.…”