高光谱遥感图像光谱解混的独立成分分析技术

Independent Component Analysis for Spectral Unmixing in Hyperspectral Remote Sensing Image

高光谱遥感在对地球陆地、海洋、大气的观测中发挥着重要作用,高光谱遥感图像分析的关键是提取像元光谱内部各物质成分及其含量,即光谱解混。独立成分分析提供了一种先进的技术手段,在很少先验知识的前提下,实现端元(物质成分)光谱及其丰度(含量)的同时提取。但丰度约束破坏了各成分独立的前提条件,导致了独立成分分析的局限性。针对这一问题,提出了丰度约束下总体相关性最小化的解决方案,并指出总体相关性最小化下的理想角度,通过设计角度修正的独立成分分析算法把各成分调整到理想角度上。利用模拟数据与真实数据算法进行检验,结果表明:经过角度修正后,独立成分分析突破了原有的局限性,有助于进一步提高独立成分分析技术在光谱分析中的有效性。

Hyperspectral remote sensing plays an important role in earth observation on land,ocean and atmosphere.A key issue in hyperspectral data exploitation is to extract the spectra of the constituent materials （endmembers） as well as their proportions （fractional abundances） from each measured spectrum of mixed pixel in hyperspectral remote sensing image,called spectral unmixing.Linear spectral mixture model （LSMM） provides an effective analytical model for spectral unmixing,which assumes that there is a linear relationship among the fractional abundances of the substances within a mixed pixel.To be physically meaningful,LSMM is subject to two constraints：the first constraint requires all abundances to be nonnegative and the second one requires all abundances to be summed to one.Independent component analysis （ICA） has been proposed as an advanced tool to unmix hyperspectral image.However,ICA is based on the assumption of mutually independent sources,which violates the constraint conditions in LSMM.This embarrassment compromises ICA applicability to hyperspectral data.To overcome this problem,the present paper introduces a solution of minimization of total correlation of the components.Interestingly,with the minimization of total correlation of the components,the angle of the direction between each components is invariable.A Parallel oblique-ICA （Pob-ICA） algorithm is proposed to correct the angle of the searching direction between the components.Two novelties result from our proposed Pob-ICA algorithm.First,the algorithm completely satisfies the physical constraint conditions in LSMM and overcomes the limitation of statistical independency assumed by ICA.Second,the last component,which is missed in other existing ICA algorithms,can be estimated by our proposed algorithm.In experiments,Pob-ICA algorithm demonstrates excellent performance in the simulative and real hyperspectral images.