基于布谷鸟搜索算法的高光谱图像解混算法  

Hyperspectral image unmixing algorithm based on cuckoo search algorithm

作　　者：孙彦慧[1] 张立毅[1,2] 陈雷[3,2] 李锵[1] 滕建辅[1] 刘静光 

机构地区：[1]天津大学电子信息工程学院,天津300072 [2]天津商业大学信息工程学院,天津300134 [3]天津大学精密仪器与光电子工程学院,天津300072

出　　处：《光电子．激光》2015年第9期1806-1813,共8页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61401307);中国博士后科学基金(2014M561184);天津市应用基础与前沿技术研究计划(15JCYBJC17100)资助项目

摘　　要：将独立成分分析(ICA)算法用于高光谱图像解混时,算法对丰度的独立性要求与实际地物分布相矛盾;同时,采用梯度算法对解混目标函数进行优化时,易收敛到局部极值点。针对上述问题,提出在非负ICA(NICA)模型的目标函数中引入丰度和为一约束(ASC),确保解混出的丰度与实际地物分布一致;同时,采用布谷鸟搜索(CS)算法,利用其优异的全局搜索性能对提出的目标函数进行优化求解。为减少参数维数并缩小CS算法的搜索范围,利用矩阵QR分解理论,将对解混矩阵的搜索转化为对一系列Gives矩阵的识别。仿真数据和真实高光谱图像数据实验结果表明,提出的算法能有效地克服上述问题,在噪声为30dB、像元纯度为0.8时,解混指标光谱角距离(SAD)和均方根误差(RMSE)达到了0.03以下,达到良好解混效果。Independent component analysis （ICA） is a typical blind source separation algorithm. Nowa- days,it has been applied to solve the problem of hyperspectral unmixing. Under linear mixture model, a pixel spectrum of hyperspectral image can be approximated to a collection of constituent spectra, called endmember and a corresponding set of fractional abundances, one set per pixel. However, the actual ground-object distribution demands that the abundance should satisfy both abundance nonnegativity con- straint （ANC） and abundance sum-to-one constraint （ASC）. Thus, the independence requirement of ICA conflicts with these two constraints and itls easy to converge to local minima with gradient algorithm to optimize the relevant objective function. To solve this problem,we propose to combine the non-negative independent component analysis model with abundance sum-to-one constraint to construct a novel objec- tive function, and introduce cuckoo search （CS） algorithm to optimize the function with its excellent global searching ability. Experimental results on synthetic data and real hyperspectral data indicate that the proposed algorithm can effectively solve the above problems and obtain more accurate results without any spectral prior knowledge. When signal-to-noise ratio （SNR） is set to 30 dB and purity of pixel is set to 0. 8, these unmixing indexes, which are spectral angel distance （SAD） and mean square error （RMSE） ,can reach below 0.03.

关 键 词：高光谱图像解混 非负独立成分分析(NICA) 丰度和为一约束(ASC) 布谷鸟搜索(CS)算法 QR分解 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]