基于支撑向量回归的高光谱混合像元非线性分解  被引量:29

Unmixing Hyperspectral Imagery Based on Support Vector Nonlinear Approximating Regression

在线阅读下载全文

作  者:吴波[1] 张良培[1] 李平湘[1] 

机构地区:[1]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079

出  处:《遥感学报》2006年第3期312-318,共7页NATIONAL REMOTE SENSING BULLETIN

基  金:国家自然科学基金(40523005);国家973研究计划基金(2003CB415205);遥感科学国家重点实验室开放基金

摘  要:提出了基于支撑向量回归的高光谱混合像元自动分解。首先利用投影迭代的方法自动寻找到影像的典型地物光谱,然后利用Hapke近似函数模拟出非线性的训练和测试数据。支撑向量回归的混合像元分解方法与基于基函数分解方法的不同点是不需要预先确定非线性的映射形式,它通过核函数,把像元矢量从低维空间映射到高维特征空间,使得在特征空间中构造的线性光谱组合对应着原始空间(像元空间)的非线性组合特性,从而揭示了典型地物光谱之间的高阶性质,提高了混合像元的分解精度。实验结果证明,这种方法具有很高的混合像元的分解精度。利用模拟数据作分解精度的评价,表明97%以上的像元分解绝对误差不大于10%,而各类总体平均平方根误差均小于3.5%。Spectral Mixture Analysis (SMA) is a straightforward and efficient approach to the spectral decomposition of hyperspectral remotely sensed scenes, Once a SMA model is developed, land cover proportions can be estimated from pixel values through model inversion. In this paper, we propose to estimate abundances from hyperspectral image using support vector regression(SVR) method. SVR method for abundance estimation can be essentially regarded as function approximation and generalization problem. Differing from other nonlinear regressive approaches which require predefined nonlinear mapping functions, this method transferred each spectral pixel into a high-dimension feature space by a kernel function, which will result in a spectral pixel in a feature space consisting of possibly many nonlinear combinations of the spectral bands of the original spectral signature. In this way the higher order relationships between the mixed pixels are exploited in the feature space. Projection iterative method has been used for endmember abstraction from the image, and then simulating nonlinear training and testing data by Hapke' s approximation function. Experiment of simulating data and real hyperspectral image (Pushbroom Hyperspectral Imager, PHI ) are conducted to validate the procedures. The experiments show that the method can provide better result of abundance estimation for hyperspectral image as compared with that of radial basis function-neural networks. In our simulating test, over 97% of the total pixels in the image lie within the bound of ±0. 1, and the RMSE are no more than 3.5%.

关 键 词:光谱分解 迭代投影 支撑向量回归 非线性 

分 类 号:TP751[自动化与计算机技术—检测技术与自动化装置]

 

参考文献:

正在载入数据...

 

二级参考文献:

正在载入数据...

 

耦合文献:

正在载入数据...

 

引证文献:

正在载入数据...

 

二级引证文献:

正在载入数据...

 

同被引文献:

正在载入数据...

 

相关期刊文献:

正在载入数据...

相关的主题
相关的作者对象
相关的机构对象