基于松弛因子改进FastICA算法的遥感图像分类方法  

Remote Sensing Image Classification Based on a Loose Modified FastICA Algorithm

作　　者：王小敏[1] 曾生根[1] 夏德深[1] 

机构地区：[1]南京理工大学计算机科学系,南京210094

出　　处：《计算机研究与发展》2006年第4期708-715,共8页Journal of Computer Research and Development

摘　　要：多波段遥感图像反映了不同地物的光谱特征,其分类是遥感应用的基础.独立分量分析算法利用信号的高阶统计信息,去除了遥感图像各个波段之间的相关性,获得的波段图像是相互独立的.然而独立分量分析算法计算量太大,影响了其在多波段遥感图像分类上的应用.MFastICA算法可以改善FastICA算法的性能,减少计算量,但是同FastICA算法一样,其收敛依赖于初始权值的选择.在MFastICA算法中引入松弛因子,使算法可以实现大范围的收敛.应用BP神经网络对独立分量分析算法预处理后的图像进行自动分类,其分类精度比原始遥感图像的精度高,并且3种独立分量分析算法的最终分类性能相当.The multi-band remote Sensing images reflect the spectral features of diverse surface features, and the classification is the basis of remote sensing applications. The independent component analysis （ICA） algorithm uses the high-order statistical information of multi-band remote sensing images. It not only removes the correlation of images, but also obtains the new band images that are mutual independent. But the computational complexity of FastlCA is too big, influencing the application of ICA in remote sensing field. M-FastlCA algorithm could improve the performance of FastlCA algorithm by reducing the computational quantum. But like FastlCA, its convergence is dependent on initial weight. By importing loose gene in the M-FastlCA algorithm, the new algorithm （LM-FastlCA） could implement convergence in large-scale. BP neural network is used in classification of the remote-sensing images which are pre-processed by ICA. The exactness rate of pre-processed images is higher than that of source images, and the performance of classification of three kinds of ICA algorithms is near.

关 键 词：独立分量分析 FASTICA M-FastICA LM-FastICA 遥感图像分类 BP神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]