以东湾流域为研究区,首先对TM影像进行主成分分析(PCA),缨帽变换,色调、亮度、饱和度彩色变换(HIS),并求得归一化植被指数、归一化建筑指数、归一化水体指数、归一化裸土指数、归一化阴影指数和第四波段灰度共生矩阵纹理特征,构造了新的特征影像;然后对其分别进行PCA、独立主成分分析(ICA)、最小噪声分离(MNF)3种变换;再对变换后的结果分别进行最大似然法、神经网络以及支持向量机(SVM)分类,以研究不同变换、不同分类方法的差异。提出了一种新的降低特征相关性的方法,并确定了最有效的一种分类—变换组合。结果表明,新特征降维方法的分类精度达到了93.457%,而单一特征降维方法最高分类精度为91.955%,证明该方法能更有效地提取影像特征。

Taking Dongwan Basin as the study area, PCA, TC, HIS transformation were performed on the TM image, and NDVI, NDBI, NDWI, BI, NDUI index and the co-occurrence matrix texture features of the fourth band were obtained, all of which composed the new features image. Performing PCA, ICA and MNF transformation on the new features image, the results of transformation were carried on the MLC, NNC, SVM classification respectively, in order to explore the differences of different transformation and different classification methods. This paper put forward a new method of reducing characteristics correlation and determined the most effective combination of classification-transform(MNF+PCA+ICA-SVM) simultaneously. The results show that the classification accuracy of new feature extraction method is 93.457%, while the highest classification accuracy of the single feature extraction method is 91.955%.The results indicate that the method can extract image features more effectively when compared with other single feature extraction methods.