论文部分内容阅读
对比了线性混合光谱分解模型(SMA)与支持向量机(SVM)在TM影像上估算不透水面覆盖率(ISP)的精度,通过SVM模型拟合TM像元光谱特征与样本ISP间的关联而获得对未知像元ISP的估算能力。对于天津市主城区的TM影像,选择学校区、工矿区和住宅区的高分辨率影像分类结果作为训练样本(7020个)和验证样本(1500个),SVM模型的ISP估算均方差(15.4%)优于SMA估算结果(19.4%);在增加缨帽变化“绿度分量”及混合光谱分解“高反射率分量”作为SVM特征变量后,ISP估算精度提高为12%。研究结果表明:SVM模型能够拟合各像元光谱组分间非线性关系且具有较好小样本泛化的性能,适用于地面样本较少的大区域ISP制图;增加与ISP相关性大的光谱特征向量作为SVM输入能提供更多的区域地物空间分布信息,能够调整无样本的地表类型的ISP估算值,提高区域ISP估算的整体精度。
The accuracy of estimating impervious surface coverage (ISP) on TM images by linear mixed spectral decomposition (SMA) and support vector machines (SVM) was compared. The SVM model was used to fit the relationship between spectral characteristics of TM pixels and sample ISPs Obtain an estimate of unknown pixel ISP. For the TM image in the main city of Tianjin, the high resolution image classification results of school area, industrial and mining area and residential area were selected as the training samples (7020) and the verification sample (1500), and the ISP estimation of the SVM model was 15.4% ) Was better than that of SMA (19.4%). The ISP estimation accuracy was improved by 12% with the increase of tassel cap change “greenness component ” and mixed spectral decomposition “high reflectance component ” as SVM characteristic variables. The results show that the SVM model can fit the non-linear relationship between spectral components of each pixel and has good performance of small sample generalization, and is suitable for ISP mapping in large area with less sample on the ground. As a SVM input, the eigenvector can provide more spatial distribution of regional features, adjust the ISP estimates of the surfaceless samples without samples, and improve the overall accuracy of regional ISP estimates.