基于背景类别分层分离的不透水面间接提取——以兰州城关区EO-1 ALI图像为例

以兰州市城关区EO-1 ALI图像作为基础数据源，对比LOOC与最邻近分类（NNC）方法提取不透水面的精度差异. LOOC方法对融合后EO-1 ALI图像进行LBV变换，将变换结果作为解译底图；设置4个尺度层次，分别对应水体、农用地、灌木林地、城市绿地和草地这5种主要背景类别，对解译底图执行4尺度面向对象分割；将上述类别对象的光谱特征和形态特征差异作为判别规则，利用决策树分类，将这5种背景类别依次从解译底图上提取、分离，生成不透水面初级提取图层；通过光谱反射率差异分析，选定EO-1 ALI图像的近红外波段8和中红外波段10作为分类特征，利用基于模糊C-均值（FCM）算法的非监督分类，从初级图层中分离出砂土、阴影这两种与高、低反照度不透水面光谱特征相近的类别，采用数学形态学开闭运算整饬图像，生成不透水面二级提取图层. 结合目视评判和总体精度、Kappa系数，定量分析LOOC方法与NNC方法的提取精度差异. 结果表明：LOOC方法提取不透水面的总体精度、Kappa系数分别为87.13%、0.830 3，较NNC方法分别提高5.91%、7.19%. LOOC方法依据各背景类别的遥感多特征知识，分两级将其分离出解译底图，从而间接、逐步逼近不透水面精准空间分布信息，辨识不透水面的效率优于NNC方法.

Impervious surface extraction based on stratified removal of contexts types-take EO-1 ALI image of Chengguan county in Lanzhou as example

The most suitable method for impervious surface extraction was adopted by using EO-1 ALI image of Chengguan County in Lanzhou. The supervised classification based on LOOC and nearest neighbor classification (NNC) was compared. LBV transform was conducted for EO-1 ALI image. Then the image was offered as interpretation base map for further processing. Four scale levels were set in interpretation base map to match 5 kinds of contexts types, including waters/farmland/shrub land/urban green space and meadow. Then object-oriented segmentation was conducted in interpretation base map to get polygon objects of each context types. The decision tree classification was conducted to dislodge out cover information of above-mentioned contexts types in interpretation base map based on spectral features and shape feature of polygon objects of each context type. Then primary layer of impervious surface was obtained from interpretation base map. The unsupervised classification based on fuzzy C-mean (FCM) algorithm was carried to dislodge out another 2 kinds of contexts types from primary layer of impervious surface, including shadow of landform and sandy soil. Band 8 and band 10 of EO-1 ALI image was selected as effective feature in the process of unsupervised classification. Then secondary layer of impervious surface was obtained. The subject and object standard were adopted to evaluate the classification precision for three methods, including LOOC/NNC. Results showed that classification accuracy and Kappa coefficient was 87.13%, 0.830 3 by using the method of LOOC. Growth of 5.91% for overall classification accuracy and 7.19% for Kappa coefficient were achieved compared with NNC. The method of LOOC can provide more accurate extraction results for impervious surface using EO-1 ALI image compared with NCC.