基于QUEST树的高寒湿地植被覆盖类型遥感分类研究

Study on remote sensing classification of highland wetlands vegetation coverage by QUEST-based decision tree

以青海省索加-曲麻河自然保护区高寒湿地密集分布区为例,进行三江源高寒湿地植被覆盖类型遥感分类方法研究。基于SPOT5影像融合后的4个波段数据和融合影像主成分分析后第一主成分PCA1,以PCA1作为数据源计算的8个纹理特征,用数字高程模型DEM和其他辅助数据构成的数据集,选用QUEST算法,对数据集进行数据挖掘,构建决策树模型对影像进行分类。分类结果表明:9×9窗口下的纹理特征均值、方差和信息熵是进行高寒湿地植被覆盖类型分类的有效纹理特征量。QUEST决策树分类方法所得分类结果精度可达84.19%,kappa系数为0.826 1。相对于传统的最大似然法监督分类所得结果,总体精度提高17.29%,总体kappa系数提高0.191 2,提高幅度近30%。所得分类结果中,矮嵩草草甸、矮嵩草+藏嵩草沼泽化草甸、杉叶藻草本沼泽、藏嵩草沼泽化草甸和藏嵩草+矮嵩草草甸等植被覆盖类型的分类精度较高,生产者精度分别可达:95.83%,95.45%,95.00%,94.74%和91.67%;用户精度分别为100%,91.30%,82.61%,78.26%和95.65%。

By taking Suojia-Qumahe Nature Reserve in Qinhai Province as a studying example,where wetlands distribute intensively,to discuss the proper method for remote sensing classification of highland wetlands vegetation coverage.Based on the data set which including spectral and texture characteristics,the first component of Principal Component Analysis,and DEM and other ancillary data,the decision tree model was established by QUEST algorithm to classify different vegetation types in SPOT5 imagery.The classification results show that the texture indices mean,variance and entropy were calculated by GLCM under 9×9 window size were effective in distinguish different vegetation communities of highland wetlands.The overall accuracy of this method was 84.19%%,and the kappa coefficient was 0.8261.It is evident that the QUEST-based Decision Tree method can improve the overall accuracy by 12.05%,and the overall kappa coefficient by 0.1407 when compared with the result obtained by traditional maximum likelihood supervised method.The vegetation communities such as： Kobresia humilis Meadow,Kobresia humilis and Kobresia tibetica swampy meadow,Hippuris vulgaris marshes,Kobresia tibetica Meadow,Kobresia tibetica and Kobresia humilis meadow had higher accuracy than other types.The producers and users accuracy of these vegetation communities can reach 95.83%,100%;95.45%,91.30%;95.00%,82.61%;94.74%,78.26% and 91.67%,95.65% respectively.