利用Landsat 8多光谱数据的湿地信息提取方法比较研究

A Comparative Study on Methods of Wetlands Information Extraction Using Landsat 8 Multispectral Data

以淮河流域蚌埠段为研究区,利用2014年3景Landsat 8多光谱数据,采用支持向量机、决策树和面向对象分类方法,对研究区中的河流、湖泊、滩地和库塘（包括水田、水库、坑塘和水渠）进行遥感分类。由于各类型湿地在Landsat 8各波段影像中的光谱特征相似度较高,不利于分类,所以,在利用影像光谱数据的同时,借助于影像数据中的空间信息和温度信息,进行分类。研究结果表明,采用面向对象分类方法提取的河流、湖泊的用户精度较高;采用决策树分类方法提取的滩地、库塘的用户精度较高;采用支持向量机分类方法提取的河流、湖泊、库塘的用户精度较高,但是在河流、湖泊的分类中出现了椒盐现象,在滩地分类中出现了错分;与另两种方法相比,面向对象分类方法提取的河流或狭长湖泊末梢信息更准确。

With the Bengbu section of Huaihe River Basin as the research area, Support Vector Machine（SVM）, decision tree and object-oriented methods were used to classify the river, lake, floodplain and pond（including paddy field, reservoir, pothole and canal） using multispectral data of Landsat 8 images in 2014. As the spectral feature of all types of the wetlands in Landsat 8 images of each band was similar, therefore, while using spectral information, the spatial and temperature information should also be used to classify wetlands.The results showed that object-oriented classification method could acquire the higher user accuracy in the extraction of rivers and lakes. Using decision tree classification method to extract floodplains and ponds, the user accuracy was higher. SVM method gave a comparative high precision on rivers, lakes and ponds, but the＂salt-and-pepper＂phenomenon appeared in the classification of rivers and lakes and the misclassification in the floodplains. Compared with the other two methods, the object-oriented method was more accurate in extracting the information of the end of rivers or narrow lakes.