基于QUEST决策树的遥感影像土地利用分类——以云南省丽江市为例

土地利用分类精度直接决定土地利用/土地覆被变化相关研究的准确性,而基于决策树的遥感影像分类是近年来提高土地利用分类精度的重要方法。QUEST决策树在影像解译和空间表达方面,运算速度和分类精度均优于普通CART等决策树方法。本文以云南丽江地区为例,应用QUEST决策树分类方法,对该地区的Landsat TM 5影像图进行分类,同时将地形因素、植被指数作为地学辅助数据的因子添加到分类波段中,进行不同特征融合,来处理目标类别间的非线性关系,该方法在处理图像理解知识方面具有更大的灵活性;同时与普通决策树分类法的遥感影像分类的结果相比较,Kappa系数值从原来的0.789提高到0.849.在地形复杂的山地地区,针对TM影像数据,选择基于QUEST决策树分类能够有效提高土地利用分类结果精度。

Research on the accuracy of TM images land-use classification based on QUEST decision tree: A case study of Lijiang in Yunnan

The accuracy of research on land use/cover change (LUCC) is determined directly by the accuracy of land use classification derived from aerial and satellite images.In analysis of the factors of accuracy of current remote sensing image classification, some methods were introduced to study new trends of classification modes.Some previous studies showed that the speed and accuracy of QUEST (Quick, Unbiased, and Efficient Statistical Tree) decision tree classification were superior to those of other decision tree classifications. On the basis of this approach, the research classified the Landsat TM-5 images in Lijiang, Yunnan province.This paper compared the result with that of maximum likelihood image classification.The overall accuracy was 90.086%, which was higher than the overall accuracy (85.965%) of CART (Classification And Regression Tree).Meanwhile, the Kappa efficient was 0.849, which was higher than the Kappa efficient (0.760) of CART. Therefore, it is concluded that in the complex terrain area such as in mountainous regions, the choice of QUEST decision tree classification on TM image would improve the accuracy of land use classification. This type of classification decision tree can precisely obtain new classification rules from integrated satellite images, land use thematic maps, DEM maps and other field investigation materials.Simultaneously, the method can also help users to find new classification rules in multidimensional information, and to build decision tree classifier models. Furthermore, the methods, including a large number of high-resolution and hyperspectral image data, integrated multi-sensor platform, multi-temporal remote sensing image, the pattern recognition and data mining of spectral and texture features, and auxiliary geographic data, will become a trend