沙漠化遥感分类方法对比分析——以新疆塔河下游为例

Comparative Analysis of Remote Sensing Classification Method about Desertification——Take the downstream of Tahe in Xinjiang as an example

新疆是我国受沙漠化危害最严重的省区,而沙漠化信息提取是遥感专题信息提取的难点之一,探索合适的沙漠化遥感分类方法可以为相关部门提供决策支持。本文利用TM&ETM遥感影像对比分析了现有主要应用的最大似然、最小距离、决策树等几大分类方法在沙漠化分类中的应用,结果显示:①传统的最大似然方法分类精度及出图效果等各方面优于SVM、最小距离及决策树方法。最大似然的分类总精度以及Kappa系数分别达到96.43%和0.95,分类精度随先验概率的增大而减小,混淆程度小,结果图清晰,能反应真实的沙漠化分布情况。②SVM径向基函数分类可以达到分类目的,当γ逐渐增大时精度增高,混分较稍严重,可以反应真实的沙漠化分布情况。③最小距离精度较低,有严重的漏分现象,能一定程度上反应沙漠化的分布情况。④决策树分类法的精度低,分类指数筛选复杂,分类结果图并不能很好的反应真实的沙漠化分布情况。

Xinjiang is the most desertification affected province in China. And desertification information extraction is one of the difficulties of remote sensing thematic information extraction. So to explore suitable desertification remote sensing classification method can provide support for the policy - making of relevant departments. In this paper, TM ＆ ETM remote sensing image is used to compare applications of several major classification methods - - Maximum likelihood, minimum distance, decision trees, etc. , in desertification classification. The results showed that：（1）Traditional maximum likelihood classification method is better than SVM, the minimum distance and decision tree method at accuracy and plot effects. Total classification accuracy and Kappa coefficient of maximum likelihood classification method respectively reached to 96.43% and 0.95. The classification accuracy reduced with the increase of priori probability. It has low degree of confusion, clear results figure, and it is able to respond to the real desertification dis- tribution. （2）SVM radial basis function classification can achieve the purposes of classification. The classification accu- racy gradually increased with increases of ~. It has a little more serious confusion classification, and can reflect real desertification distribution. （3）Minimum distance has lower classification accuracy, and serious missing phenomenon. It can reflect desertification distribution in a certain extent. （4）Decision tree classification has lower classification accu- racy too, and complex filter of sub - index but it can not reflect real desertification distribution.