证据推理的遥感图像分类方法与应用

Evidential Reasoning Remote Sensing Image Classification Method and Application

遥感图像分类一直是遥感研究领域的核心问题。然而,传统遥感分类方法在地物复杂地区不能取得满意效果。不但分类精度不够理想,分类灵活性也存在不足。本文尝试引入证据推理"软"分类方法,选择小兴安岭山区一景TM遥感图像,基于用户知识和经验,通过人机交互处理,以累计信任度(CBV)最大为划分像元归类的原则得到证据推理方法的分类图像。结果表明,整体分类精度从最大似然法的78.74%提高到82.28%,kappa系数从0.67提高到0.71。但该方法对于裸地分类精度不高,通过人为设定各类别CBV阈值的方法,获得了人为干预的证据推理方法分类图,其整体分类精度达到了87.80%,kappa系数也达到了0.81,所有类别的生产者和用户精度相比于最大似然法都有提高。研究表明,证据推理方法在遥感分类精度和分类灵活性方面都具有优越性。

Remote sensing classification is always a core problem in the remote sensing research field.However,traditional classification method should be caught into question when the context information is complex.Besides the classification accuracy is not ideal,there is also a classification flexibility problem.This paper introduced an evidential reasoning ＂soft＂ classification method which has a high degree of flexibility for classification.The results can be adjusted easily for improved classification accuracy based on users＇ knowledge and experience combined with visual interpretation.We chose a TM image of Xiaoxinganling mountain area to assess the performance of evidential reasoning classification method.The TM image spectral characteristic of typical surface categories of the study area was analyzed,and the results showed that the spectra overlap phenomenon was serious for each band.We chose the 4,5,7 band with light spectra overlap phenomenon as evidence sources.Through the combination rule of evidence,we got distribution maps of cumulative belief value for different categories.Then we produced a theme classification map with the maximum cumulative belief value as pixels divided rule.We compared classification accuracy between maximum likelihood classification（MLC） and evidential reasoning（ER） based on confusion matrix,and the results illustrated that the overall accuracy improved from 78.74% of MLC to 82.28%,the kappa coefficient also improved from 0.67 to 0.71.For a better accuracy,we set different CBV thresholds of different categories based on visual interpretation and ＂soft＂ classification mode,and produced a higher classification accuracy map with 87.80 % overall accuracy and 0.81 kappa coefficient.This study proved that the evidential reasoning method has more advantages than traditional the method,not only in its classification accuracy,but also in its flexibility in remote sensing classification processing.