基于自适应高斯混合模型的遥感影像分类方法研究——以武汉地区遥感影像分类为例  

Methods of Remote Sensing Image Classification Based on Adaptive Gaussian Mixture Model

作　　者：李登朝 吴健 许凯[2] 

机构地区：[1]湖北省地质局第一地质大队,湖北大冶435100 [2]中国地质大学(武汉)信息工程学院,湖北武汉430074

出　　处：《资源环境与工程》2015年第6期1014-1021,共8页Resources Environment & Engineering

摘　　要：利用遥感影像中地物目标的光谱和纹理特征,基于自适应高斯混合模型对地物特征的概率密度函数进行建模,通过自适应消除最小权重高斯子分量的方法获取分类样本最佳高斯分量数,并基于贝叶斯分类准则进行地物类别的判别,实现遥感影像分类。最后,通过对TM、Quickbird分类实验证明了该算法的有效性,并与传统的图像分类算法进行试验对比,验证了该算法的可行性和优越性。The paper develops an innovation classification method for remote sensing images which can adaptively obtain the optimal number of Gauss components based on spectral feature of target objects. The method uses the excellent mathematical properties of the GMM, i. e. its ability to approximate any kind of probability density distribution, to integrate spectral and texture features of remote sensing images firstly. Then retrieving optimal Gaussian components through A- daptive Gaussian elimination of minimum weight classification sub-component sample. Lastly the authors do Gaussian mixture model clustering based on Bayesian classification criteria, which can finalize the classification result. Finally, the new classification method has been successfully applied to TM （Moderate Resolution Images）, Quickbird （high spatial resolution images） and other different types of remote sensing images. By a series of comparison to traditional classifica- tion algorithms, new method also outperform among all of them.

关 键 词：高斯混合模型 EM算法 遥感 模式识别 分类 

分 类 号：P237[天文地球—摄影测量与遥感]