结合纹理特征的SVM样本分层土地覆盖分类

SVM Land Cover Classification based on Spectral and Textural Features Using Stratified Samples

支持向量机(SVM)分类在精度、泛化性、高维数据处理等方面都具有较强的优势,在遥感影像分类中也得到了广泛应用。由于遥感影像"同物异谱"和"异物同谱"现象的影响,结合纹理特征提高SVM分类精度已成为遥感应用研究的热点。但不同尺度的纹理特征突出的信息不一,在同一尺度上难以区分的地物在多尺度空间则更容易区分,因此,采用多尺度纹理特征进行SVM分类,并从分类样本和纹理特征的选取两个方面探讨SVM土地覆盖分类的方法。首先,以ALOS影像为例,通过灰度共生矩阵提取不同尺度、不同方向的几种纹理特征;然后在光谱分类结果基础上,借助地类特征曲线,选取有效的多尺度纹理特征,最后进行样本分层分类。样本分层分类是选取首层样本进行分类,再从"漏分和错分"地块中选取新样本加入到首层样本中,得到第二层样本并对整个影像进行分类;用同样的方法选出第三层样本或更高层样本进行分类,直到结果满意为止。结果表明:该方法比仅用光谱特征的SVM分类总精度提高了8.11%,Kappa系数增加了0.11。其中,纹理特征的引入使分类总精度提高了4.13%,且对纹理特征较明显的地类更有效;采用样本分层后的分类总精度进一步提高了3.98%,且各单一地类的精度也都有不同程度的提高。借助地类特征曲线选择合适的纹理特征具有一定的可行性,并且采用样本分层的方法能够提高SVM分类的精度。

Support vector machine(SVM)shows great performance in many classification algorithms,with the merits of high precision,generalization ability and high-dimensional data processing ability.Therefore, It has been widely used in remote sensing classifications.SVM classification,combining with texture features,has been the research focus of remote sensing applications.Since texture features can overcome the phenomena of'the same thing with different spectrum and different things with the same spectrum'in remote sensing images.Multi-scale texture features were used to distinguish features in different scales space,which were difficult to distinguish in single scale.The study was mainly focus on texture features selection and classification with stratified samples.Firstly,using ALOS pan and multispectral remote sensing images,8kinds of texture features in different scales and directions were extracted,based on the Gray Level Co-occurrence Matrix;Secondly,with the help of the characteristic curve of land types,texture features of mean,homogeneity and dissimilarity in multi-scale were selected,based on the spectral classification results.Finally,the sample stratification method was used in the SVM classification of land cover,which combined spectral with these three kinds of multi-scale texture features in different directions.The sample stratification was implemented as follows:firstly,select the initial training samples to classify the land types;secondly,select new samples from the misclassified plots of the initial classification results,then put these new samples and the initial training samples together constituted the second-level training samples.If the second-level training samples met the needs of classification,they were the final training samples.If not,using the same method to select higher-level training samples.The experimental results show that the overall accuracy and Kappa coefficient of the SVM classification,which combined spectral with multi-scale texture features,using the third-level training samples,which are improved by 8.11%and 0.11respectively,compared with that based only on spectral features,using the initial level training samples.Texture features made overall classification accuracy improved 4.13%,especially for those land cover types whose texture features is strong;while stratified sample contributed to 3.98%,and the accuracy of single classification has different degrees of improvement.Results illustrates the method used in this paper is effective.