融合空间关系的遥感图像分类

Classification of remote sensing images based on the fusion of spatial relationship

针对光谱纹理特征分类方法的不足,提出了一种融合空间关系的遥感图像分类方法。利用直方图提取像斑特征,采用G统计量构建单像斑概率,通过迭代统计方法计算地物类别邻接概率,利用地物类别邻接概率表达像斑邻域概率,加权组合单像斑概率与像斑邻域概率构建像斑联合概率,依据最大后验概率准则获取图像分类结果。在QucikBird图像上的试验结果表明:与传统的光谱纹理分类方法相比,该方法能够提高图像分类的精度;整体分类精度与Kappa系数分别提高了1.5%和2.1%。

In order to overcome the disadvantages of the classification method based on spectral and texture features,the authors put forward a classification method based on the fusion of spatial relationship in this paper.Single object probability was built by G statistic after image object feature was extracted by histogram. The neighborhood object probability was described by land cover adjacency probability which was calculated by iterative statistics method. The joint probability of the object was built by the weighted combination of single object probability and neighborhood object probability. The classification result of the image was obtained according to the maximum a posteriori. The experimental results based on Quick Bird image show that the proposed method can improve the classification accuracy compared with the traditional classifier using spectral and texture features. The overall classification accuracy and kappa coefficient are increased by 1. 5% and 2. 1%,respectively.