【摘要】 充分表达和利用目标空间上下文及语义信息是提高高空间分辨率影像分类精度的关键技术,而条件随机场(CRFs)在目标空间上下文建模以及分类预测方面有其独特优势。但是,基于单一尺度分析的CRFs模型存在不能反映目标多层次空间结构及语义关系的问题,因此针对城区高分辨率影像土地利用/覆盖分类问题,在面向对象分类框架下,提出了一种多级空间上下文LRCRFs模型。该模型定义如下:首先,将影像进行对象层、目标层及场景层的分层表达及分层特征提取,并进行"对象-目标-场景"的逐层关联;其次,采用逻辑回归(LR)分类器定义CRFs模型的关联势函数,利用分层特征加权的Potts函数定义交互势函数;采用最大-积消息传递算法对该模型进行近似推理。利用IKONOS多光谱影像及大比例尺真彩色航空影像进行试验的结果表明:多级空间上下文LR-CRFs模型分类精度高于单一尺度的基于像素层或对象层分割的LR-CRFs模型,其精度平均分别提高了4.63%和2.22%;该方法在一定意义上也缓解了面向对象分类方法中分类结果对分割尺度的依赖程度。更多还原

【Abstract】 Expressing and utilizing objective spatial context and semantic information adequately is a key technology to improve the classification precision of high spatial resolution image,but conditional random fields(CRFs)have the unique advantages for modeling objective spatial context and predicting the classification.However,CRFs model with single scale can not show the multi-level spatial structure of object and semantic relationship,so multi-level spatial context LR-CRFS model is proposed under the framework of object-oriented classification for the land use/cover classification of urban high resolution image.The definition of the model is that image is classified into object layer,target layer and scene layer,and features of each layer are extracted,and"object-target-scene"is layer by layer collected;second,correlation potential function of CRFs model is defined with logistic regression(LR)classifier,and interaction potential function is defined with Potts function weighted by hierarchical features;max-product message passing algorithm is used to infer the model approximately.The experiments on IKONOS multi-spectral image and large scale true color aerial image indicate that the classification precision of multi-level spatial context LR-CRFs model is averagely 4.63% and 2.22% higher than that of LR-CRFs model with single scale pixel layer or object layer, respectively;in a certain sense,the model lessens the dependent of classification result calculated by object-oriented classification method on segmentation scale.更多还原