针对“基于像素的条件随机场( conditional random fields,CRFs)模型能否在m级分辨率的多光谱遥感图像分类中表现良好”的问题,提出了集成图像的光谱、方向梯度直方图和多尺度多方向Texton纹理等多种线索的CRFs模型定义方法。利用上述特征,选择随机森林( random forests,RF)定义CRFs关联势函数；利用特征对比度加权的Potts函数定义CRFs交互势函数,并且建立了多标签的RF-CRFs模型；对该模型进行分项参数训练以及基于图割的α-膨胀算法推理；利用典型城区的QuickBird多光谱图像进行模型的验证与精度评价。结果表明RF-CRFs模型的分类精度可达82.52%以上,比RF分类器的分类精度提高了3.35%。
The classification accuracy of superpixel - based conditional random fields ( CRFs ) model greatly depends on segmentation scale parameters, which constitutes a problem that should be solved. Therefore, to answer the question “whether a pixel-based CRFs model performs well in HSR image classification with m level spatial resolution or not”,the authors proposed a pixel-based CRFs model with the association term defined as an output of random forests classifier and the interaction potential defined as Potts function weighted by contrast function, and the definition of association and interaction terms adopted multi-cue features such as histogram of gradient, multi-scale and multi-direction Texton filter and multi-spectral information from HSR imagery. Finally, the proposed model was trained using piecewise training method and inferred using α-expansion algorithm based on graph cut. Experiments on a typical urban scene from QuickBird multi-spectral satellite imagery have shown that the proposed RF-CRFs model shows the classification accuracy of over 82. 52%. In addition, the classification accuracy of the model is higher than that of the RF classifier by 3. 35% on average.