基于空-谱先验条件随机场的高分辨率遥感影像变化检测方法

Change detection based on conditional random field model with spectral-spatial prior information for high spatial resolution remote sensing imagery

高空间分辨率遥感影像能够提供丰富的空间细节信息,使利用遥感影像进行精细变化检测成为可能.为充分挖掘高分辨率影像中的光谱、空间信息,本文提出一种基于影像空-谱先验信息的条件随机场(Conditional Random Field based on Spectral-Spatial Prior,SSPCRF)模型,该方法使用显著性检测方式自动提供先验光谱-空间样本信息,提高一元势能构建精度,有效缓解一元势能构建不准确导致的推理过程中的误差传递问题,并在二元势能中综合考虑标记场与观察影像的空间上下文信息以保持变化地物轮廓信息.最后,使用基于消息传递机制的推理方法将模型进行全局优化.在2组高分辨率影像数据集上的实验结果表明该方法能够提供较精确的初始变化检测信息,使得在减少变化检测结果中虚警点的同时保持变化地物细节信息.

In this paper,a conditional random field model based on spectral-spatial prior information（ SSPCRF） is proposed to perform the task of change detection for high spatial resolution remote sensing images. The proposed method firstly introduces a saliency based sample selection strategy which considers the spectral-spatial information of observed difference image to improve the accuracy of modeling initial change detection result.Then a pairwise potential with boundary constraint is used to help keep the boundary of changed objects. Finally an inference method based on loopy belief propagation（ LBP） algorithm is introduced to perform efficient optimization of the proposed model and get the final change map. The proposed SSPCRF model can greatly improve change detection accuracy while keeping detailed boundary information of changed objects. The proposed method is tested on two high resolution datasets and outperforms the commonly used change detection methods.