一种高分影像随机森林变化检测方法

A method of random forest change detection based on high resolution image

针对现有对象级变化检测方法在高分辨率影像上受影像配准误差的影响而表现不佳的现状,提出了一种顾及像元邻域的遥感影像随机森林变化检测新方法。该方法首先在像元尺度上提取前后两时相的光谱特征图和LBP纹理特征图,在此基础上考虑像元邻域关系计算对应的差分特征图;接着采用多尺度分割技术对两时相的叠合影像进行分割;最后利用随机森林算法模型实现变化检测,并以鹤地水库2014和2018年两期Spot卫星影像为数据源验证方法的有效性。结果表明:(1)综合考虑对象的同质性指数HI和对象的异质性指数MI的综合评价函数F能为最优分割尺度的选择提供客观依据;(2)考虑像元邻域建立匹配关系可以削弱由于不同时相影像间的配准误差所引起的像元误匹配风险,变化检测的精度随着邻域窗口的增大呈现出先升后降的特征;(3)本文方法的变化检测精度约为95%,优于现有的面向对象的变化矢量分析检测方法。

The existing object-level change detection method shows poor performance to high-resolution remote sensing images due to the registration error,to overcome this problem,this paper proposed a new method for change detection of high-resolution remote sensing images using random forest considering pixel neighborhood.The proposed method first extracted spectral feature and LBP texture feature at pixel scale and generated the corresponding feature maps for them.Then an multi-scale segmentation technique was used to segment the two-phase stacked images,and the difference feature maps were generated based on the feature maps and considering the spatial neighborhood relationships.Finally the random forest model was applied to obtain the final change detection results.A real experiment was used to validate the effectiveness of the proposed method by using two-phase images from 2014 to 2018 in Hedi reservoir area.The results indicated that:(1)A comprehensive evaluation function F that comprehensively considers homogeneity index HI and heterogeneity index MI for image objects can be effective in the selection of the optimal segmentation scale.(2)Matching strategy with pixel neighborhoods can weaken the region mismatch risk caused by registration error between different temporal images.The change detection accuracy shows a trend of rising first and then decreasing with the increase of the neighborhood window size.(3)The change detection accuracy of the proposed method can reach approximately 95%,and shows promotion when compared with the existing object-oriented change detection methods.