遥感图像数字表面模型建筑物点过滤方法研究

Research on building point filtering method of remote sensing image digital surface model

针对图像密集匹配生产的数字表面模型(DSM)进行点云滤波,算法对地形依赖大,参数设置复杂,精度不高,后续人工编辑修饰的工作量大、效率低的问题,该文设计了第一套针对DSM滤波、涵盖多种样本形式(栅格、矢量)的航空图像建筑物数据集。针对航空图像建筑物尺度较大等特点,将膨胀卷积加入U-Net构成Dilated U-Net,并综合运用其进行建筑物语义分割,利用分割结果在相应图像密集匹配得到的DSM上滤除建筑物点,然后采用投票插值策略得到过滤掉建筑物点的DSM。实验证明:利用该文网络DU-Net将DSM中非地面建筑物点滤除,Ⅰ类误差在5.8%以内,Ⅱ类误差在2.4%以内,其可以在30 s内完成超过9000万个建筑点与非建筑物点位置的预测,效率高、成本低。DU-Net网络建筑物语义分割过程不受地形、高差的限制,对于其他非地面点的滤波具有一定的借鉴意义。

Aiming at the problem that the digital surface model(DSM)produced by image dense matching is subjected to point cloud filtering,and the algorithm relies heavily on terrain,the parameter setting is complex,the precision is not high,and the subsequent manual editing and modification work is heavy and inefficient,in this paper,we designed the first set of aerial image building data sets for DSM filtering,covering a variety of sample forms(raster,vector).In view of the large scale of aerial image buildings,this paper added dilated convolution to U-Net to form dilated U-Net network.According to the characteristics of large scale of aerial image buildings,dilated convolution was added to U-Net to form Dilated U-Net.In this paper,we used DU-Net to segment buildings semantically,filter out building points on DSM obtained by dense matching of corresponding images using segmentation results,and then used voting interpolation strategy to obtain DSM filtered out building points.Experimental results showed:using the convolution neural network DU-Net would find not the ground buildings point in DSM,Ⅰ class error within 5.8%,Ⅱ class error within 2.4%.This method could predict the location of more than 90 million building points and non-building points in 30 seconds,which was efficient and lowcost.Moreover,the building segmentation process was not limited by the terrain,which had certain reference significance for the filtering of other nonground points.