基于瓦片重叠法的在线高分遥感图像目标智能提取方法研究

Research on Intelligent Extraction Method of Spatial Objects from Online High-Resolution Remote Sensing Images based on Tile-overlapping Strategy

在线遥感影像上目标智能提取是地表机器智能感知的重要途径。然而,如何利用对单个图像大小限制的Mask R-CNN模型在线遥感影像上指定空间范围内有效提取空间目标,是亟待解决的关键问题。为此,提出了一种瓦片重叠法以实现在线高分遥感影像上大范围目标进行智能提取。首先,依据高分遥感影像上空间目标大小,设定检测单元的宽度和高度均为512个像素,并按一定的步长将相邻瓦片拼接成一系列相互重叠的检测单元;然后,利用Mask R-CNN提取局部范围的空间目标,并将其转换为空间坐标;最后,通过空间重叠融合相邻检测单元的重复目标。实验发现,当重叠率为50%时,能以最优的性能达到较高的提取精度。以在线地图上的运动场、城市路口、跨河大桥为例,开展了中国部分城市的在线高分遥感影像目标提取实验,平均目标提取率为80%左右,为地表机器智能感知提供了新途径。

The intelligent extraction of spatial objects from online remote sensing images is an important approach for the machine intelligence perception of earth surface.However,how to use the Mask R-CNN model which limits the size of a single image to effectively extract the spatial objects in the specified spatial range on the online remote sensing images is a key problem to be solved.In this paper,an extraction method of spatial objects based on tile-overlapping strategy is proposed to realize the intelligent extraction of spatial objects from online high-resolution remote sensing images.Firstly,according to the size of the spatial objects on the high-resolution remote sensing images,the width and height of the detection unit are set to 512 pixels,and the adjacent tiles are spliced into a series of overlapping detection units according to a certain overlap rate.Then,Mask R-CNN is used to extract the spatial objects from each of the detection units.The local coordinates of each extracted spatial object are converted to spatial coordinates through the coordinates and spatial resolution of the upper left corner of the prediction box.Finally,the repeated objects of adjacent detection units are fused through spatial overlap analysis to obtain the final spatial object sequence.Taking part of Chengdu urban area(with an area of 33.4 km~2)as the experimental area,set different overlap rates for spatial object extraction,and evaluate the performance and accuracy of different tile overlap rates.The experimental result indicated that the higher the overlap rate,the more opportunities there are to repeatedly identify the same object from different"fragments"at the same location,so as to reduce the leakage.On the other hand,repeated identification of the same location also increases the number of false identification.When the overlap rate is 50%,it can achieve high extraction accuracy with the best performance.Taking the playground,urban intersection and river crossing bridge on the online map as examples,the spatial objects of some cities in China were extracted from online remote sensing images with the the overlap rate set to50%.The results show that the average object extraction rate is about 80%,which provides a new way for the intelligent perception of earth surface.