序列卷积神经网络支持下线状地图目标的分段方法  

Segmentation of linear map objects using sequential convolutional neural network

作　　者：杨敏[1] 陈果 李连营[1] 黄浩然 苗静 晏雄锋 YANG Min;CHEN Guo;LI Lianying;HUANG Haoran;MIAO Jing;YAN Xiongfeng(School of Resource and Environmental Sciences,Wuhan University,Wuhan 430079,China;College of Surveying and Geo-Informatics,Tongji University,Shanghai 200092,China;Wuhan Geomatics Institute,Wuhan 430022,China)

机构地区：[1]武汉大学资源与环境科学学院,湖北武汉430079 [2]同济大学测绘与地理信息学院,上海200092 [3]武汉市测绘研究院,湖北武汉430022

出　　处：《测绘学报》2023年第1期108-116,共9页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42071450,42001415);自然资源部数字制图与国土信息应用重点实验室开放研究基金(ZRZYBWD202101)。

摘　　要：依据形态特征差异实施分段处理是实现线状地图目标自适应综合的重要步骤。传统方法主要采用长度、角度、曲率等参量描述局部曲线段的几何形态变化,通过人工设置规则或常规机器学习方法建立模式判别模型。本文提出基于卷积神经网络的线状目标模式识别与分段方法。首先,以相邻坐标点构成的线元为基本单元,以线元端点横、纵坐标差为特征,将线状目标离散化为二维序列;然后,建立序列卷积神经网络进行线元特征序列学习与预测,实现线元层次的模式类型判别;最后,利用迭代融合方法将拓扑相邻且模式类型相同的线元合并,从而输出不同形态模式的分段结果。以1∶5万行政区界线和1∶25万山区道路数据开展试验,本文方法分段结果与人工分段结果的一致性比率分别达到91.25%和85.65%,相较传统方法有一定提升。同时,本文方法通过深度学习获取模式判别的深层次特征,能够有效避免人工选择特征带来的主观性影响,对不同尺度、不同类型的线状目标分段问题适应性更好。Segmentation of linear objects based on their morphological characteristics is a pivotal pre-step for adaptive generalization. Existing studies mainly use hand-crafted features, such as length, angle, and curvature, to describe the local structures of linear objects, and further to identify different patterns based on manual-defined rules or machine learning methods. In this study, we propose a structural recognition and segmentation method for linear objects using deep learning. First, a linear unit(also known as lixel) composed of two adjacent points is considered as the processing unit, and each linear object is discretized into a two-dimension sequence in which the differences between the horizontal and vertical coordinates of each lixel are encoded. Then, a sequential convolutional neural network(SCNN) is established to predict the types of each lixel. Finally, the segmentation results of different morphological characteristics are obtained by merging the adjacent lixels with the same type using an iteration method. Experiments were conducted on two datasets of 1∶50 k administrative boundaries and 1∶250 k mountain roads, and the consistency ratios of segmentation results were 91.25% and 85.65%, respectively, outperforming the traditional methods based on backpropagation artificial neural network and Na6ve Bayes. Overall, our method can effectively avoid the subjectivity that exists when designing the hand-crafted features, and is more adaptable to the segmentation of linear objects with different scales and types.

关 键 词：线状地图目标 分段 序列卷积神经网络 深度学习 

分 类 号：P208[天文地球—地图制图学与地理信息工程]