PSPNet和形态学运算相结合的夜光遥感影像城市建成区提取方法  

A Method of Extracting Urban Built-up Areas From Nightlight Remote Sensing Images by Combining PSP Net and Morphological Operations

作　　者：王家琪 龚志辉[1] 郭海涛[1] 卢俊[1] 余东行 林雨准 WANG Jiaqi;GONG Zhihui;GUO Haitao;LU Jun;YU Donghang;LIN Yuzhun(PLA Strategic Support Force Information Engineering University,Institute of Geospatial Information,Zhengzhou 450001,China)

机构地区：[1]信息工程大学地理空间信息学院,河南郑州450001

出　　处：《测绘与空间地理信息》2022年第9期73-76,85,共5页Geomatics & Spatial Information Technology

基　　金：国家自然科学基金(41671410)资助。

摘　　要：城市建成区的提取对城市发展及其他经济建设具有重要作用。针对传统提取方法需要人工设置阈值、实用性差等问题,提出了一种结合PSPNet和形态学运算的夜光遥感影像城市建成区提取方法。该方法选用PSPNet作为原始建成区提取的算法模型,采用基于深度学习的语义分割技术对目标城市进行预测获得原始建成区,针对灯光溢出效应,提出Boundary Restrain结构元对原始提取结果进行形态学运算,最终获得更为精确的城市建成区。利用郑州、西安、济南等地区的珞珈一号夜光遥感影像进行了实验,该方法的平均精度为99.45%,Kappa系数为0.988。结果表明,本研究提出的PSPNet与形态学运算相结合的城市建成区提取方法,具有很高的提取精度和较好的泛化能力。The extraction of urban built-up areas plays an essential role in urban development as well as economic construction.In this study,a method of extracting urban built-up areas from nightlight remote sensing images combining PSPNet and morphological operations is proposed to address the problems of traditional extraction methods requiring manual setting of threshold values and poor practicality.Firstly,PSPNet is chosen as the algorithm model for the extraction of the original built-up area,and the semantic segmentation technique based on deep learning is applied to predicting the target city to obtain the original built-up area.Secondly,we proposed to use Boundary Restrain structural elements to perform morphological operations on the original extraction results for the light spillover effect.Finally,a more accurate built-up area of the city is obtained.Experiments were conducted using the nightlight remote sensing images of Luojia No.1 in Zhengzhou,Xi′an and Ji′nan.The results show that the average accuracy of the method is 99.45%and the kappa coefficient reaches 0.988.The proposed method of urban built-up area extraction by combining PSPNet and morphological operations with high extraction accuracy and excellent generalization ability.

关 键 词：深度学习 夜光遥感影像 城市建成区 语义分割 

分 类 号：P237[天文地球—摄影测量与遥感]