关键词: Sentinel-1A, 哨兵数据, 合成孔径雷达, 城市建设用地, 全卷积网络, 深度学习

Abstract:

Timely and accurate extraction of urban construction land is essential for assessing the impacts of urban expansion on the environment.Optical remote sensing is susceptible to weather conditions and contains insufficient information on the texture of urban features,and is thus not conducive to urban construction land extraction.Synthetic aperture radar (SAR) is unique in its potential to be used in all weather conditions and at any time.Furthermore,its backscattering and polarimetric information is sensitive to the dielectric and geometric properties of the urban land surface,providing a new means for rapid and accurate extraction of urban construction land.Among existing SAR data,the Sentinel-1A SAR data is widely used due to its free access.Nonetheless,existing methods are poorly adapted to fully utilize the Sentinel1A SAR data,thereby limiting its application.The fully convolutional network is a deep learning network developed on the basis of the convolutional neural network,which adopts “pixel to pixel” image recognition and has become an effective method for extracting urban construction land.This paper aims to develop a method for detecting urban construction land based on Sentinel1A SAR data,and thus a fully convolutional network.Firstly,a fully convolutional network consisting of five Inception modules was developed with reference to GoogleNet.Each Inception module includes three convolutional layers with convolutional kernel sizes of 1×1,3×3,and 5×5 respectively,three corresponding activation layers,and a concatenation layer.The Ganzhou District in Zhangye City,Gansu Province,China,where urban construction land and bare land show similar spectrum features,was used as a case study area to verify our fully convolutional network.The extraction results show that the area of construction land in the Ganzhou District was 4 006 hectares in 2018,accounting for about 20% of the total area.The patch area of urban construction land was generally found to be between 0.01 and 1 hectare.Urban construction land was mainly distributed in the west and northeast,and the circle of 2-4 km from the city center exhibited the most concentrated urban construction land.Our accuracy assessment shows an overall accuracy of 92.50% and a Kappa coefficient of 0.85.By comparing our results to extraction results using the “KTH-Pavia” method (the most widely used method for extracting urban construction land from SAR data),it was found that extraction results based on the fully convolutional network are closer to the real urban construction land in spatial patterns.Furthermore,the overall accuracy and Kappa coefficient were,respectively,11% and 37% higher than the “KTH-Pavia” method.The principal reason for this higher accuracy is that the fully convolution network can better integrate multi-polarization and multi-scale texture information from SAR data.Furthermore,the fully convolutional network contains multiple convolutional structures and supports multi-source data inputs.The method developed in this study has both greater accuracy than existing methods and is applicable to urban construction land extraction based on different SAR data in different regions.It therefore has potential for widespread application.

Key words: Sentinel-1A, sentinel , data, synthetic aperture radar, urban construction land, fully convolutional network, deep learning