基于Faster R-CNN和MorphACWE模型的SAR图像高原湖泊提取  

The extraction of plateau lakes from SAR images based on Faster R-CNN and MorphACWE model

作　　者：董天成 杨肖 李卉[2] 张志 齐睿 DONG Tiancheng;YANG Xiao;LI Hui;ZHANG Zhi;QI Rui(Institute of Geophysics & Geomatics, China University of Geosciences, Wuhan 430074, China;School of Earth Science, China University of Geosciences, Wuhan 430074, China;32023 Troops, Dalian 116032, China)

机构地区：[1]中国地质大学(武汉)地球物理与空间信息学院,武汉430074 [2]中国地质大学(武汉)地球科学学院,武汉430074 [3]32023部队,大连116023

出　　处：《国土资源遥感》2021年第1期129-137,共9页Remote Sensing for Land & Resources

基　　金：中国地质调查局项目“全国矿山环境恢复治理状况遥感地质调查与监测”(编号:DD20190705);国家自然科学基金项目“环境示踪剂辅助的高寒山区融雪径流过程模拟研究”(编号:41401076);青海省青藏高原北部地质过程与矿产资源重点实验室开放课题(编号:2019-KZ-01)共同资助。

摘　　要：青藏高原湖泊是高原生态环境中最重要的自然要素之一,实现青藏高原湖泊调查与监测是现阶段迫在眉睫的任务。由于水体在SAR图像上呈现出独特的镜面反射特征,使得利用SAR图像进行湖泊的提取与分析成为当下研究热点。为进一步排除干扰地物影响、提高分类准确度,采用欧空局Sentinel-1A干涉宽幅模式的斜距单视复数产品(SLC)为主要数据源,Sentinel-2A多光谱影像Level-1C产品作为参考数据源,提出一种结合改进Faster R-CNN和MorphACWE轮廓模型的SAR图像湖泊提取算法(Faster Region-based Convolution Neural Network-MorphACWE,FR-MorphACWE)。该算法结合深度学习目标检测算法的高维特征分析和MorphACWE模型的边界提取,从综合干扰多湖泊提取角度进行分类实验评价,充分利用高原湖泊的形态学和雷达反射特征,实现西藏自治区那曲市南部至日喀则市北部高原湖泊高精度提取。实验结果表明,该算法在综合干扰多湖泊情境下准确率可达99.71%,精准率和召回率均高于98%,可作为SAR图像高原湖泊提取的新算法加以推广和应用。Lakes in the Tibetan Plateau constitute one of the most important natural factors in the plateau ecological environment.So,it is an urgent task to investigate and monitor lakes in the Tibetan Plateau.Because of the unique backscatter characteristics of water body in the image,the extraction and analysis of the lake based on SAR image has become a research hotspot.In order to further eliminate the interference of surface features and improve the classification accuracy,this paper proposes a high-precision lake extraction FR-MorphACWE(Faster Region-based Convolution Neural Network-MorphACWE)model of SAR image.The Interferometric Wide Swath(IW SLC)of the European Space Agency's sentinel-1A interference wide-band mode is used as the main data source,and the sentinel-2a multispectral image level-1c product is used as the reference data source.This model combines the morphological analysis advantages of Faster R-CNN target detection algorithm and the contour extraction advantages of MorphACWE model.The classification experiments were carried out from extraction of comprehensive interference multi-lake.The target detection algorithm was applied to eliminate non-lake surface disturbance.On such a basis,the active contour model was used to extract the lake boundary,and the morphological characteristics and radar reflection characteristics of plateau lakes were fully utilized to achieve high-precision extraction of plateau lakes from the south of Naqu City to the north of Xigaze City in Tibet.The experimental results show that the accuracy of the algorithm can reach 99.71%and the accuracy and recall rate are higher than 98%in the situation of multi-lake interference.

关 键 词：目标检测 Faster R-CNN CV轮廓模型 合成孔径雷达 高原湖泊提取 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]