基于深度学习的Sentinel-1A影像冰川识别  

Glacier Recognition from Sentinel-1A Satellite Image Based on Deep Learning

作　　者：王梓霏 柯长青[1] WANG Zifei;KE Changqing(School of Geography and Ocean Science,Nanjing University,Nanjing 210023,China)

机构地区：[1]南京大学地理与海洋科学学院,南京210023

出　　处：《遥感信息》2022年第4期43-50,共8页Remote Sensing Information

基　　金：国家自然科学基金项目(41830105)。

摘　　要：冰川监测对于气候变化研究及区域可持续发展有重要意义,利用遥感影像提取冰川边界是冰川监测的关键。利用Sentinel-1A结合地形数据,通过基于VGG16、MobileNetV2的UNet和DeepLabV3+卷积神经网络对喀喇昆仑地区的冰川进行识别,并比较VH极化和VV极化下的识别精度。结果表明,VH极化的识别精度整体高于VV极化。基于MobileNetV2网络的识别精度不如VGG16高,但实现了精度相当的同时提高了运行效率。基于相同的主干网络,DeepLabV3+较UNet网络识别精度高,即基于VGG16的DeepLabV3+网络精度最高,在VH极化下其识别总体精度可达95.18%,交并比IoU可达84.33%,均交并比mIoU达到88.91%。卷积神经网络对纯净冰川、表碛冰川及冰川湖都有较好的识别效果,且识别出部分前进冰川,为大区域山地冰川的快速且半自动化识别提供了技术基础。Glacier monitoring is of great significance to climate change research and regional sustainable development.Extracting glacier boundary from remote sensing images is essential to glacier monitoring.UNet and DeepLabV3+convolutional neural networks based on VGG16,MobileNetV2 are utilized to identify glaciers separately in the Karakoram region by using Sentinel-1A data combined with terrain data,and the recognition accuracy of VH and VV polarization is compared.The results show that the recognition accuracy of VH polarization is higher than that of VV polarization.The accuracy of glacier recognition based on the MobileNetV2 is close to that of VGG16,but MobileNetV2 can improve operating efficiency.DeepLabV3+has higher recognition accuracy than UNet based on the same backbone.That is,DeepLabV3+network based on VGG16 has the highest accuracy,which can reach an overall recognition accuracy of 95.18%with VH polarization mode,meanwhile,the intersection over union can reach 84.33%,and the mean intersection over union can reach 88.91%.Convolutional neural network has a good recognition effect on pure glaciers,debris-covered glaciers and glacial lakes,and can recognize advancing glaciers,which provides a technical basis for rapid and semi-automatic identification of mountain glaciers in a large area.

关 键 词：Sentinel-1A 冰川 深度学习 喀喇昆仑 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]