摘要
由于山体坡度、光照角度、传感器成像角度等因素,遥感图像中的山体阴影影响了冰川识别的精度.现有方法一般是先去除阴影再进行冰川识别,既繁琐又可能破坏图像的光谱信息.本文在U-Net框架中集成金字塔池化模块以增强多尺度特征提取能力,提出了一种U-PSP-Net结构的卷积神经网络,可以实现阴影区冰川识别.在自制的含阴影冰川数据集上进行验证,与PSP-Net、SegNet和U-Net的性能比较表明,提出的U-PSP-Net的平均像素精度为95.84%,平均交并比(IoU)为92.79%.与U-Net相比,分别提升了0.61%和0.92%;与PSP-Net和SegNet相比分别提高了1.41%、2.54%和2.85%、2.86%.以上结果证明了神经网络结构在含阴影遥感影像中识别冰川的可行性和有效性.
Mountain shadows, resulted from certain mountain slope, illumination angle, and sensing angle, notably affect the accuracy of glacier identification in remote sensing image analysis.Generally, existing methods remove the shadows first and then perform glacier identification, which is cumbersome and alters the spectral information inside the shadows.By integrating a pyramid pooling module with U-Net to enhance feature extraction ability across scales, this paper proposed a U-PSP-Net convolutional neural network to simplify the process of glacier identification with shadows.The proposed network was validated with a self-made glacier dataset of highly shadowed regions.The average pixel accuracy of the proposed U-PSP-Net is 95.84%,and the mean intersection over union(IoU) is 92.79%.The improvements are 0.61% and 0.92%,respectively, compared with U-Net, 1.41% and 2.54% over PSP-Net, and 2.85% and 2.86% over SegNet.The results prove the feasibility and effectiveness of the proposed network in glacier identification with shadows.
作者
张大奇
范慧颖
康宝生
高健
李铁键
ZHANG Daqi;FAN Huiying;KANG Baosheng;GAO Jian;LI Tiejian(School of Water Resources and Electric Power,Qinghai University,Xining 810016,China;School of Information Science and Technology,Northwest University,Xi'an 710127,China;Department of Hydraulic Engineering,Tsinghua University,Beijing 100084,China)
出处
《应用基础与工程科学学报》
EI
CSCD
北大核心
2022年第4期806-818,共13页
Journal of Basic Science and Engineering
基金
国家自然科学基金项目(51769027)
青海省重大科技专项(2021-SF-A6)。
作者简介
张大奇(1976-),男,博士,副研究员.E-mail:dqzhang2005@aliyun.com;通信作者:范慧颖(1996-),女,硕士研究生.E-mail:1395632010@qq.com。
