基于融合多模态的遥感影像冰川识别方法  

A Glacier Identification Method Based on Fused Multimodal Remote Sensing Images

作　　者：张昊 张秀再[1] 杨昌军[2,3,4] 许岱 ZHANG Hao;ZHANG Xiu-zai;YANG Chang-jun;XU Dai(School of Electronic and Information Engineering,Nanjing University of Information Science&Technology,Nanjing 210044,China;National Satellite Meteorological Center,Beijing 100081,China;Xu Jianmin Meteorological Satellite Innovation Center,Beijing 100081,China;China Meteorological Administration,China Key Laboratory of Remote Sensing Satellite Radiometric Measurement and Calibration,Beijing 100081,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,江苏南京210044 [2]国家卫星气象中心,北京100081 [3]许健民气象卫星创新中心,北京100081 [4]中国气象局遥感卫星辐射测量和定标重点开放实验室,北京100081

出　　处：《中国电子科学研究院学报》2024年第5期419-431,共13页Journal of China Academy of Electronics and Information Technology

基　　金：国家社会科学基金一般项目(22BZZ080);第二次青藏高原综合科学考察研究项目(2019QZKK0105)。

摘　　要：冰川对气候变化极为敏感,冰川变化与区域生态、自然灾害、水资源等息息相关。高原冰川遥感信息提取及实时监测是监测冰川变化不可或缺的手段。为有效识别多尺度高分辨率遥感影像中的冰川,设计一种Glacier-Unet模型。(1)针对现有的基于Landsat卫星遥感影像高原冰川提取算法因缺乏应对复杂地物干扰影响的有效方法,导致反射目标信息丢失的问题。以青藏高原阿尼玛卿雪山为试验对象,选取基于Landsat-9遥感卫星高分辨率影像制作数据集。对高分辨率冰川遥感影像进行数据预处理,采取特征级融合和像素级融合制作多模态遥感数据影像,通过滑动切片、数据增强手段丰富语义分割数据集,保证模型训练准确性和鲁棒性;(2)针对零散、细小冰川识别能力不足的问题,设计门控多尺度过滤层(Gated Multi-scale Filter Layer,G-MsFL)滤除无用特征信息,使模型具备多尺度特征提取和特征融合能力,有效识别复杂地物环境中的冰川;(3)针对冰川轮廓模糊问题,设计并联双通道注意力模块(Paralleling Dual Attention Module,P-DAM)。将冰川边界丰富的上下文信息进行编码作为特征图的局部特征,从而增强其特征表达能力。对改进的Glacier-Unet模型在阿尼玛卿测试数据集中的实验结果进行定性、定量分析,发现整体分割精度较对比方法提升6.1%,且能有效识别零散、细小冰川,对高原地区冰川识别工作具有重要意义。Glaciers are extremely sensitive to climate change,and changes in glaciers are closely related to regional ecology,natural disasters and water resources.Real-time monitoring and information extraction from remote sensing of highland glaciers is an indispensable means to monitor glacier changes.In order to effectively identify glaciers in multi-scale high-resolution remote sensing images,a Glacier-Unet model is designed.The specific work is(1)to address the existing Landsat satellite remote sensing image based plateau glacier extraction algorithms due to the lack of response to the impact of complex feature in-terference,resulting in the loss of reflective target information.Taking the Tibetan Plateau Animaqing Snow Mountain as a test object,a dataset based on Landsat-9 remote sensing satellite high-resolution im-age production was selected.Data preprocessing is carried out on high-resolution glacier remote sensing images,feature-level fusion and pixel-level fusion are adopted to produce multimodal remote sensing data images,and semantic segmentation datasets are enriched by sliding slices and data enhancement means to ensure the accuracy and robustness of the model training;(2)Aiming at the insufficient recognition abil-ity of scattered and tiny glaciers,a gated multi-scale filter layer(G-MsFL)is designed.filter layer(G-MsFL)is designed to filter out useless feature information.The model is equipped with multi-scale fea-ture extraction and feature fusion capabilities,which can effectively identify glaciers in complex feature environments;(3)Paralleling Dual Attention Module(P-DAM)is designed to address the problem of fuzzy glacier contours.The rich contextual information of the glacier boundary is encoded as the local fea-tures of the feature map,thus enhancing its feature expression ability.Qualitative and quantitative analy-ses of the experimental results of the improved Glacier-Unet model in the test dataset reveal that the over-all segmentation accuracy is improved by 6.1%compared with the comparison method,a

关 键 词：高分辨率遥感影像 多模态融合 注意力机制 多尺度特征提取 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]