极地研究

极地研究 ›› 2025, Vol. 37 ›› Issue (3): 464-476.DOI: 10.13679/j.jdyj.20240095

• 研究论文 • 上一篇    下一篇

基于多源遥感和深度学习的南极拉斯曼丘陵小型湖泊水深反演

朱婷婷1,李加程1,崔祥斌2,束禅方1,张宇3   

  1. 1南京工业大学测绘科学与技术学院, 江苏 南京 211816; 
    2中国极地研究中心(中国极地研究所), 上海 200136; 
    3武汉大学中国南极测绘研究中心, 湖北 武汉 43007

  • 收稿日期:2024-11-07 修回日期:2024-11-29 出版日期:2025-09-30 发布日期:2025-09-25
  • 通讯作者: 崔祥斌
  • 基金资助:
    国家自然科学基金、“高分”国家重大科技专项项目、自然资源部极地科学重点实验室开放基金和江苏省科技项目

Bathymetric inversion of small lakes in the Larsemann Hills, Antarctica, based on multi-source remote sensing and deep learning

ZHU Tingting1, LI Jiacheng1, CUI Xiangbin2, SHU Chanfang1, ZHANG Yu3   

  1. 1College of Geomatics Science and Technology, Nanjing Tech University, Nanjing 211816, China;
    2Polar Research Institute of China, Shanghai 200136, China;
    3Chinese Antarctic Center of Surveying and Mapping, Wuhan University, Wuhan 430079, China
  • Received:2024-11-07 Revised:2024-11-29 Online:2025-09-30 Published:2025-09-25

PDF (PC)

26

摘要： 由于降水和冰雪消融在南极冰盖表面及裸岩区形成的湖泊, 其时空分布和水储量的变化影响着南极冰盖边缘冰川和冰架的稳定性, 是全球气候变化的重要指标。本文结合ICESat-2激光测高、Sentinel-2多光谱影像和航空影像等多源遥感数据, 提出多源异质遥感影像融合反演水深框架。利用ICESat-2激光测高数据提取沿测线水深, 并结合光学数据线性经验模型和卷积神经网络(convolutional neural network, CNN)深度学习方法建立湖泊光谱信息与湖泊水深反演模型, 估算拉斯曼丘陵小型湖泊水深, 最后利用中国第39次南极考察获取的高分辨率航空遥感影像数据和俄罗斯第63次南极夏季考察的实测原位水深数据对反演结果进行精度验证。结果表明, 拉斯曼丘陵Reid湖采用经验模型得到的水深与原位水深的均方根误差(Root Mean Square Error, ERMSE)为0.58 m、平均绝对误差(Mean Absolute Error, EMAE)为0.49 m、平均偏差(Average Bias, BA)为−0.36 m、偏差标准差(Bias Standard Deviation, DBSD)为0.46 m、相关系数(R2)为0.51; CNN深度学习算法得到的水深与实测水深的ERMSE为0.37 m、EMAE为0.32 m、BA为−0.19 m、DBSD为0.32 m、R2为0.75, 从4个指标上均体现出基于深度学习算法能够显著提升水深反演精度。结果表明, 采用ICESat-2激光测高和多光谱影像等多源异质遥感影像融合可以实现高精度水深反演, 其中, 基于CNN模型反演的南极小型湖泊深度比线性经验模型误差更小、精度更高。

关键词: 多源异质遥感, ICESat-2, Sentinel-2, 卷积神经网络(CNN), 水深

Abstract: The spatial-temporal distribution and changes in water storage of lakes formed on the surface of the Antarctic Ice Sheet (AIS) and in exposed bedrock due to precipitation and melting snow and ice are important indicators of global climate change, as they affect the stability of glaciers and ice shelves at the margins of the AIS. In this paper, multi-source remote sensing data, such as ICESat-2 laser altimetry, Sentinel-2 multispectral imagery, and aerial imagery, are fused to derive water depths in the proposed framework. ICESat-2 laser altimetry data are employed to calculate water depth along the survey line. Multispectral information derived from optical data is input into the linear empirical model and the deep learning convolutional neural network (CNN) models to establish the corresponding relationships for small lakes in the Larsemann Hills. The derived water depths are verified by combining high-resolution airborne remote sensing image data acquired by helicopters on 39th Chinese National Antarctic Research Expedition and measured bathymetric data from the 63rd Russian Antarctic Summer Expedition. The experimental results show that the root mean square error (ERMSE) of the image inversion bathymetry and in situ bathymetry at Lake Reid derived using the empirical model is 0.58 m, the mean absolute error (EMAE) is 0.49 m, the average bias (BA) is −0.36 m, the bias standard deviation (DBSD) is 0.46 m, and R2 is 0.51. The accuracy evaluation results obtained by the CNN deep learning algorithm are 0.37 m for ERMSE, 0.32 m for EMAE, –0.19 m for BA, and 0.32 m for DBSD, with an R2 value of 0.75, which indicates that the deep learning-based algorithm can realize significant improvement of the bathymetric inversion accuracy from the four accuracies. Therefore, this paper utilizes ICESat-2 laser altimetry and multispectral remote sensing images to construct a multi-source heterogeneous remote sensing image fusion model, on the basis of which a deep learning CNN model is developed to realize the Antarctic small lake bathymetry task, with lower error and higher accuracy compared with the linear empirical model.

Key words: multi-source heterogeneous remote sensing, ICESat-2, Sentinel-2, CNN, water depth