X img





作者投稿 专家审稿 编辑办公 编委办公 主编办公

冰川冻土 ›› 2022, Vol. 44 ›› Issue (5): 1419-1428.doi: 10.7522/j.issn.1000-0240.2022.0128

• 第二次青藏高原综合科学考察研究 • 上一篇    


王习敏1,2(), 黄荣刚1, 徐志达1,2, 焦志平1,2, 江利明1,2()   

  1. 1.中国科学院 精密测量科学与技术创新研究院 大地测量与地球动力学国家重点实验室, 湖北 武汉 430077
    2.中国科学院大学 地球与行星科学学院, 北京 100049
  • 收稿日期:2022-05-01 修回日期:2022-06-01 出版日期:2022-10-25 发布日期:2022-11-05
  • 通讯作者: 江利明 E-mail:wxm_2833740760@apm.ac.cn;jlm@apm.ac.cn
  • 作者简介:王习敏,硕士生,主要从事冰缘地貌遥感智能解译研究. E-mail: wxm_2833740760@apm.ac.cn
  • 基金资助:

Extraction of solifluction terraces in the eastern Qinghai-Tibet Plateau based on deep learning and high-resolution remote sensing images

Ximin WANG1,2(), Ronggang HUANG1, Zhida XU1,2, Zhiping JIAO1,2, Liming JIANG1,2()   

  1. 1.State Key Laboratory of Geodesy and Earth’s Dynamics,Innovation Academy for Precision Measurement Science and Technology,Chinese Academy of Sciences,Wuhan 430077,China
    2.College of Earth and Planetary Sciences,University of Chinese Academy of Sciences,Beijing 100049,China
  • Received:2022-05-01 Revised:2022-06-01 Online:2022-10-25 Published:2022-11-05
  • Contact: Liming JIANG E-mail:wxm_2833740760@apm.ac.cn;jlm@apm.ac.cn


融冻泥流阶地是一种典型的斜坡冷生型冰缘地貌,在青藏高原东部发现大量古融冻泥流阶地,其空间分布对于重建该区古多年冻土分布和古气候环境具有重要意义。融冻泥流阶地纹理复杂、几何形态不一、表面覆盖多样化,导致融冻泥流阶地遥感解译和自动提取困难,然而,深度学习方法能获取上下文多尺度语义信息,提高特征表达的能力,为融冻泥流阶地的大范围提取提供了重要手段。因此,本文提出了一种基于DeepLab V3+深度学习模型和高分辨率光学遥感影像的融冻泥流阶地自动提取方法,并在四川甘孜州新都桥周边地区开展了实验研究。结果表明:与人工解译结果对比,本方法提取结果的综合精度达到0.68以上,并经野外调查验证了其有效性;在该区共识别了9 203条融冻泥流阶地,主要分布在新都桥镇附近的山谷两侧;泥流阶地主要朝西北方向,坡度集中分布在20°~25°,海拔高程大部分位于3 650~3 750 m间,主要地表覆盖类型是草地。

关键词: 深度学习, 高分辨率遥感, 融冻泥流阶地, 青藏高原东部


Solifluction terraces is a typical slope cryogenic glacial landform. A large number of ancient paleo-solifluction terraces are found in the eastern Tibetan Plateau, and their spatial distribution is important for reconstructing the distribution of ancient permafrost and paleoclimatic environment in the region. The complex texture, different geometric shapes and diverse surface coverage of solifluction terraces make remote sensing interpretation and automatic extraction of solifluction terraces very difficult. However, deep learning methods can acquire contextual multi-scale semantic information and improve feature representation, providing an important means for large-scale extraction of solifluction terraces. Therefore, this paper proposed an automatic extraction method of solifluction terraces based on the DeepLab V3+ deep learning model and high-resolution optical remote sensing images, and conducted experimental research in the surrounding area of Xinduqiao, Ganzi Prefecture, Sichuan. The results show that: (1) Compared with the visual interpretation results, the comprehensive accuracy of the extraction results of this method is above 0.68, and its effectiveness were verified by field investigation; (2) A total of 9 203 solifluction terraces were identified in this area, mainly distributed on both sides of the valley near Xinduqiao Town; (3) The solifluction terraces are mainly in the northwest direction, with a concentrated distribution of slopes ranging from 20° to 25°, most of the elevation is from 3 650 to 3 750 m, and the main surface cover type is grass.

Key words: deep learning, high-resolution satellite images, solifluction terraces, eastern Qinghai-Tibet Plateau


  • P642.14