地球信息科学学报 ›› 2022, Vol. 24 ›› Issue (4): 597-616.doi: 10.12082/dqxxkx.2022.210512

• 综述 •    下一篇

国土调查遥感40年进展与挑战

舒弥1,2(), 杜世宏1,2,*()   

  1. 1.自然资源部城市国土资源监测与仿真重点实验室,深圳 518034
    2.北京大学遥感与GIS研究所,北京 100871
  • 收稿日期:2021-08-26 修回日期:2021-09-23 出版日期:2022-04-25 发布日期:2022-06-25
  • 通讯作者: *杜世宏(1975— ),男,甘肃靖远人,研究员,主要从事时空数据智能理解方法研究。E-mail: smilegis@163.com
  • 作者简介:舒 弥(1994— ),男,湖北汉川人,博士,主要从事遥感影像智能解译、深度学习算法研究。E-mail: shumimi@pku.edu.cn
  • 基金资助:
    自然资源部城市国土资源监测与仿真重点实验室开放基金项目(40271090)

Forty Years' Progress and Challenges of Remote Sensing in National Land Survey

SHU Mi1,2(), DU Shihong1,2,*()   

  1. 1. Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources, Shenzhen 518034, China
    2. Institute of Remote Sensing and GIS, Peking University, Beijing 100871, China
  • Received:2021-08-26 Revised:2021-09-23 Online:2022-04-25 Published:2022-06-25
  • Supported by:
    Open Fund of Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources(40271090)

摘要:

运用遥感技术进行土地资源调查,摸清其数量及分布状况,长期以来都是遥感领域研究的重要内容。本文首先回顾了过去40年来遥感技术在我国国土调查中的应用情况,然后围绕高分辨率影像的特征提取、大范围影像的样本获取、多时相/多传感器影像的迁移学习以及多源异构数据融合4个方面介绍了相关进展情况;接着归纳总结了现有遥感信息提取技术在国土调查中面临的4个挑战:① 高分辨率影像分类存在如何定义、选择、挖掘高级特征的问题;② 国土调查中的遥感数据集规模庞大,存在着类间不平衡和类内多样性,为这种复杂数据集获取足够、均衡、多样化的样本集是一个巨大挑战;③ 对于多传感器/多时相影像,如何低成本、及时地实现土地利用分类是值得考虑的问题;④ 从土地覆盖到土地利用存在语义鸿沟,如何合适地引入语义信息以弥合语义鸿沟需要被考虑。最后,本文对国土调查遥感技术的未来发展方向和应用点进行了展望。

关键词: 国土调查, 土地利用分类, 遥感信息提取, 特征优选, 小样本, 样本平衡, 样本多样性, 迁移学习, 语义鸿沟

Abstract:

The national land survey is a major component of evaluating national conditions and strength. Its main purpose is to master the detailed national land use status and natural resource changes. It is of great significance to cultivated land protection and sustainable social and economic development. With the development of remote sensing technology, investigating the status, quantity, and distribution of land resources has always been the focus of remote sensing applications. This article reviews the application of remote sensing in national land survey over the past four decades. Until now, remote sensing technology has shown broad prospects in national land survey. However, the remote sensing information extraction in national land survey still mainly relies on visual interpretation and is not automated enough. In recent years, the remote sensing data tend to have the characteristics of high-resolution, large-scale, multi-temporal, and multi-sensor. However, the existing automated information extraction technology does not fully integrate those characteristics, hindering the application in national land survey. This article first introduces the relevant progress in national land survey from four aspects: feature extraction using very-high-resolution images, samples acquisition from large-scale images, transfer learning in multi-temporal/multi-sensors images, and multi-source heterogeneous data fusion. Then, four challenges in the existing remote sensing information extraction technology in the national land survey are summarized: ① Image feature is the key to image classification. There are questions on how to define and select features. In addition, high-resolution images put forward higher requirements for advanced feature extraction; ② Remote sensing data in the national land survey are usually large in scale, and there are inter-class imbalance and intra-class diversity. Therefore, it is a challenge to obtain sufficient, balanced, and diverse sample sets from such complex data set; ③ Generally, the efficiency of sample collection cannot catch up with the accumulation speed of remote sensing data, thus the labeled samples are relatively small compared with the data. For multi-sensor/multi-temporal imagery, how to realize land use classification in a low-cost and timely manner is a question worth considering; ④ There is a semantic gap between land cover and land use. Since remote sensing images mainly reflect land cover information, how to properly introduce semantic information to bridge the semantic gap and realize land use classification is a problem. Finally, the future development and application of remote sensing technology in national land survey are prospected, such as transformation from visual interpretation to artificial intelligence technology, accuracy and consistency assessment of remote sensing classification products in land survey, crowdsourcing methods for large-scale land use production, and update of large-scale land use data.

Key words: land survey, land use classification, remote sensing information extraction, feature optimization, small sample, sample balance, sample diversity, transfer learning, semantic gap