地球信息科学学报 ›› 2020, Vol. 22 ›› Issue (12): 2445-2455.doi: 10.12082/dqxxkx.2020.200338

• 遥感科学与应用技术 • 上一篇    

LSTM支持下时序Sentinel-1A数据的太白山区植被制图

杨丹1(), 周亚男1,*(), 杨先增1, 郜丽静2, 冯莉1   

  1. 1.河海大学水文水资源学院,南京 211100
    2.中国科学院空天信息创新研究院 遥感科学国家重点实验室,北京 100010
  • 收稿日期:2020-06-30 修回日期:2020-09-15 出版日期:2020-12-25 发布日期:2021-02-25
  • 通讯作者: 周亚男 E-mail:danyanggis@163.com;zhouyn@hhu.edu.cn
  • 作者简介:杨 丹(1996— ),女,安徽枞阳人,硕士生,主要从事遥感植被分类研究。E-mail: danyanggis@163.com
  • 基金资助:
    国家重点研发计划项目(2019YFC1804301);中央高校基本科研业务费项目(B200202008);遥感科学国家重点实验室开放基金(OFSLRSS201919);空间数据挖掘与信息共享教育部重点实验室(福州大学)开放基金(2019LSDMIS04)

Vegetation Mapping in Taibai Mountain Area Supported by LSTM with Time Series Sentinel-1A Data

YANG Dan1(), ZHOU Yanan1,*(), YANG Xianzeng1, GAO Lijing2, FENG Li1   

  1. 1. College of Hydrology and Water Resources, Hohai University, Nanjing 211100, China
    2. State Key Laboratory of Remote Sensing Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100010, China
  • Received:2020-06-30 Revised:2020-09-15 Online:2020-12-25 Published:2021-02-25
  • Contact: ZHOU Yanan E-mail:danyanggis@163.com;zhouyn@hhu.edu.cn
  • Supported by:
    National Key Research and Development Program(2019YFC1804301);Fundamental Research Funds for the Central Universities(B200202008);Open Fund of State Key Laboratory of Remote Sensing Science(OFSLRSS201919);Opening Foundation of Key Lab of Spatial Data Mining & Information Sharing, Ministry of Education (Fuzhou University)(2019LSDMIS04)

摘要:

植被分类是森林资源调查与动态监测的基础与前提。当前植被分类研究大都利用光学遥感影像,然而,光学遥感成像易受到云雨覆盖的影响,难以构建完整时间序列,植被分类精度有限。微波遥感具有全天时全天候、时间序列完整的优势,在植被调查与分析中具有巨大的应用潜力。本文利用2018年Sentinel-1A微波遥感时间序列数据和深度循环网络方法,对秦岭太白山区的森林植被进行分类制图。首先利用Sentinel-2光学影像与数字高程数据对研究区进行多尺度分割;然后将处理后的时间序列Sentinel-1A数据空间叠加到分割地块上,构建地块的多元时间序列曲线;最后利用深度循环网络提取与学习多元时间序列的时序特征并分类。实验结果表明:① 与传统机器学习方法(如RF、SVM)相比,本文提出的深度循环网络方法的分类精度提高10%以上;② 在Sentinel-1A微波极化特征组合中VV+VH表现最好,与VV+VH+VV/VH极化特征组合的精度相近;③ 使用全年的时间影像构建时间序列分类精度最高,达到82%。研究表明,利用深度循环网络与时间序列Sentinel-1A数据的方法能够有效提高植被分类的精度,从数据源与分类方法上为森林植被分类研究提供了新的思路。

关键词: 植被分类, 太白山, 时间序列, Sentinel-1A数据, 深度循环网络, 微波遥感, 机器学习

Abstract:

Vegetation classification is the basis and premise of forest resource investigation and dynamic monitoring. Remote sensing techniques have long been important means of forest monitoring with their ability to quickly and efficiently collect the spatial-temporal variability of vegetation. Vegetation classification is a key issue for forest monitoring and is critical to many remote sensing applications in the domain of precision forestry such as vegetation area estimation. Remote sensing applications in vegetation classification have traditionally focused on the use of optical data such as MODIS. However, due to cloud and haze interference, optical images are not always available at phenological stages that are essential to vegetation identification, making it difficult to construct complete time-series vegetation growth and limiting the vegetation classification accuracy. Unlike passive visible and infrared wavelengths which are sensitive to cloud and light, active SAR (Synthetic Aperture Radar) is particularly attractive for vegetation classification due to its all-weather, all-day imaging capabilities. In addition, SAR provides information on the stem and leaf structures of vegetation and is sensitive to soil roughness and moisture content, making it effective in forest applications. In this study, a deep-learning-based time-series analysis method employing multi-temporal SAR data is presented for forest vegetation classification in the Taibai Mountain (the main peak of Qinling Mountains). Firstly, Sentinel-2 optical images and digital elevation data in the study area were used for multi-scale segmentation to produce a precise farmland map. Then pre-processed SAR intensity images were overlaid with the farmland map to construct time-series vegetation growth for each parcel. Finally, a deep-learning-based classifier using the Long Short-Term Memory (LSTM) network was employed to learn time-series features of vegetation and to classify parcels to produce a final classification map. The experimental results show that: (1) Compared with traditional machine learning methods (such as Random Forest and Support Vector Machine), the classification accuracy of the deep-learning-based method proposed in this paper was improved by more than 10%; (2) Among different combinations of Sentinel-1A polarizations, VV+VH performed best, having a similar accuracy with the VV+VH+VV/VH; (3) Time-series classification using all images in the whole year achieved the best performance, with an overall accuracy of 82% using VV+VH. The study shows that the combination LSTM network and time-series Sentinel-1A data can effectively improve the accuracy of vegetation classification and provide new ideas for forest vegetation classification from the perspectives of data source and classification method.

Key words: vegetation classification, Taibai Mountain, time series, Sentinel-1A data, LSTM network, microwave remote sensing, machine learning