欢迎访问《生态学杂志》官方网站,今天是 分享到:

生态学杂志 ›› 2022, Vol. 41 ›› Issue (2): 404-416.doi: 10.13292/j.1000-4890.202202.003

• 技术与方法 • 上一篇    

基于Sentinel-1/2的大兴安岭木本沼泽信息提取方法

赵宇欣1,2,张冬有1*,毛德华2,杜保佳2,孙俊杰3   

  1. 1哈尔滨师范大学寒区地理环境监测与空间信息服务黑龙江省重点实验室, 哈尔滨 150025;2中国科学院东北地理与农业生态研究所湿地生态与环境重点实验室, 长春 130102; 3正元地理信息集团股份有限公司潍坊分公司, 山东潍坊 261000)
  • 出版日期:2022-02-10 发布日期:2022-08-10

Extent extraction method of swamp in the Greater Khingan Mountains based on Sentinel-1/2 images.

ZHAO Yu-xin1,2, ZHANG Dong-you1*, MAO De-hua2, DU Bao-jia2, Sun Jun-jie3   

  1. (1Heilongjiang Provincial Key Laboratory of Geographic Environment Monitoring and Spatial Information Service in Cold Regions, Harbin Normal University, Harbin 150025, China; 2Key Laboratory of Wetland Ecology and Environment, Northeast Institute of Geography and Agroecology, Chinese Academy of Sciences, Changchun 130102, China; 3Weifang Branch, Zhengyuan Geographic Information Group Co., Ltd., Weifang 261000, Shandong, China).
  • Online:2022-02-10 Published:2022-08-10

摘要: 木本沼泽的遥感信息提取一直是湿地研究的难点之一,在复杂环境地区传统调查方法无法深入,而利用遥感影像提取湿地分布信息大大提高了研究效率,对于理解全球变化具有重要意义。本研究选取位于黑龙江省西北部大兴安岭地区额木尔河流域的一景影像为研究案例,融合应用Sentinel1的雷达波段和Sentinel-2的红边波段,基于红边光谱特征和雷达特征等变量组合的特征波段集,采用深度学习和支持向量机方法对湿地分布信息进行提取,以期提高湿地信息提取和识别的精度。结果表明:在深度学习方法中,加入红边特征比只使用光谱特征总体精度提高了2.3%,达84.3%;加入雷达特征后,精度再度提高1.8%。即精度最高的融合光谱特征、红边特征以及雷达特征的方案,总体精度达86.5%,Kappa系数达0.84,其中林地精度达99.9%、草本沼泽达85.4%、木本沼泽达69.9%、水域达89.3%。最优深度学习模型迭代精度最高达95.7%。利用支持向量机方法进行对照试验,只使用光谱特征以及依次加入红边特征、雷达特征总体精度分别为74.4%、75.4%、77.3%,各项方案的总体精度均低于深度学习方法精度。本文采用深度学习方法对Sentinel-1/2结合的影像进行湿地信息精准提取,最高精度达86.5%,为今后更大尺度上木本沼泽的遥感信息精准提取研究提供了方法参考。

关键词: 木本沼泽, Sentinel, 红边波段, 深度学习, 支持向量机

Abstract: Due to environmental and ecosystem complexity, it is difficult to accurately mapping swamp in mountainous regions, especially in the inaccessible areas. Remote sensing provides a potential powerful tool for extracting the extent distribution of swamp, with great significance for understanding global change. In this study, we selected a scene extent of Sentinel imagery covering the Emur River Basin located in the Greater Khingan Mountains as the study area. The radar bands of Sentinel-1 with the red-edge bands of Sentinel-2, and other multispectral bands were combined to establish the feature bundles for delineating swamp. In order to improve the accuracy of swamp delineation, deep learning and support vector machine (SVM) were tested and compared. The results showed that: (1) For the deep learning method, the overall accuracy was improved by 2.3% to 84.3% when adding the red feature to the multispectral feature alone. Furthermore, when radar features were added, the accuracy was improved by another 1.8%. In other words, the combination of multispectral, rededge, and radar features got an overall accuracy of 86.5% and a Kappa coefficient of 0.84. The highest producer accuracy of each category was 99.9% for woodland, 85.4% for marsh, 69.9% for swamp, and 89.3% for waterbody. The maximum iteration accuracy of the optimal deep learning model was 95.7%. (2) The support vector machine method was used for the control test. The overall accuracies of only multispectral features, followed by the addition of red-edge features and radar features were 74.4%, 75.4% and 77.3%, respectively. The overall accuracy of each scheme was lower than that of deep learning method. In this study, deep learning method could extract the swamp extent information accurately from the Sentinel-1/2 images, with the highest accuracy of 86.5%, which could provide a method reference for future research on the accurate extraction of swamp at broader scale.

Key words: swamp, sentinel, red-edge band, deep learning, support vector machine.