地球信息科学学报 ›› 2020, Vol. 22 ›› Issue (10): 2010-2022.doi: 10.12082/dqxxkx.2020.190622

• 专栏:城乡生态环境综合监测 • 上一篇    下一篇

基于改进U-Net网络的高分遥感影像水体提取

何红术1,2(), 黄晓霞1,*(), 李红旮1, 倪凌佳1,2, 王新歌3, 陈崇3, 柳泽3   

  1. 1.中国科学院空天信息创新研究院,北京 100094
    2.中国科学院大学,北京 100049
    3.住房和城乡建设部城乡规划管理中心,北京 100835
  • 收稿日期:2019-10-24 修回日期:2020-02-16 出版日期:2020-10-25 发布日期:2020-12-25
  • 通讯作者: 黄晓霞 E-mail:hhs10010@163.com;huangxx@aircas.ac.cn
  • 作者简介:何红术(1993- ), 男,重庆人,硕士生,研究方向为深度学习遥感信息提取、地理信息系统。E-mail:hhs10010@163.com
  • 基金资助:
    国家重点研发计划项目(2017YFB0503905);国家自然科学基金项目(41971363)

Water Body Extraction of High Resolution Remote Sensing Image based on Improved U-Net Network

HE Hongshu1,2(), HUANG Xiaoxia1,*(), LI Hongga1, NI Lingjia1,2, WANG Xinge3, CHEN Chong3, LIU Ze3   

  1. 1. Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
    3. Urban and Rural Planning Management Center of the Ministry of Housing and Urban-Rural Development,Beijing 100835, China
  • Received:2019-10-24 Revised:2020-02-16 Online:2020-10-25 Published:2020-12-25
  • Contact: HUANG Xiaoxia E-mail:hhs10010@163.com;huangxx@aircas.ac.cn
  • Supported by:
    National Key Research and Development Program of China(2017YFB0503905);National Natural Science Foundation of China(41971363)

摘要:

传统基于光谱信息的水体提取未能考虑水体形状、纹理、大小、相邻关系等问题,且存在同物异谱、异物同谱现象,导致水体提取精度较低。而传统基于分类提取水体方法设计特征过程较为繁琐,且不能挖掘深度信息特征。因此,本文提出改进的U-Net网络语义分割方法,借鉴经典U-Net网络的解编码结构对网络进行改进:① 将VGG网络用于收缩路径以提取特征;② 在扩张路径中对低维特征信息进行加强,将收缩特征金字塔上一层的特征图与下一层对应扩张路径上的特征图进行融合,以提高提取结果分割精度;③ 在分类后处理中引入条件随机场,以将分割结果精细化。在保持相同训练集、验证集和测试集的情况下,分别用SegNet、经典U-Net网络和改进的U-Net网络做对照试验。试验结果表明,改进的U-Net网络结构在IoU、精准率和Kappa系数指标上均高于SegNet和经典U-Net网络,与SegNet相比,3项指标分别提升了10.5%、12.3%和0.14,与经典U-Net网络结果相比,各个指标分别提升了5.8%、4.4%和0.05。改进的网络水体提取结果较为完整,对小目标水体能够准确提取。改进的U-Net网络能够有效地实现水体提取任务。

关键词: U-Net, 水体提取, 高分遥感影像, 条件随机场, 图像分割, VGG16, 青岛, 西宁

Abstract:

There are two main methods of traditional water body extraction: a method based on spectral information and a method based on classification. Traditional water body extraction methods based on spectral information fail to take into account features such as water body shape, internal texture, water body size, and adjacent relations of water body. Also, there is a common phenomenon of “same object with different spectra and same spectrum with different objects”, which could result in low accuracy of water body extraction. Thus, the traditional methods that design features based on classification to extract water body is complex and impossible to capture the deep information of water body features. This paper proposed an improved U-Net network semantic segmentation method, which uses the de-encoding structure of the classic U-Net network to improve the network: ① Use the VGG network to shrink the path and increase the depth of the network to extract deep features of the water; ② Strengthen the low-dimensional feature information in the expansion path, fuse the feature map on the next layer of the shrinking feature pyramid with the feature map on the corresponding expansion path in the next layer, and enhance the model's low-dimensional feature information to improve the classification accuracy of the model; and ③ The Conditional Random Feld (CRF) was introduced in the post-classification process to refine the segmentation results and improve the segmentation accuracy. In the study of Qingdao area, SegNet, classic U-Net network, and improved U-Net network were selected as controlled experiments while maintaining the same training set, validation set, and test set. The test results show that the improved U-Net network structure performed better than SegNet and classic U-Net networks in terms of IoU, accuracy rate and Kappa coefficient. Compared with SegNet, the three indicators increased by 10.5%, 12.3%, and 0.14, respectively. Compared with the results of the classic U-Net network, each indicator increased by 5.8%, 4.4% and 0.05, respectively. The results demonstrate the effectiveness of the improved method in this paper. In addition, the method proposed in this paper has more advantages than the other two networks in the extraction of small targets in the test area, the completeness of water body extraction, the distinction between shadows and water bodies, and the accuracy of boundary segmentation. In order to verify the space-time scalability of the model, this paper chose western Qingdao and Xining, Qinghai as the verification areas. The verification results show that the water body extraction was good for areas similar to the geographical environment of the experimental area, and the effect of water body extraction needs to be further improved in places that differ greatly from the geographical environment of the experimental area. In general, the improved U-Net network can effectively achieve the task of water extraction.

Key words: U-Net, water extraction, high-resolution remote sensing image, conditional random field, image segmentation, VGG16, Qingdao, Xining