地球信息科学学报 ›› 2022, Vol. 24 ›› Issue (4): 766-779.doi: 10.12082/dqxxkx.2022.210489

• 遥感科学与应用技术 • 上一篇    下一篇

融合边缘特征与语义信息的人工坑塘精准提取方法

杨先增1(), 周亚男1,*(), 张新2, 李睿1, 杨丹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 100101
  • 收稿日期:2021-08-20 修回日期:2021-09-30 出版日期:2022-04-25 发布日期:2022-04-13
  • 通讯作者: *周亚男(1987— ),男,河南漯河人,博士,副教授,主要从事基于深度学习的遥感信息提取、遥感时序分析与农业 遥感。E-mail: zhouyn@hhu.edu.cn
    *周亚男(1987— ),男,河南漯河人,博士,副教授,主要从事基于深度学习的遥感信息提取、遥感时序分析与农业 遥感。E-mail: zhouyn@hhu.edu.cn
    *周亚男(1987— ),男,河南漯河人,博士,副教授,主要从事基于深度学习的遥感信息提取、遥感时序分析与农业 遥感。E-mail: zhouyn@hhu.edu.cn
  • 作者简介:杨先增(1997— ),男,安徽六安人,硕士生,主要从事深度学习遥感应用研究。E-mail: yangxz19970909@163.com
  • 基金资助:
    国家自然科学基金项目(42071316);中央高校业务费项目(B200202008);自然资源部地理国情监测重点实验室开放基金项目(2020NGCM03);农业产业数字化地图(21C00346)

Accurate Extraction of Artificial Pit-pond Integrating Edge Features and Semantic Information

YANG Xianzeng1(), ZHOU Ya'nan1,*(), ZHANG Xin2, LI Rui1, YANG Dan1   

  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 100101, China
  • Received:2021-08-20 Revised:2021-09-30 Online:2022-04-25 Published:2022-04-13
  • Supported by:
    National Natural Science Foundation of China(42071316);Fundamental Research Funds for the Central Universities(B200202008);Open Foundation of Key Laboratory of National Geographic Census and Monitoring, Ministry of Natural Resources(2020NGCM03);Chongqing agricultural industry digital map projec(21C00346)

摘要:

针对高空间分辨率遥感影像目标提取中定位精度低、边缘粗糙等问题,提出一种融合目标边缘特征与语义信息的人工坑塘提取网络模型。方法首先利用改进的U-Net语义分割网络模块来提取遥感影像中丰富的目标语义信息,然后拓展上述语义分割网络构建边缘提取子网络来获取遥感影像的多尺度边缘特征,最后借助于编码-解码子网络融合边缘特征与语义信息,实现遥感影像目标的精准提取。将该方法运用到雷州半岛复杂背景条件下人工坑塘提取实验中,实验结果中本文提出的方法在F分数以及边界F分数等评价指标上表现最优,达到97.61%与83.01%,验证了融合高层语义信息结合低层的边缘特征在提升遥感目标提取精确度上的有效性。

关键词: 高分遥感, 深度学习, 语义分割, 边缘提取, 目标提取, 人工坑塘, 特征融合, 多任务学习

Abstract:

High-resolution remote sensing images have more detailed spatial, geometric, and textural features, which provides useful visual description features such as spot position, shape, and texture, and reliable and abundant data sources for accurate extraction of spatial elements. However, traditional methods require the researchers to extract these features manually and have some limitation such as low positioning accuracy and rough edges. With the development of deep learning, it can extract typical elements such as water bodies, buildings, and roads from remote sensing images with higher accuracy and without the support of prior knowledge. The extracted element information can provide a data basis for innovative applications in urban and rural land resource actuarial calculation and planning, disaster risk assessment, and industrial output evaluation and estimation. However, traditional deep learning semantic segmentation methods focus more on the improvement of semantic segmentation accuracy in the extraction process of remote sensing elements and pay less attention to boundary accuracy. In view of the existing problems of deep learning methods in target extraction from high resolution remote sensing images, such as rough edge and much noise, a network model combined with edge and semantic features of targets was proposed to extract the artificial pit-pond. The improved U-Net semantic segmentation network was used to extract rich semantic information of targets in remote sensing images, which could be developed in edge structure and sub-network extraction, thus acquiring multi-scale edge features in remote sensing image. In this case, an encoding-decoding subnetwork combined with edge features and semantic information were applied to extract remote sensing image objects accurately. Meanwhile, the synchronous extraction of boundary information was also realized, and feature fusion and noise screening were realized through the encoding-decoding subnetwork. The proposed method was used to extract artificial pit-pond in a complicated background condition in Leizhou Peninsula. First, we designed labeled training and testing images for the experiment and performed data augmentation to increase the number of samples. Second, we provided a series of evaluation indicators for the extraction effect. Finally, we evaluated the performance of the model from multiple perspectives including semantic accuracy and boundary. Results show that the method proposed in this paper had the best performance in the evaluation, the F score and boundary F score reached 97.61% and 83.01%, respectively, which demonstrated the effectiveness of the fusion of high-level semantic information and low-level edge features in improving the accuracy of remote sensing target extraction.

Key words: high-resolution remote sensing, deep learning, semantic segmentation, edge extraction, target extraction, artificial pit-pond, feature fusion, multi-task learning