The browser you are using is not supported by this website. All versions of Internet Explorer are no longer supported, either by us or Microsoft (read more here: https://www.microsoft.com/en-us/microsoft-365/windows/end-of-ie-support).

Please use a modern browser to fully experience our website, such as the newest versions of Edge, Chrome, Firefox or Safari etc.

photo of Zheng Duan on Lund webpage

Zheng Duan

Associate senior lecturer

photo of Zheng Duan on Lund webpage

Mapping inundation extents in Poyang Lake area using Sentinel-1 data and transformer-based change detection method

Author

  • Zhen Dong
  • Zifan Liang
  • Guojie Wang
  • Solomon Obiri Yeboah Amankwah
  • Donghan Feng
  • Xikun Wei
  • Zheng Duan

Summary, in English

Accurate and timely mapping of inundation extents during flood periods is essential for disaster evaluation and development of rescue strategies. With unique advantages over the optical sensors (e.g., little effect of clouds, and observations at day and night), Synthetic aperture radar (SAR) sensors provide an important data source for mapping inundation, particularly during flood periods. Freely available SAR images from Sentinel-1 have been increasingly used for many applications. This study applied an efficient transformer-based change detection method, bitemporal image transformer (BiT) with bitemporal Sentniel-1 images, to map inundation extents and evolution in Poyang Lake area in 2020. The transformer-based change detection method firstly adopted ResNet for high-level semantic features extraction, and applied a transformer mechanism to refine these features pixel-wise, followed by employing a FCN as the prediction head for generating the results of change detection. Besides, we constructed a water change detection dataset with spatial-and-temporal generalization from bitemporal Sentinel-1 images; this dataset consists of the seasonal variation water samples of Poyang Lake for years. We compared the results from the BiT method with other convolutional neural network (CNN) based methods (STANets and SNUNet). Mapped inundation extents were evaluated with the ground truth visually derived from high spatial resolution images. The evaluation showed the BiT method generated high accurate mapped inundation extents with the F1-score of 95.5%. The BiT model has proven its superior performance in detecting increased water. Based on the results of the BiT method, the variation of inundation extents in Poyang Lake during May-November 2020 was further analyzed. It was found that the water surface coverage of Poyang Lake is the smallest in late May; it gradually increased to the maximum on 14th July, and then began to stabilize and show a significant downward trend before November. The flood distribution map shows that cultivated land has been inundated with the largest area of approximately 600 km2.

Department/s

  • Dept of Physical Geography and Ecosystem Science
  • BECC: Biodiversity and Ecosystem services in a Changing Climate
  • MERGE: ModElling the Regional and Global Earth system

Publishing year

2023

Language

English

Publication/Series

Journal of Hydrology

Volume

620

Document type

Journal article

Publisher

Elsevier

Topic

  • Oceanography, Hydrology, Water Resources

Keywords

  • Change detection
  • Deep learning
  • Flood monitoring
  • Poyang Lake
  • SAR
  • Transformer

Status

Published

ISBN/ISSN/Other

  • ISSN: 0022-1694