Performance Evaluation and Improvement of Shoreline Detection Using Sentinel-1 SAR and DEM Data

Abstract

This study improved the shoreline detection performance based on the U-Net model by combining Sentinel-1 synthetic aperture radar (SAR) and digital elevation model (DEM) data. The U-Net network was first modified to enhance feature extraction by using the MobileNetV3 backbone architecture and convolutional block attention module. To alleviate the performance degradation of shoreline detection caused by radar shadow, especially in coastal areas with large terrain undulations, SAR and DEM data were combined as input to U-Net. Furthermore, this study evaluated the shoreline detection performance using the statistical analysis based on the proposed probabilistic model of distance difference between the detected shoreline and reference data which was provided by Construction and Planning Agency Ministry of the Interior, Taiwan government. The experiment was conducted based on two self-built datasets, one containing 4061 SAR images and the other containing 3822 SAR images and corresponding DEM data, both collected in the coastal areas of Taiwan from 2016 to 2019. The experimental results showed that compared with the U-Net network using SAR data, the modified U-Net has achieved superior performance in shoreline detection for various coastal landforms. Moreover, the addition of DEM data reduced the influence of radar shadow, making shoreline detection results more consistent with reference data. Finally, the generalization ability of the modified U-Net in shoreline detection was also verified by testing images from regions not included in the built dataset.