TY - JOUR
T1 - Super-resolution reconstruction of the 1 arc-second Australian coastal DEM dataset
AU - Zhang, Bo
AU - Shi, Zekai
AU - Hong, Danfeng
AU - Wang, Qi
AU - Yang, Jian
AU - Ren, Haoyuan
AU - Zhang, Meng
N1 - Publisher Copyright:
© 2025 Wuhan University. Published by Informa UK Limited, trading as Taylor & Francis Group.
PY - 2025
Y1 - 2025
N2 - Seafloor mapping to create a coastal DEM dataset of the oceans is significant for various applications. Land-based DEM monitoring and acquisition are more practical as compared to ocean DEM. The highest resolution free available DEM data in the Australian land area is 30 m (1 arc-second), whereas for the ocean area, it is only 450 meters (15 arc-seconds) due to the slow and costly nature of sonar data collection. Deep learning-based image super-resolution (SR) has demonstrated high computational efficiency and optimal performance in enhancing bathymetric resolution. However, direct training a robust network for the entire Australian DEM-SR remains a challenge due to the difficulty of obtaining sufficiently high-resolution ocean DEM data. Therefore, we proposed a novel method that incorporates two key measures to address this issue. First, we designed a deep gradient prior AUSDEM-SRNet network based on numerous land sample sets to acquire prior knowledge from land data. Second, we introduced transfer learning to leverage the similarities between land and bathymetric (ocean) data, compensating for the scarcity of high-resolution bathymetric information. Through experiments using bathymetric (ocean) data around Australia, the proposed method demonstrated superior performance compared to naive bicubic interpolation, achieving a 11.64% reduction in root mean square error (RMSE) and a 1.93% increase in peak signal-to-noise ratio (PSNR) averages. The AUSDEM-SRNet model has been meticulously developed and evaluated for its capability to amplify the spatial resolution of existing DEM dataset by up to 15 times without requiring additional data or information related to the original dataset. As a result, we generated an open-access Australian coastal DEM dataset with 1 arc-second, named AUSDEM_2023. This method can significantly reduce the number of sea areas or points that must be measured, thereby accelerating the detailed mapping of the coastal seabed and the creation of high-resolution sounding maps for various coastal countries.
AB - Seafloor mapping to create a coastal DEM dataset of the oceans is significant for various applications. Land-based DEM monitoring and acquisition are more practical as compared to ocean DEM. The highest resolution free available DEM data in the Australian land area is 30 m (1 arc-second), whereas for the ocean area, it is only 450 meters (15 arc-seconds) due to the slow and costly nature of sonar data collection. Deep learning-based image super-resolution (SR) has demonstrated high computational efficiency and optimal performance in enhancing bathymetric resolution. However, direct training a robust network for the entire Australian DEM-SR remains a challenge due to the difficulty of obtaining sufficiently high-resolution ocean DEM data. Therefore, we proposed a novel method that incorporates two key measures to address this issue. First, we designed a deep gradient prior AUSDEM-SRNet network based on numerous land sample sets to acquire prior knowledge from land data. Second, we introduced transfer learning to leverage the similarities between land and bathymetric (ocean) data, compensating for the scarcity of high-resolution bathymetric information. Through experiments using bathymetric (ocean) data around Australia, the proposed method demonstrated superior performance compared to naive bicubic interpolation, achieving a 11.64% reduction in root mean square error (RMSE) and a 1.93% increase in peak signal-to-noise ratio (PSNR) averages. The AUSDEM-SRNet model has been meticulously developed and evaluated for its capability to amplify the spatial resolution of existing DEM dataset by up to 15 times without requiring additional data or information related to the original dataset. As a result, we generated an open-access Australian coastal DEM dataset with 1 arc-second, named AUSDEM_2023. This method can significantly reduce the number of sea areas or points that must be measured, thereby accelerating the detailed mapping of the coastal seabed and the creation of high-resolution sounding maps for various coastal countries.
KW - Australian coastal
KW - deep gradient prior network
KW - Digital elevation model (DEM) dataset
KW - super-resolution
KW - transfer learning
UR - http://www.scopus.com/inward/record.url?scp=105002458314&partnerID=8YFLogxK
U2 - 10.1080/10095020.2025.2487143
DO - 10.1080/10095020.2025.2487143
M3 - 文章
AN - SCOPUS:105002458314
SN - 1009-5020
JO - Geo-Spatial Information Science
JF - Geo-Spatial Information Science
ER -