Deep convolutional neural network approaches for the super-resolution of bathymetric maps
Abstract
Ocean bathymetry data at high resolution is essential for planning with regards to the prevention and mitigation of natural disasters, environmental stewardship, habitat mapping, resource extraction, and many other topics. However, approximately 80% of ocean seafloor topography has not yet been mapped. Social needs for the finer scale of bathymetric maps and the resolution of existing bathymetry maps have a major gap. JAMSTEC has launched the new research program "Mathematical Seafloor Geomorphology" to tackle the problem and to establish a method to provide high-resolution seafloor topography from existing low-resolution data. Super-resolution techniques based on deep convolutional neural networks (DCNNs) have increasingly been applied to the production of high-resolution images from low-resolution data. A range of different DCNN algorithms were tested on acoustically-acquired seafloor topographical data, including a Super-Resolution Convolutional Neural Network (SRCNN), a Fast Super-Resolution Convolutional Neural Network (FSRCNN), an Efficient sub-pixel convolutional neural network (ESPCN), a Super-Resolution Generative Adversarial Network (SRGAN) and an Enhanced SRGAN (ESRGAN). Sets of 100m and 50m mesh grid bathymetric maps of the same areas were prepared and were fed to each model to improve the resolution of the data obtained from the middle part of the Okinawa Trough (southwest Japan). 32x32 pixel images were extracted from the map with 100m mesh grid (Low_Res) while for the corresponding area in the 50m mesh grid map, 64x64 pixel images were extracted (High_Res), so as to sample exactly the same area of seafloor but at different resolutions. Compared with bicubic interpolation, all the DCNN models produced better high super-resolution performance for both peak signal-to-noise ratio (PSNR) and DSSIM (Structural Dissimilarity). Data preparation is a critical issue to develop a "practical model", so the effects of training data selection on the super-resolution performance were also investigated and are presented here.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMP004.0003H
- Keywords:
-
- 1906 Computational models;
- algorithms;
- INFORMATICS;
- 1914 Data mining;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS;
- 1952 Modeling;
- INFORMATICS