9:00 AM - 9:15 AM
[SCG51-01] Data Augmentation of Ocean Bathymetry Data by Point Cloud Super-Resolution with Graph Convolution
Keywords:bathymetric charts, Deep learning, Point Cloud Super-Resolution
Bathymetric charts of the ocean floor are important not only for voyages but also for other fields of human activity such as energy, fisheries, environment, and disaster prevention. However, even today, observations do not cover the entire oceans. It is estimated that the GEBCO_2021 grid in the Seabed 2030 project (the Nippon Foundation-GEBCO) maps only about 20.6% of the ocean floor. The reason for this is that it is necessary to measure the seafloor depth at many points to create a highly accurate bathymetric charts, which is costly and time-consuming. Additionally, sonars used for bathymetry have advantages and disadvantages. For example, a single beam sonar is inexpensive, but only one point can be measured at a time, while a multi-beam sonar can simultaneously measure a wide range of points but is expensive.
A process of seafloor depth map creation consists of the following steps: (1) measure the seafloor depth using an underwater sonar and GPS for determination of the horizontal position; (2) reduction in the measurements noise through the sonar calibration (which considers water temperature), and removal of outliers; (3) making a map from measured points by kriging interpolation.
Efforts have been made to improve the resolution of bathymetric charts in step 3 by treating the gridded bathymetric data as digital images and using a neural network image processing technique called super-resolution to estimate high-resolution images from low-resolution images. Since bathymetric data is important for the detailing of bathymetric charts, it is worthwhile to approach research using neural networks also in step 1.
The field of image processing that handles three-dimensional point clouds has recently reached advanced levels in various fields. It is a reason for the appearance of a new data processing technology called point cloud super-resolution. The application of this technology permits the increase in the number of points in a point cloud while retaining its original characteristics. Note that data from seafloor depth measurements by sonar can be treated as three-dimensional point clouds. Therefore, in this study, we investigated the possibility of increasing the observation data by applying point cloud super-resolution technologies using graph-convolution neural networks.
For point cloud super-resolution, we use graph convolutional networks, AR-GCN and PU-GCN, and created a model that outputs twice as many points as the input points. The three-dimensional point cloud data used for a training and an evaluation is JAMSTEC seafloor topographic map data. From this original data, we extracted a high-resolution point cloud that mimics multi-beam sonar with parallel wakes, and then thinned out the data to create a low-resolution point cloud. Two types of thinning methods were assumed for the wake spacing: equal and unequal. For each model, we evaluated the effectiveness of point cloud super-resolution by comparing the bathymetry of the low-resolution and super-resolution point clouds with that of the high-resolution point cloud. Mean square error and SSIM were used as quantitative accuracy metrics, and we also evaluate quality of the super-resolution by using isobath lines.
As a result, the accuracy of the quantitative evaluation was generally higher when the low-resolution point cloud was used as the bathymetric charts, but the accuracy of the point cloud super-resolution method tended to be higher when the data had undulations. In addition, the bathymetric charts created by point cloud super-resolution seemed to express the characteristics of the seafloor topography locally better when isobath lines were used for the qualitative evaluation.
This suggests the possibility of obtaining bathymetric charts with more detailed topography by application point cloud super-resolution. Therefore, in the future, we will continue to study this technique in the context of bathymetry data.
A process of seafloor depth map creation consists of the following steps: (1) measure the seafloor depth using an underwater sonar and GPS for determination of the horizontal position; (2) reduction in the measurements noise through the sonar calibration (which considers water temperature), and removal of outliers; (3) making a map from measured points by kriging interpolation.
Efforts have been made to improve the resolution of bathymetric charts in step 3 by treating the gridded bathymetric data as digital images and using a neural network image processing technique called super-resolution to estimate high-resolution images from low-resolution images. Since bathymetric data is important for the detailing of bathymetric charts, it is worthwhile to approach research using neural networks also in step 1.
The field of image processing that handles three-dimensional point clouds has recently reached advanced levels in various fields. It is a reason for the appearance of a new data processing technology called point cloud super-resolution. The application of this technology permits the increase in the number of points in a point cloud while retaining its original characteristics. Note that data from seafloor depth measurements by sonar can be treated as three-dimensional point clouds. Therefore, in this study, we investigated the possibility of increasing the observation data by applying point cloud super-resolution technologies using graph-convolution neural networks.
For point cloud super-resolution, we use graph convolutional networks, AR-GCN and PU-GCN, and created a model that outputs twice as many points as the input points. The three-dimensional point cloud data used for a training and an evaluation is JAMSTEC seafloor topographic map data. From this original data, we extracted a high-resolution point cloud that mimics multi-beam sonar with parallel wakes, and then thinned out the data to create a low-resolution point cloud. Two types of thinning methods were assumed for the wake spacing: equal and unequal. For each model, we evaluated the effectiveness of point cloud super-resolution by comparing the bathymetry of the low-resolution and super-resolution point clouds with that of the high-resolution point cloud. Mean square error and SSIM were used as quantitative accuracy metrics, and we also evaluate quality of the super-resolution by using isobath lines.
As a result, the accuracy of the quantitative evaluation was generally higher when the low-resolution point cloud was used as the bathymetric charts, but the accuracy of the point cloud super-resolution method tended to be higher when the data had undulations. In addition, the bathymetric charts created by point cloud super-resolution seemed to express the characteristics of the seafloor topography locally better when isobath lines were used for the qualitative evaluation.
This suggests the possibility of obtaining bathymetric charts with more detailed topography by application point cloud super-resolution. Therefore, in the future, we will continue to study this technique in the context of bathymetry data.