Oceanographic Visualization
Visualization Tool Prototype
During the summer of 2021 I interned in the physics department at Johns Hopkins. I worked on a project to develop an oceanographic visualization tool to allow for quick viewing of large oceanographic simulation datasets. The visualization tool is primarily designed to allow for viewing of a ~2.5 km resolution 3D full earth ocean circulation simulation. When completed the dataset will be several petabytes of data, meaning an efficient visualization tool will be essential for quickly viewing and drawing conclusions about the data.
The simulation dataset is in an unusual coordinate system divided into 13 tiles, each 4320 by 4320 data points. There are also 90 separate depth levels and hourly time steps with the whole dataset storing a year of time steps. For each datapoint the local temperature, salinity, and ocean velocity are stored. The visualization tool projects the data into Web Mercator coordinates and allows individual tiles of the earth to be dynamically generated from the original dataset without any precomputation.
Dataset tile arrangement in lat-lon coordinates
Delaunay triangulation of antarctica
Image tiles are built from the dataset using 'interpolator' objects. Each image tile in the visualization tool has its own interpolator object that tracks which datapoints in the dataset are needed to build that tile. Images are built by generating a Delaunay triangulation of the datapoints and linearly interpolating each pixel value from the three datapoints enclosing it.
As the user zooms in on the visualization, higher resolution image tiles are generated from the dataset. This is achieved by resampling the dataset with higher resolutions for higher zoom levels and thus building the Delaunay triangulation with more closely spaced datapoints. Image tiles are always 256 by 256 pixels, but at the lowest zoom level a single tile renders that entire earth, while at the highest zoom level the earth is comprised of 64 by 64 image tiles, for a total resolution of 16384 by 16384. At any one time the viewer tool will need to render ~40 tiles to provide a full resolution display for the user.
In order to improve the rendering speed of the images we exclude datapoints that appear on land and are not relevant to the interpolation of any ocean pixel value. This was achieved by masking out all these data points before the calculation of the Delaunay triangulation. Further, for each tile, points not relevant to that tile are also masked out so that each interpolator maps the minimum possible number of data points from the dataset to the image.
Delaunay triangulation with land datapoints excluded
Green points are pixel values, blue points are data points, and red points are data points relevant to the interpolation of the pixel values.
Example Web Mercator tile at the north pole
The simulation dataset is stored as a Zarr file, a format that allows it to be broken into smaller chunks, each of which can be read independently of one another. This allows for rapid loading of small regions of the dataset; in the case of the simulation data each 4320 by 4320 tile is a single chunk. To further speed up the rendering process each interpolator was modified to contain only the indices of the datapoints relevant to that tile. This allows for the subset of data values to be efficiently extracted from dataset for calculation of the image values. The interpolation weights for each pixel value are also stored in the interpolator object so no recomputation of the Delaunay triangulation is necessary to build an image tile. The Zarr format of the dataset means for high zoom level tiles often only a single Zarr chunk needs to be read to extract the necessary data values for pixel value computation.
The original method of image rendering required repeated recomputation of pixel interpolation weights from data points. This method of image rendering results in a significant speed up over the original rendering technique. This allows the visualization tool to relatively smoothly transition between timesteps and depth levels for a fully animated view of the simulated data.
Different Fields in the Dataset
Ocean salinity
Ocean temperature
Curl of ocean velocity
