Python Machine Learning Conference & GeoPython 2020

»Visualizing the sea ice of tomorrow - Converting Sentinel-1 radar images into high resolution forecast«
2020-09-21, 18:00–18:20, Room 2

Today's sea ice forecasts are at a coarse spatial resolution of several kilometers, whilst tactical navigation planning requires forecasts of much higher resolution. We present an approach to couple satellite and model information, to produce high resolution forecast images.

The talk focuses on the means we use to derive short term, high resolution sea ice forecasts in polar regions, aiming to assist navigation in icy waters. A test case will be presented as well as challenges affiliated with this task. The outline of the talk will be as follows:

  1. Motivation: State of the art in sea-ice forecasts. This includes data-type, sources, availability and policy. Motivation for high resolution forecasts, stating the current drawbacks and the need for tools that allow quick decision making.

  2. The method explained. The basic concept on how to propagate weather model information to high resolution satellite imagery. We use ice drift modeled data from TOPAZ4 ocean-sea ice data assimilation system which we couple to a static Sentinel-1 Synthetic Aperture Radar image. We use a computer vision algorithm from OpenCV to geometrically transform the image.

  3. The tools we use for the entire processing explaining the scope of each tool: python modules and bindings e.g. GDAL, netCDF4, shapely to read the radar dat, the SNAP module of the European Space Agency to enhance the image. This processing includes Radiometric calibration terrain correction, noise reduction. Finally we use OpenCV to propagate the ice drift information into the satellite image.

  4. Test case example and validation. We will present a forecast of a Synthetic Aperture Radar image taken from Sentinel-1 satellite. In this experiment we've selected an area and a time window in which we had subsequent satellite recordings. This allows us to use the first image as an input for the forecast and the image from the later recording we use it as a "known truth" to validate the forecast.

  5. Current challenges and possible solutions. CPU time is one issue we face. Potential solutions we examine include to move some time consuming functions to C# and the use of machine learning technics. Further due to this approach artifacts near the shore are present, which we need to consider.