Automated Layer Tracking: Soliciting interest in being involved in the project
We are interested in soliciting feedback and potentially joining efforts with other groups on the following layer tracking projects at CReSIS (Indiana University leading). Here we mean both internal layers and ice surface/bottom layers.
These two projects are 1) a layer tracking tool set that leverages image processing techniques to automate the process and 2) a geospatial database server and client architecture for storing, serving, and modifying layer information.
We would like to know if there are similar efforts underway at other institutions. If so, does it make sense to collaborate?
We would also like to find collaborators who currently manually pick layers that would like to help in testing the software and database services.
More detailed information about these two efforts:
I. Automated and Semi-automated Layer Tracking with Image Processing Technique
We are developing a tool for finding and labeling ice layer boundaries in ground-penetrating radar echograms. The tool attempts to locate the layer boundaries automatically using computer vision and image processing techniques, but can
also incorporate interactive user feedback to quickly correct mislabelings. The technique is based on a statistical framework called Markov Random Field models, which have recently been applied to a variety of problems in the computer vision community. These
models allow evidence from both local features (e.g. that layer boundaries should lie along edges in the echograms) and global features (e.g. that layer boundaries should be continuous and relatively smooth) to be combined together in a single probabilistic
framework. An efficient inference algorithm is then used to find the layer boundaries that best fit all of this evidence. The probabilistic framework has a number of advantages over other techniques based on edge detection or snake-based algorithms. First,
our framework postpones making hard classification decisions until all evidence has been combined together; this is especially helpful when local evidence is weak (e.g. very faint layer boundaries). Second, the models can naturally incorporate additional
constraints, such as known layer depths from bore samples or layer positions marked by a user. Finally, the probabilistic models give a natural measure of confidence in the estimated solutions, and can be used to sample multiple high-confidence hypotheses
instead of producing just a single solution.
II. Layer Tracking Database Server and Client
We have developed a geospatial database for storing and serving our ice layer information. Currently just ice surface and ice bottom for the radar depth sounder data are being served from it through a web interface:
http://polargrid.org/polargrid/software-release. We are working on expanding this database to support internal layer tracking with a number of features including: unlimited layers, linking and unlinking
layer instances in the database, tagging layers with a variety of information (such as tool used to generate, user who entered the layer information, allow duplicate picks of the same layer, layer notes/names, etc). We are also porting our current picker from
Matlab to Python and merging it to work with the database and the internal layer tracking tools. The intention is to make all of these tools cross platform and open source (including the backend database, map server, etc) so that distribution and licensing
are not an issue.