Intelligent Trace Analysis for Automatic Extraction of Analog Seismographic Data

Seismic detection systems are one of the primary means used to monitor subsurface nuclear explosions. Consequently, the study of previous seismic events provides a critical underpinning for ongoing nonproliferation efforts. In this proposal, Retriever Technology will continue its efforts to develop accurate and automated software to extract digital X-Y trace information from historical analog seismograph data. Expanding on our previous work which significantly advanced understanding of the current problem, we have identified areas of research whose completion will lead to a successful and complete working model. Our work will focus on three main areas: We will use a variety of image pre-processing and edge and line detection techniques to isolate the traces as line features (moving beyond our previous work where they were broken into polygonal elements). Techniques will include (at least) variants of the Canny edge detector and Hough transforms, as well as selective frequency domain filters. While trace separation from background is important, the key step is assigning properties to each of those traces that will allow us to follow them across the seismogram when there are crossings, line drops and other features that make identification difficult. Using MatLab as our core analysis tool, we will identify line crossings, line drops and other discontinuous features. Once identified, we will use a variety of techniques to accurately follow the traces including linearized segmentation (which is much easier to use on traces-as-lines vs. traces-as- polygons that we have studied previously), curve fitting, in-range down selecting (i.e. amplitude-based selection) to identify and tag quiescent traces, and start/finish locations, all of which will provide input for heuristic accuracy checks. The outcome of this work will be to create a self-satisfying and robust set of criteria that will identify traces accurately and with a high degree of confidence, and in particular correctly track them across gaps and trace intersections. Incorporating our previous discoveries including distortion correction, FFT techniques to identify periodic features, automatic trace counting, and other important groundwork, we will provide a start-to-finish demonstration of automatic seismogram digitization.