Structured Sparse Methodology for Early Tsunami Warning Systems

• Activity Number: 81
• Type: Contributed
• Date/Time: Sunday, August 4, 2013 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics and the Environment
• Author(s): Daniel Percival*+ and Donald Percival
• Companies: Google Inc. and University of Washington
• Keywords: Sparsity ; Lasso ; Tsunami ; Natural Disasters ; Warnings Systems ; Time Series

Abstract

In response to hazards posed by earthquake-induced tsunamis, the National Oceanographic and Atmospheric Administration has developed a system for issuing timely warnings to coastal communities. This system in part involves matching data collected in real time from deep-ocean buoys to a database of pre-computed geophysical models, each of which is associated with a geographical location. Currently, trained operators must handpick models from the database using the epicenter of the earthquake as guidance, which can be delay issuing of warnings. In this paper, we introduce an automatic procedure to select models to improve the timing and accuracy of these warnings. This procedure uses a lasso-like penalized and constrained linear least squares estimator in conjunction with a sweeping window to ensure spacially connected models, which is desirable from geophysical considerations. Test data from the 2006 Kuril Islands and the devastating 2011 Japan tsunamis show that the automatic procedure yields model ts and verication equal to or better than those from a time-consuming hand-selected solution.