[S07-4] Hazard and risk assessment, and data processing strategies

[S07-4-04] Automatic detection of earthquakes, quarry blasts, rockfalls and avalanches on the Swiss permanent broadband network

Conny Hammer, ^○Donat Faeh (Swiss Seismological Service, ETH Zurich, Switzerland)

In common observatory practice, seismic events are detected automatically using a classical STA/LTA trigger. Then, a trained expert manually assigns a specific seismic signal class such as tectonic event or quarry blast to verified detections. However, this approach absorbs manpower and happens generally with considerable time-delay. To overcome these problems we test an automatic classification approach on the Swiss permanent broadband network. The approach is based on a probabilistic description of the signals, called hidden Markov models, that allows the robust identification of corresponding signals and has been successfully applied in various environments. The automatic system is set up at 15 stations of which most are located in alpine regions. Once a transient is detected by the preceding trigger, the signal is fed into the automatic system and labeled as one of the following signal types: earthquake, quarry blast, rockfall, avalanche or noise. So far, the procedure is tested on a data period of one year. To evaluate the system performance automatically and manually assigned labels are compared. Achieved results show large differences between individual stations. While at some stations classification rates of up to 95 % are achieved, classification rates are much lower at other stations. The overall performance can be further improved by using the complete network information. For this, the classification results of different stations needs to be combined. We present various approaches to combine the outcome of the automatic system dependent on source-receiver distance and confidence of classification result. A majority based voting system completes the post processing. In the next step, we aim to detect signals directly on the continuous data stream, in order to avoid missing small local events due to the preceding trigger algorithm. The performance provides a base for deciding about implementation of an early information system.