干旱区生态环境知识资源中心

Automatic classification of local seismic events which are only recorded at single stations poses great challenges because of weak hypocentre constraints. This study investigates how single-station event clusters relate to geographic hypocentre regions and common source processes. Typical applications arise in local seismic networks where reliable ground truth by a dense temporal network precedes or follows a sparse (permanent) installation. The seismic signals for this study comprise a 3-month subset from a field campaign to map subduction below northern Chile (PISCO ’94). Due to favourable ground noise conditions in the Atacama desert, the data set contains an abundance of shallow and deeper earthquakes, and many quarry explosions. Often event signatures overlap, posing a challenge to any signal processing scheme. Pattern recognition must work on reduced seismograms to restrict parameter dimensionality. Continuous parameter extraction based on noise-adapted spectrograms was chosen instead of discrete representation by, for example, amplitudes, onset times or spectral ratios to ensure consideration of potentially hidden features. Visualization of the derived feature vectors for human inspection and template matching algorithms was hereby possible. Because event classes shall comprise earthquake regions regardless of magnitude, clustering based on amplitudes is prevented by proper normalization of feature vectors. Principal component analysis is applied to further reduce the number of features used to train a self-organizing map (SOM). The SOM will topologically arrange prototypes of each event class in a 2-D map. Overcoming the restrictions of this black-box approach, the arranged prototypes could be transformed back to spectrograms to allow for visualization and interpretation of event classes. The final step relates prototypes to ground-truth information, confirming the potential of automated, coarse-grain hypocentre clustering based on single-station seismograms. The approach was tested by a twofold cross-validation whereby multiple sets of feature vectors from half the events are compared by a one-nearest neighbour classifier in combination with an Euclidean distance measure resulting in an overall correct geographic separation rate of 80.5 per cent.