Timo Tiira

Detecting teleseismic events using artificial neural networks

Abstract Artificial neural networks (ANNs) have been trained to recognize seismic signal onsets from vertical channel data. ANNs were trained using previously analysed events and noise samples recorded at 3 short period stations in central Finland. Separate nets were trained for each station. Comparisons were made between different net configurations and between different types of neural nets. The input to the nets consisted of four different STA/LTA values computed in seven frequency bands. The training data base was obtained from P-wave signals of 193 teleseismic events. The ANNs were trained to give high output values at onset and low output for noise and coda of events. After training the ANNs could produce a time series in which the signal onsets were shown as sharp peaks. When compared with the Murdock–Hutt detector the ANN detector could find 25% more events of the Reviewed Event Bulletins (REB) of the International Data Center. The detectors were tuned to produce the same total number of detections. When tuned to detect the same number of REB events as the Murdock–Hutt detector, the ANN detector produced over 50% less detections indicating a smaller false alarm rate. The type of ANN used was a multi-layer-perceptron (MLP) with one hidden layer. MLPs with two hidden layers and with a linear output layer were also tested but they clearly gave weaker results. Also, partially recurrent Elman and Jordan networks were tried but they showed a weaker detection capability than MLP. This type of detector could be used as a post-detector processor using outputs of other detectors of different types as its input combining the best features of each of the detectors.

Discrimination of nuclear explosions and earthquakes from teleseismic distances with a local network of short period seismic stations using artificial neural networks

Abstract Seismic discrimination capability of artificial neural networks (ANNs) was studied using earthquakes and nuclear explosions from teleseismic distances. The events were selected from two areas, which were analyzed separately. First, 23 nuclear explosions from Semipalatinsk and Lop Nor test sites were compared with 46 earthquakes from adjacent areas. Second, 39 explosions from Nevada test site were compared with 27 earthquakes from close-by areas. The basic discriminants were complexity, spectral ratio and third moment of frequency. The spectral discriminants were computed in five different ways to obtain all the information embedded in the signals, some of which were relatively weak. The discriminants were computed using data from six short period stations in Central and southern Finland. The spectral contents of the signals of both classes varied considerably between the stations. The 66 discriminants were formed into 65 optimum subsets of different sizes by using stepwise linear regression. A type of ANN called multilayer perceptron (MLP) was applied to each of the subsets. As a comparison the classification was repeated using linear discrimination analysis (LDA). Since the number of events was small the testing was made with the leave-one-out method. The ANN gave significantly better results than LDA. As a final tool for discrimination a combination of the ten neural nets with the best performance were used. All events from Central Asia were clearly discriminated and over 90% of the events from Nevada region were confidently discriminated. The better performance of ANNs was attributed to its ability to form complex decision regions between the groups and to its highly non-linear nature.

Automatic classification of seismic events within a regional seismograph network

This paper presents a fully automatic method for seismic event classification within a sparse regional seismograph network. The method is based on a supervised pattern recognition technique called the Support Vector Machine (SVM). The classification relies on differences in signal energy distribution between natural and artificial seismic sources. We filtered seismic records via 20 narrow band-pass filters and divided them into four phase windows: P, P coda, S, and S coda. We then computed a short-term average (STA) value for each filter channel and phase window. The 80 discrimination parameters served as a training model for the SVM. We calculated station specific SVM models for 19 on-line seismic stations in Finland. The training data set included 918 positive (earthquake) and 3469 negative (non-earthquake) examples. An independent test period determined method and rules for integrating station-specific classification results into network results. Finally, we applied the network classification rules to independent evaluation data comprising 5435 fully automatic event determinations, 5404 of which had been manually identified as explosions or noise, and 31 as earthquakes. The SVM method correctly identified 94% of the non-earthquakes and all but one of the earthquakes.The result implies that the SVM tool can identify and filter out blasts and spurious events from fully automatic event solutions with a high level of accuracy. The tool helps to reduce the work-load and costs of manual seismic analysis by leaving only a small fraction of automatic event determinations, the probable earthquakes, for more detailed seismological analysis. The self-learning approach presented here is flexible and easily adjustable to the requirements of a denser or wider high-frequency network. Fully automatic method for classification of seismic events.The method is based on Support Vector Machine.Effective in filtering out blasts and spurious events from automatic event bulletins.The method is flexible and easily adjustable to denser or wider networks.

Seismic lithosphere–asthenosphere boundary beneath the Baltic Shield

The problem of the existence of the asthenosphere for old Precambrian cratons is still discussed. In order to study the seismic lithosphere–asthenosphere boundary (LAB) beneath the Baltic Shield, we used records of nine local earthquakes with magnitudes ranging from 2.7 to 5.9. To model the LAB, original data were corrected for topography and Moho depth using a reference model with a 46-km-thick crust. For two northern events at Spitsbergen and Novaya Zemlya, we observe a low-velocity layer, 60–70-km-thick asthenosphere, and the LAB beneath Barents Sea was found at depth of c. 200 km. Sections for other events show continuous first arrivals of P-waves with no evidence for “shadow zone” in the whole range of registration, which could either be interpreted as the absence of the asthenosphere beneath the central part of the Baltic Shield, or that the LAB in this area occurs deeper (>200 km). The relatively thin low-velocity layer found beneath southern Sweden, 15 km below the Moho, could be interpreted as small-scale lithospheric heterogeneities, rather than asthenosphere. Differentiation of the lower lithosphere velocities beneath the Baltic Shield could be interpreted as regional heterogeneity or as anisotropy of the Baltic Shield lithosphere, with high velocities approximately in the east–west direction, and slow velocities approximately in the south–north direction.

Slowness vector correction for teleseismic events with artificial neural networks

Abstract The slowness anomalies cause serious location errors. The objective of this study is to create a mapping from observed slowness values to corrected values, which will provide more accurate locations. Artificial neural networks (ANNs) are efficient tools for mapping one multidimensional space to another. ANNs have been applied to compute slowness vector corrections for teleseismic events. Separate databases were used for training, testing and validating the networks. The training data set consisted of 2218 events in the period 1988–1992. An independent test database consisted of 1091 events from the year 1993 and the first half of 1994. The observed slowness vectors were computed using a three-station array of short period stations, KEF, SUF and KAF, in central Finland. The type of neural network was multi-layer perceptron. To improve the learning capability of the networks, a set of region-dependent extra inputs, resembling bias inputs, were added to the input layer. Several nets of different sizes were tested. The smallest net with only two hidden nodes gave best results. The median of error of the validation database dropped from 523 to 138 km. The median of error after correction is smaller than achieved with the method previously used with these stations. Due to the good interpolation capability of the neural net, the corrections decreased the location error even on areas which had no previous events in the training database. The method can be applied to slowness vector correction at any type of station or array, which produces slowness and azimuth values, if the mapping from the observed slowness values to calculated values is unambiguous.

Examining Three‐Dimensional Crustal Heterogeneity in Finland

The marriage of several high-quality seismic experiments in Finland over the past 30 years has shown that the saying “something old, something new, something borrowed” can result in the cost-efficient analysis of large-scale, three-dimensional (3-D) seismic structures. Standing alone, each data set gives a partial view of complex 3-D structures. When combined, they reveal a 3-D block structure embedded in a layered crust and enable the analysis of dynamics involved in forming stable cratonic crust. Efforts to collect large 3-D data sets around the globe include EarthScope (funded by the U.S. National Science Foundation (NSF)), the European Science Foundations (ESF) 4-D Topography Evolution in Europe: Uplift, Subsidence and Sea Level Change (TOPO-EUROPE), and the European Space Agencys Gravity Field and Steady-State Ocean Circulation Explorer (GOCE). Such endeavors are fundamental to modern crustal research. Huge emphasis is placed on collecting and archiving these data, but often only a fraction of data are used in initial studies. Fortunately, new data sets can be complemented with vintage ones (e.g., the NSF-funded Consortium for Continental Reflection Profiling (COCORP) and ESFs European GeoTraverse (EGT), as well as continent-wide science programs on continental evolution in Canada (LITHOPROBE), Europe (ESF-funded EUROPROBE), and the Himalayas (NSF-funded International Deep Profiling of Tibet and the Himalaya (INDEPTH)). Because fieldwork and archiving have already been completed for these vintage projects, new information can be extracted by new methods, with considerably less effort and funding.

Crustal and upper mantle velocity model along the DOBRE-4 profile from North Dobruja to the central region of the Ukrainian Shield: 2. geotectonic interpretation

This part of the paper addresses the geotectonic interpretation of the velocity model obtained from the results of seismic studies under the DOBRE-4 project in Ukraine. The velocity field does not show distinct lateral changes from the Precambrian platform towards the younger tectonic structures in the southwest. Hence, based on the seismic data alone, it is not possible to recognize the tectonic units that are known on the surface. The Moho has an undulating pattern over an interval with a length of ~150 km. The amplitude of the undulations reaches 8 to 17 km. The similar wavelike behavior, although on a shorter spatial scale and lower amplitude, is also typical of the upper crust and upper mantle. The presence of several separate horizons in the folded crust revealed by the velocity model is consistent with the presence of the folded systems which have different extensions on the different depth levels in the Earth’s crust. This situation is believed to be typical of folding on the lithospheric scale and to reflect the rheological stratification of the crust. The DOBRE-4 velocity section of the crust and adjacent part of the mantle promotes a clearer view of the geodynamical model describing the formation of the southwestern part of East European Platform in the Early Precambrian from the plate’s tectonic standpoint.

Crustal and upper mantle velocity model along the DOBRE-4 profile from North Dobruja to the central region of the Ukrainian Shield: 1. seismic data

For studying the structure of the lithosphere in southern Ukraine, wide-angle seismic studies that recorded the reflected and refracted waves were carried out under the DOBRE-4 project. The field works were conducted in October 2009. Thirteen chemical shot points spaced 35–50 km apart from each other were implemented with a charge weight varying from 600 to 1000 kg. Overall 230 recording stations with an interval of 2.5 km between them were used. The high quality of the obtained data allowed us to model the velocity section along the profile for P- and S-waves. Seismic modeling was carried out by two methods. Initially, trial-and-error ray tracing using the arrival times of the main reflected and refracted P- and S-phases was conducted. Next, the amplitudes of the recorded phases were analyzed by the finite-difference full waveform method. The resulting velocity model demonstrates a fairly homogeneous structure from the middle to lower crust both in the vertical and horizontal directions. A drastically different situation is observed in the upper crust, where the Vp velocities decrease upwards along the section from 6.35 km/s at a depth of 15–20 km to 5.9–5.8 km/s on the surface of the crystalline basement; in the Neoproterozoic and Paleozoic deposits, it diminishes from 5.15 to 3.80 km/s, and in the Mesozoic layers, it decreases from 2.70 to 2.30 km/s. The subcrustal Vp gradually increases downwards from 6.50 to 6.7–6.8 km/s at the crustal base, which complicates the problem of separating the middle and lower crust. The Vp velocities above 6.80 km/s have not been revealed even in the lowermost part of the crust, in contrast to the similar profiles in the East European Platform. The Moho is clearly delineated by the velocity contrast of 1.3–1.7 km/s. The alternating pattern of the changes in the Moho depths corresponding to Moho undulations with a wavelength of about 150 km and the amplitude reaching 8 to 17 km is a peculiarity of the velocity model.