The role of instrumental versus macroseismic locations for earthquakes of the last century : a discussion based on the seismicity of the North-Western Apennines ( Italy )

Many seismological observatories began to record and store seismic events in the early years of the twentieth century, contributing to the compilation of very valued databases of both phase pickings and waveforms. However, despite the availability of the instrumental data for some of the events of the last century, an instrumental location for these earthquakes is not always computed; moreover, when available, the macroseismic location is strongly preferred even if the number of points that have been used for it is low or the spatial distribution of the observations is not optimal or homogeneous. In this work I show how I computed an instrumental location for 19 events which occurred in the Garfagnana-Lunigiana region (Northern Tuscany, Italy) beginning from 1902. The location routine is based on a Joint Hypocentral Determination in which, starting from a group of master events, the systematic errors that may affect the data are summed up in the corrective factors complementing the velocity propagation model. All non-systematic errors are carefully checked and possibly discarded by going back to the original data, if necessary. The location is then performed using the classic approach of the inverse problem and solved iteratively. The obtained locations are then compared to those already available from other macroseismic studies with the aim to check the role to be attributed to the instrumental locations. The study shows that in most cases the locations match, in particular when considering the different significance of the location parameters, especially for the strongest events: the instrumental location provides the point where the rupture begins, while the macroseismic one is an estimate of the area where the earthquake possibly took place. This paper is not meant to discuss the importance and the necessity of macroseismic data; instead, the aim is to show that instrumental data can be used to obtain locations even for older seismic events, without any intention to define which location is better or more reliable. Mailing address: Dr. Stefano Solarino, INGV-CNT c/o Dipartimento Studio del Territorio e delle sue Risorse (DipTeRis), Università degli Studi di Genova, V.le Benedetto XV 5, 16132 Genova, Italy; e-mail: peter@dipteris.unige.it


Introduction
In this paper I will show that when enough seismic recordings exist it is possible to compute an instrumental location of major earthquakes.The computation is carried out on an area of midhigh seismicity (Garfagnana, Northern Tuscany, box with dashes in fig. 3) and could be conducted on any other area of similar characteristics; the results are used for a discussion on the role of the instrumental data of historical earthquakes.
Many efforts have been made in the recent past to recover the historical seismicity of the Italian Peninsula with the aim to distinguish hazardous areas through the appropriate algorithm.The goal of designing a map of hazard for the Italy has been an impressive incentive in recent decades and has led to the compilation of several parametric catalogues (Camassi and Stucchi, 1997;Gruppo di Lavoro CPTI, 1999;Boschi et al., 2000;Gruppo di lavoro CPTI, 2004).
These catalogues differ for both period of coverage and number of events included, but they are especially distinct in the philosophy that guided their compilation.CFT (Boschi et al., 2000) is based on macroseismic data and locations only, contains 605 seismic events selected on the basis of the minimum intensity threshold of VII and contains earthquakes which occurred in the period 461 B.C. to 1997. Conversely, NT4.1.1, (Camassi andStucchi, 1997) contains all earthquakes exceeding I o= =5/6 or Ms=4.0 in the period 1000-1980, and this sums up to 2421 records.The compilation was then enlarged to cover the period 1981-1992.The catalogue is mainly based on macroseismic data, a few of which are novel, but it also contains several locations whose parameters have been previously published in other catalogues; the macroseismic data for these events are not always available.In particular, one catalogue that has been extensively used for merging data in NT4.1.1 is that published by Postpischl (1985), which also contained instrumental locations.NT4.1.1 is probably the most flexible and widely used historical catalogue in Italy and is thus adopted in this work as a reference.In this frame, it is especially important to emphasize that the records extracted from other catalogues and included in NT4.1.1 are indicated as CP: this code actually means that it is no straightforward to determine where the parameters come from (macroseismic or instrumental location) and especially that there is no direct information of the quality of these parameters.In the next paragraphs I will show that this lack biases the comparison, since more than 50% of the resulting instrumental locations that do not match with the existing ones are marked CP.
Finally, CPTI is the result of a joint project to merge and recompute part of the mismatch-ing locations of the mentioned (CFT and NT) catalogues: it has been recently updated to version CPTI04 (Gruppo di Lavoro CPTI, 2004).
All these inventories cover almost the whole of human history and, being updated to very recent years, they overlap the period in which many seismological instruments were already operating.As an example, in fig. 1 I show the number of installed instruments versus years in the period 1900-1950 as derived from the results of the TROMOS (http://storing.ingv.it/tromos)project.The histogram does not show two important features: stations were almost homogeneously spread over the Italian Peninsula and a few of them never interrupted they recording activity, even if performed by updated and more modern instruments.Figure 3, even if only showing the position of the stations used in this work, gives a frame on the distribution of historical observatories.More information can be retrieved from the TROMOS web pages.
The existence of several seismic instruments provided good waveforms for the main earthquakes of the century; however, despite the richness of the database, an instrumental location for these earthquakes is not always computed and when a macroseismic one is available it is strongly preferred, even when based on a poor and or not optimal distribution of the localities.In theory, it is more correct to input, in whatever computation, data obtained with the same means or obtained applying similar methodology, that is in this case to always use macroseismic data for homogeneity.But catalogues are already «biased» by a few instrumental locations as discussed above, so it does not seem so difficult to use alternative locations especially if they prove to be of acceptable quality.The study is not meant to discuss the acknowledged importance and the necessity of macroseismic location; instead, the aim is to show that instrumental data can be used to obtain locations even for older seismic events, without any intention to define which location is better or more reliable.In fact, throughout the paper the corresponding macroseismic locations will be used as a reference, and the discussion will be based on the differences between epicentral position rather than on quality of data and result.
In particular, I introduce the basics of the method, I make a short description of the main problems encountered in assembling the data and finally I discuss and compare the obtained locations to try to find out which is the role of instrumental data for less recent earthquakes.

Towards an instrumental location: Joint Hypocentral Determination
The methodology applied in this work (Solarino et al., 1996) has been extensively described in Solarino (2002).It is based on the concept that it is possible to account for all systematic errors that may interfere with the location of an earthquake by computing corrective factors in conjunction with the model used for locations; all non-systematic errors may instead be discarded by checking the single datum.For example, site effects or non perfect knowledge of the position of the instrument can be included in the correction factor, which will bias all data with the same importance, while a picking error can be avoided by a careful check of the single datum, especially if original waveforms are still available.This part of the work proves to be the most delicate, due to the difficulties in finding original seismic bulletins, digitized or on paper waveforms, original scripts by seismologists running the instruments.In a few cases data that would have represented a potential source of error were finally recovered by going back to the original seismogram.
One of the most important aspects in this frame is that of the time synchronization.It must be taken into account that the technology of clocks was very similar for all seismic observatories, although provided by different brands, as happens today.It is then reasonable to believe that the performances (and thus the errors) of these timing device were very similar.Data synchronized with these machineries can then be considered comparable.
It is also well known that in the early days of seismology scientists used mechanical clocks, the drift of which was corrected at established intervals using more sophisticated radio timing systems, and ensuring a more accurate absolute time.In case of major earthquakes, the phase picking was corrected to take into account the drift.Sometimes this correction was directly annotated on the seismogram, sometimes it was annotated apart, but in general the published and shared data was already corrected for the drift.However, in some cases, different phase pickings are found on diverse bulletins, perhaps reporting corrected and non-corrected data respectively: in such cases either it is possible to realize which is the correct timing or the reading must be discarded.Finally, since many location programs use S-P times, for this kind of data the correctness of absolute time does not really matter provided that no malfunctioning occurs during the recording of the event.
While random errors can be avoided by careful control of the data, systematic errors can be included in the computation by applying a Joint Hypocentral Determination (JHD thereinafter) technique, the aim of which, as stated above, is to sum up all errors in a corrective factor.As known, the JHD uses a master event to check the compatibility of its location with the rest of the data.In this case, the JHD technique has been based on a group of very recent and well located events that acted as masters in successive steps.
The starting point of the whole routine consisted in the computation of a very reliable velocity propagation model using a dataset of recent events (01/1999-12/2001, fig.2); the computation of this propagation model, later indicated as «reference model», has been widely described by Ellsworth (1977) and Kissling (1988).The 1D model results from a combined, simultaneous inversion where the inverse problem is solved by the full inversion of the damped least squares matrix A T A+λ.As the inverse problem is non-linear, the solution is obtained iteratively, where one iteration consists of solving both the complete forward problem and complete inverse problem once.VELEST (Kissling et al., 1995), the code that performs the computation, can also be run in single event mode when only the location problem is carried out.In this case, an additional singular value decomposition of the symmetric matrix A T A is performed to calculate the eigenvalues.As known, this way of treating the inverse problem provides supplementary information on the resolution: they will not be discussed in detail in this paper, but supplemental information can be found in Solarino (2002).At the end of the computation, different models with about the same residual variance and location precision can usually be obtained from the same data set.The model that coincides best with the surface geology and a priori information on the nearsurface structure has to be chosen as the final model, and will be called «minimum» (Kis-   , 1988) to indicate that it provides minimum average (rms) values for all earthquakes used in the inversion.Additionally, relocation of known sources (shots, blasts) facilitates the choice of the best model and gives hints on its reliability.
The 1D model resulting from the computation with VELEST (fig.3, left panel) is accompanied by P and S station corrections that are first calculated only for the stations that are still operating and that contributed to the location of the master events, while those relative to seismic ob-Fig.3. Velocity propagation model (left panel) and position of stations used for this study.For each station, the correction factor is reported (circle for negative values, triangle for positive; size is proportional to the amount) as after the computation process resumed in table I.The distribution of delays is close to the optimal, being very nicely distinct the positive and negative areas (Kissling et al., 1995).The box with dashes shows the approximate extension of the Lunigiana-Garfagnana area.
Table I.Evolution of the computation of delays for the seismic observatories the readings of which are used in this work.At increasing steps the number of known stations corrections (in percentage with respect to the total number) expands, and even observatories that no longer operate can be taken into account.
At the end of this routine, the model is complete with all station delays (fig.3) and can be used for the location of the «historical» events of the Garfagnana area.

Data and results
Apart from the limitation of the availability of instrumental data, in theory no constraints act in selecting the data to be relocated.In practice there are several limitations that must be taken into account.
First, the events must be included in some other (macroseismic) catalogue for comparison.This is not going to be necessary in a future application, but it is of paramount importance to test the results, and cannot be neglected at this stage.Second, events must have at least 4 usable phase pickings that in some cases may mean 6 to 7 readings to be conservative enough to consider mistakes, picking errors and non-matching information.Third, there must exist more than one source for the data to cross-check the entries, as men-tioned in the last paragraph.By combining all these aspects, it turned out that the best selection should take into account events which occurred later than 1900, that are included in the NT4.1.1Catalogue (the most complete) and whose data are reported on two or more of the sources quoted in Appendix A. It is noteworthy that some of the references reported in Appendix A are compilations, and should then be comprehensive of all data available for an event as collected from the various bulletins and observatories.The careful cross checking between these compilations and some other original data showed that in a few cases the compilations contain inconsistencies and differences, and this is another important issue towards the use of more sources.
The initial selection of 56 seismic events (as extracted from NT4.1.1 for the area 43.6-44.6 and 9.1-10.5)was dramatically reduced to 30 when looking for the multiplicity of sources but the number dropped to 19 when considering the minimum number of readings required for locations upon careful selection of the blunder-free data (the compiled phase readings for these data are reported in Appendix B).Of course, this aspect diminishes the potential of the instrumental locations versus the macroseismic ones, it being possible in this case only for fewer than 40% of the events, and will be discussed in the final paragraph.The characteristics of the selected events are reported in table II (second column from left); their location properties are taken from the NT4.1.1Catalogue.
After checking for data completeness on all available sources, events have been located applying VELEST in single event mode; the final locations are reported in table II together with the relevant location errors, while table III shows a comparison of the obtained locations with the original positions.The new locations are compared with the macroseismic ones in fig.4, where it is possible to see at a glance how far the epicentres have been shifted: the vectors connecting each two points indicate the change in position from macroseismic to instrumental and show very different lengths for the diverse couples.The figure also shows that the original locations were sometimes very clustered (see for example events 08/1902, 1920, 1968, 1961, 1939).
The characteristics of the newly obtained locations are reported in table II: the depth, marked by bold characters, has been computed for more than 50% of the earthquakes.At first glance, the obtained distribution of earthquakes seems compatible with the knowledge of the recent seismicity of the area (Solarino et al., 2002) and with the proposed position for the main seismogenic structures of the area (e.g., the Garfagnana-north fault, Valensise and Pantosti, 2001); moreover, the less clustered aspect of the seismicity seems less «artificial» than the macroseismic one.
The errors associated with the new locations (ERX, ERY, ERZ) are on average less than 5 km, and the overall RMS is sometimes less than 1 s.The usage of the standard errors in the hypo format may appear inadequate or incomplete to describe the uncertainties of the final locations and, as shown in Solarino (2002), could be substituted by a less compact form by showing the full resolution matrix or its diagonal elements.
However, at this stage the errors are only used to distinguish between locatable and nonlocatable events and to give a rough idea about the final quality and reliability of the results and in such an attempt the standard errors are enough.Finally, the gap shows that for some events the station coverage is very good, while in a few cases it exceeds 200°.
The analysis of the differences between locations (table III, difference between latitudes and between longitudes in km) shows three different trends.There are seismic events (05/03/1902, 07/09/1920, 1931, 1939, 1951, 1959, 1972) for which the differences between the two location methods do not exceed 16 km for each coordi-929 Table II.Index of the events relocated in this work.The list is the final result of a selection done in the area 43.6-44.6 and 9.1-10.5 after checking for multiplicity of data sources and minimum number of phase pickings (see text).The locations parameters of the second column are those of the NT4.1.1 (Camassi and Stucchi, 1997) Catalogue.The remaining columns are relative to instrumental locations and report the position, the depth and the errors associated to the locations.NC is used for depths that are not adequately constrained.

Date
Location nate.The choice of this reference value is somewhat arbitrary, but it represents a good compromise between the errors associated with the instrumental location and the uncertainty associated with the macroseismic one.It is then reasonable to say that the two locations match, if one considers that the instrumental location gives the point of rupture while the macroseismic one is more an average of the area the rupture started from.This statement is especially true for events exceeding M 5.5 (Gasperini and Ferrari, 2000), but the most widely shared definition of the macroseismic epicenter («the baricentre of the area featuring the maximum effects», Gasperini and Ferrari, 2000) involves a similar bias for all magnitude levels, though of course in proportion to them, since the maximum effects may depend upon several factors (site effects, wrong or inaccurate description, distribution of the localities).On average, the number of points used for the macroseismic locations of these events is large, say greater than 30, except for a couple of them that directly come from other sources (CP).
A second group contains events that do not match: their differences not only exceed 16 km in one or both coordinates, but at least for one can be up to 80 km, which cannot be simply explained by the diverse meaning of the location parameters.These are the 08/1902, 1921, 1926, 1927, 07/1928, 08/1928, 1934, 1961, 1965, 1968, 1973 events.Almost all of them have been taken from other catalogues (CP), while those based on macroseismic observations have been located with 10 (1921), 13 (1928), 13 (1965), 29 (1934) and 39 (1927) localities.This second group is probably much more difficult to locate using macroseismic data with the currently available information; on average they have a lower magnitude (down to 3.7).Moreover, the original catalogue which contained those for which macroseismic data are not available (CP) was probably not accurate enough.
It is important to emphasize that, if the locations proposed in this work are reliable, some of these earthquakes actually occurred far from where they have been positioned, and this may have an impact on the microzonation and hazard computation.I discuss for example the event which occurred in 1934.Its original location has been obtained using 29 macroseismic observations (Monachesi and Stucchi, 1997), which is rather a good number, positioned as in fig. 5.It is evident that there is a very uneven distribution of localities and that they are especially lacking on the eastern side, which can have biased the macroseismic location (grey star) towards the western area.Conversely, the instrumental location (black star in the figure) moves the epicenter to the north-east, that is in the same direction where macroseismic data are most lacking.If I accept that the macroseismic location may be biased by the poor coverage on the eastern side and I take into account the location errors of the instrumental determination, all this would render plausible a shift of the lo-Table III.Absolute differences (in km) between the latitude and the longitude of the macroseismic and instrumental locations.Bold numbers are used when the differences exceeds 16 km for any of the values.The magnitude is taken from the NT4.1.1Catalogue.The last column of the table displays the number of points of intensity available for the macroseismic location.CP indicates the events extracted from other catalogues: for these events no quality information can be derived.

Date
Ms   cation of the earthquake towards the east, that would mean on the other side of the watershed of the Apennines.The third group is made up of one event only (1914): its location consistently differs from the macroseismic one but, the latter being obtained by more than 500 localities, it is reasonable to believe that instrumental data are not able to constrain the solution adequately.This is shown by both the location errors associated with this hypocenter (they are among the highest) and by the resolution matrix (not reported), whose diagonal elements are very far from optimal (0.9604, 0.9584, 0.9525 and 0.8055 respectively for origin time, X, Y and Z coordinates).
The presented analysis, although not exhaustive, gives a rough estimate of the potential and limitations of both data and method and thus represents a good starting point for the discussion on the function of instrumental locations.

Discussion
The obtained instrumental locations proved to be very similar, in many cases, to those proposed using macroseismic data.Moreover, it is always possible to define how reliable they are by looking at the associated location errors (within the limitations that these errors may introduce) or at the resolution matrices, and finally they may provide information on the depth of the events, which cannot be easily derived with macroseismic data, although several methodologies have been proposed to infer depth from intensity observations.Provided thus that, under particular conditions, it is possible to obtain an instrumental location for older earthquakes, it is important to define the limits of their usage, that is to define what their role is.
It is evident that instrumental locations cannot substitute macroseismic locations as they do not completely overlap the recent part of the parametric catalogues: instrumental data are largely available only since the sixties, while as stated above they lack for the first decades of the twentieth century, except for very large earthquakes.But instrumental data give a much better chance to distinguish different events in seismic sequences, which is not required in the macroseismic database (it is selected to avoid clusters that would bias many of the applications which it is used for) but may be helpful in other studies (seismic cycles, seismotectonics and so on).For example, the event of September 1920 was preceded the day before by an important shock that is not included in the published macroseimic catalogues but was recorded by many seismographic stations: how did this first shock affect the macroseismic observations of the following large event?Could this important event be neglected when studying the way energy is released in the area?In this frame, it is very important to have at least a rough estimate of the depth of the earthquake, which can emerge, sometimes with severe limitations, from instrumental locations.In fact several methods have been proposed to infer depth from the attenuation of macroseismic data, but seldom are they applied and in no Italian macroseismic catalogue is depth reported.
At the moment, the usage of instrumental data for the mentioned purposes is limited by the fact that seismic bulletins and waveforms for historical events are not in digital format, but this will soon change thanks to national (SISMOS, Michelini and the INGV SISMOS Group, 2004) and international (EUROSISMOS, Ferrari and Pino, 2004) projects that aim at compiling information for both instruments and seismic recordings.
The correct role of the instrumental locations is then to support and complement the macroseismic data and, in the few cases in which the macroseismic data are poor, to be used instead, provided that the instrumental data have been carefully checked.This should not sound too odd since at the moment the most widely used parametric catalogue already contains a small percentage of instrumental locations, and adding a few more events or some more information could only improve it.I owe special thanks to Alberto Ansaloni who provided very important information about the Observatory of Chiavari and let me thumb through the many bulletins, journals and scientific reports archived there.(1920,1927,1934) throughout the period di Venezia (formerly of operation (1817-1938) Osservatorio Meteorologico e Geodinamico)

Fig. 1 .
Fig. 1.Stations operating in Italy in the period 1900-1950 (source TROMOS).Since the project made an inventory of observatories established before 1940, the histogram is updated to that time.The last bars of the plot takes into account only the stations that still existed in 1950 and does not include those established in the decade 1940-1950.

Fig. 2 .
Fig. 2. Earthquakes used for the computation of the reference model applied in this work.They cover the period 1999-2001. sling

Fig. 5 .
Fig. 5. intensities for the 1934 event (DOM,Monachesi and Stucchi, 1997).The grey star shows the macroseismic location, the black one marks the instrumental one proposed in this study.

Fig. 4 .
Fig. 4. Vectors connecting the original macroseismic location with the instrumental one as obtained in this work.