Pure and Applied Geophysics

, Volume 170, Issue 9, pp 1385–1395

Tsunami Early Warning Within Five Minutes

Article

DOI: 10.1007/s00024-012-0512-6

Cite this article as:
Lomax, A. & Michelini, A. Pure Appl. Geophys. (2013) 170: 1385. doi:10.1007/s00024-012-0512-6

Abstract

Tsunamis are most destructive at near to regional distances, arriving within 20–30 min after a causative earthquake; effective early warning at these distances requires notification within 15 min or less. The size and impact of a tsunami also depend on sea floor displacement, which is related to the length, L, width, W, mean slip, D, and depth, z, of the earthquake rupture. Currently, the primary seismic discriminant for tsunami potential is the centroid-moment tensor magnitude, MwCMT, representing the product LWD and estimated via an indirect inversion procedure. However, the obtained MwCMT and the implied LWD value vary with rupture depth, earth model, and other factors, and are only available 20–30 min or more after an earthquake. The use of more direct discriminants for tsunami potential could avoid these problems and aid in effective early warning, especially for near to regional distances. Previously, we presented a direct procedure for rapid assessment of earthquake tsunami potential using two, simple measurements on P-wave seismograms—the predominant period on velocity records, Td, and the likelihood, T50Ex, that the high-frequency, apparent rupture-duration, T0, exceeds 50–55 s. We have shown that Td and T0 are related to the critical rupture parameters L, W, D, and z, and that either of the period–duration products TdT0 or TdT50Ex gives more information on tsunami impact and size than MwCMT, Mwp, and other currently used discriminants. These results imply that tsunami potential is not directly related to the product LWD from the “seismic” faulting model, as is assumed with the use of the MwCMT discriminant. Instead, information on rupture length, L, and depth, z, as provided by TdT0 or TdT50Ex, can constrain well the tsunami potential of an earthquake. We introduce here special treatment of the signal around the S arrival at close stations, a modified, real-time, Mwpd(RT) magnitude, and other procedures to enable early estimation of event parameters and tsunami discriminants. We show that with real-time data currently available in most regions of tsunami hazard, event locations, mb and Mwp magnitudes, and the direct, period–duration discriminant, TdT50Ex can be determined within 5 min after an earthquake occurs, and T0, TdT0, and Mwpd(RT) within approximately 10 min. This processing is implemented and running continuously in real-time within the Early-est earthquake monitor at INGV-Rome (http://early-est.rm.ingv.it). We also show that the difference mb − log10(TdT0) forms a rapid discriminant for slow, tsunami earthquakes. The rapid availability of these measurements can aid in faster and more reliable tsunami early warning for near to regional distances.

Keywords

Earthquake tsunami early warning earthquake location real-time seismology body waves 

1 Introduction

Tsunamis are most destructive at near to regional distances (i.e. <1,000 km) from the causative earthquake, arriving within 20–30 min after the earthquake origin time (OT); effective early warning at these distances requires notification within 15 min or less after OT (Tsushimaet al. 2011; Newmanet al. 2011; Madlazim2011). Currently, rapid assessment of the tsunami potential of an earthquake by organizations such as the Japan Meteorological Agency (JMA), the German–Indonesian tsunami early warning system (GITEWS), or the West Coast and Alaska (WCATWC) and Pacific (PTWC) Tsunami Warning Centers relies mainly on initial estimates of earthquake location, depth, and moment, M0, or the corresponding moment magnitude, Mw. For the regional scale, WCATWC and PTWC issue warning messages within approximately 5–10 min after OT for shallow, underwater events using the P-wave moment-magnitude discriminant, Mwp (Tsuboiet al. 1995, 1999) if Mwp ≥ 7.5 (e.g., Hirshorn and Weinstein2009).

M0 is of interest for tsunami warning because the efficiency of tsunami generation by a shallow earthquake depends on the amount of sea floor displaced, which can be related to a finite-faulting model expressed by the seismic potency, LWD, where L is the length, W the width, and D the mean slip of the earthquake rupture (Kanamori1972; Abe1973; Kajiura1981; Lay and Bilek2007; Polet and Kanamori2009). Then, because M0 = μLWD, where μ is the shear modulus at the source, the sea-floor displacement and thus tsunami potential should scale with LWD = M0/μ.

Mw is found to be a good discriminant for many past tsunamigenic earthquakes, but not all, especially not for slow “tsunami earthquakes” (TsE), which, by definition, cause larger tsunami waves than would be expected from their Mw (Kanamori1972; Satake2002; Polet and Kanamori2009; Newmanet al. 2011). The discrepancy for these earthquakes can be related to rupture at shallow depth where μ can be very low, an-elastic deformation, for example ploughing and uplift of sediments may occur, and the fault surface may be non-planar with splay faulting into the accretionary wedge (Fukao1979; Mooreet al. 2007; Lay and Bilek2007). Additionally, a reliable estimate of Mw for large earthquakes is usually provided by a centroid-moment tensor magnitude, MwCMT (Dziewonskiet al.1981; Ekströmet al.2005), which requires waveform inversion, varies with rupture depth, earth model, and other factors, and is only available 20–30 min or more after an earthquake (Hayeset al. 2011; Duputelet al. 2011). Thus rapid magnitude estimators such as Mwp are used for tsunami warning, but Mwp performs poorly compared with MwCMT and other discriminants for tsunami potential (Lomax and Michelini2011a, LM2011 hereinafter).

To avoid these problems and aid in effective early warning, especially for near to regional distances, we suggest the use of more direct and rapid discriminants for tsunami potential and for identification of TsEs. We have presented (Lomax and Michelini2009b; LM2011; Lomax and Michelini2011b) a direct procedure for rapid assessment of earthquake tsunami potential using two, simple measurements on P-wave seismograms—the predominant period on velocity records, Td, and the likelihood, T50Ex, that the high-frequency, apparent rupture-duration, T0, exceeds 50–55 s. T0 for large earthquakes is related primarily to rupture length, L, and both Td and T0 will increase as rupture depth, z, decreases, because of the effects of a reduced shear modulus and rupture velocity, vr. We show in LM2011 that either of the period–duration products TdT0 or TdT50Ex gives more information on tsunami impact and size than MwCMT, Mwp, Mwpd (Lomax and Michelini2009a, LM2009A hereinafter), and other currently used discriminants. These results imply that tsunami potential is not directly related to the product LWD from the “seismic” faulting model, as is assumed with the use of Mw discriminants, and suggest instead that information on rupture length and depth can constrain well the tsunami potential of an earthquake. This information on rupture length and depth is provided by TdT0 and TdT50Ex, in which case explicit estimates of rupture length and depth, which are difficult or impossible to determine rapidly, are not required.

Effectively, as shown schematically in Fig. 1, for a fault of fixed potency LWD, as rupture depth decreases the quantity TdT0 increases, reflecting the increased tsunami potential of the shallowest, underwater earthquakes, including TsEs. This behaviour suggests that the TdT0 discriminant captures the “tsunami” faulting model corresponding to the observed tsunami waves (Satake1994), as opposed to the “seismic” faulting model as given by M0. In this case TdT0 may also be a valuable aid in defining the finite-faulting description of the source and sea floor displacement for real-time tsunami forecasting.
Fig. 1

Simplified diagram of a subduction zone mega-thrust showing two interplate thrust ruptures 1 and 2 with the same seismic potency LWD (light grey patches) but different vertical seafloor displacement (uplift areas shown in dark grey). The long, shallow rupture 1 results in greater total seafloor uplift than the deeper rupture 2. Because M0 = μLWD and μ increases with depth, M0, the “seismic” faulting model, will be smaller for rupture 1 than for rupture 2. In contrast, because L1 > L2, and the rupture velocity, vr, is lower at shallow depths, T0 ∝ L/vr will be larger for rupture 1 than for rupture 2. Because Td may give additional information on z, W, or D, the quantities TdT0 and TdT50Ex can be larger or much larger for rupture 1 than for rupture 2, reflecting the “tsunami” faulting model and correctly identifying the greater seafloor uplift and tsunami potential of the long, shallow rupture 1

We introduce here special treatment of the signal around the S arrival at close stations, a modified, real-time, Mwpd magnitude (Mwpd(RT)), and other procedures to enable early estimation of multiple event parameters. We show that, with current, real-time seismogram data, rapid calculation of the direct, period–duration discriminant, TdT50Ex, can be completed within 5 min after OT, and calculation of T0, TdT0, and Mwpd(RT) magnitudes within approximately 10 min. We also show that the difference mb − log10(TdT0) is a useful, rapid discriminant for TsEs. We illustrate that the rapid availability of these measurements can aid in faster and more reliable tsunami early warning for near to regional distances.

2 Rapid, Direct Assessment of Tsunami Potential

To rapidly assess T0 for an earthquake we use the duration-exceedance (DE) procedure of Lomax and Michelini (2009b) which, by analysis of high-frequency (HF), vertical-component seismograms, determines whether T0 for an earthquake is likely to exceed 50–55 s (Fig. 2). On the 1–5 Hz band-pass filtered seismogram we form the ratio of the rms amplitude, A50, from 50 to 60 s after the P arrival time, TP, with the rms amplitude, A25, for the first 25 s after TP to obtain a station DE level, l50 = A50/A25, available approximately 60 s after TP. We define an event DE level, T50Ex, as the median of the station l50 values. If T50Ex is greater (less) than 1.0, then T0 is likely (unlikely) to exceed 50–55 s.
Fig. 2

Schematic diagram of single-station, period–duration processing for the 2010.10.25, Mw7.8, Mentawai earthquake at station MS.BTDF at 6° GCD. Top trace raw, broadband velocity with Td period estimation. Middle trace HF seismogram showing estimation of the station DE level, l50 = A50/A25. Bottom trace HF seismogram showing estimation of T0 when T0raw ends after TS

We define the dominant period, Td, for an event as the median of the dominant period values for each station given by the peak of the τc algorithm (Nakamura1988; Wu and Kanamori2005) applied with a 5 s sliding time-window from 0 to 55 s after the P arrival on velocity seismograms (Fig. 2; LM2011). Then, taking T50Ex as a substitute for T0, the TdT0 discriminant for tsunami potential becomes TdT50Ex (i.e., TdT50Ex ≥ 8.0 s).

To enable calculation of TdT50Ex, T0, mb, Mwp, and Mwpd within 5 min after an earthquake occurs we modify the procedures described in LM2011, including reduction of the minimum station distance to 5° for all measurements, and special treatment of the signal around the S arrival time, TS, at close stations:
  1. 1.

    For distances less than 30° significant S signal may remain on the HF seismograms (LM2009A). Consequently, if the raw, HF duration (T0raw) determined following LM2011 ends later than (TS − TP)/4 after TS, we assume that the HF signal after TS is primarily S wave energy, and that a best estimate of T0 is measured starting at TS. That is, if T0raw > (TS − TP) + (TS − TP)/4 then T0 = T0raw − (TS − TP), (Fig. 2).

     
  2. 2.

    To prevent including S energy in the magnitude estimates, the Mwp and Mwpd analysis windows are truncated at TS.

     
We also modify Mwpd from the formulation of LM2009A to enable simple and robust real-time application without event type determination by:
  1. 1.

    using the Mwp constant 4πρα3r/FP from Tusboi et al. (1995, 1999), where ρ is density, α is P velocity, r is station-source distance, and FP a radiation pattern correction;

     
  2. 2.

    applying the depth correction but no other geometrical or attenuation corrections from LM2009A; and

     
  3. 3.

    applying the moment correction of LM2009A to all event types if T0 > 90 s; to prevent discontinuous jumps in magnitude, this correction is applied using a linear ramp, weight factor from 0 at 90 s to 1.0 at and above 110 s.

     

We identify this magnitude as Mwpd(RT).

We also calculate mb on simulated WWSSN-SP velocity traces, following the Vmax procedure of Bormann and Saul (2008), we identify this magnitude as mb(Vmax).

The processing procedures discussed here are implemented and running continuously in real-time within the Early-est software and earthquake monitor at INGV-Rome (http://early-est.rm.ingv.it), including rapid calculation of event locations, TdT50Ex and mb(Vmax) and Mwp magnitudes within approximately 5 min, and T0, TdT0, Mwpd(RT) magnitudes and P first-motion, fault-plane solutions within approximately 10 min.

3 Application to Recent Large Earthquakes

We consider a reference set of 120 large earthquakes (6.4 ≥ Mw ≥ 9.1; 103 shallow, under water) since 1992, when high-quality, broadband seismograms became widely available, along with measurements of the impact and size of any generated tsunamis (Table S1). This reference set includes most tsunamigenic earthquakes listed in the NOAA/WDC Historical Tsunami Database (http://www.ngdc.noaa.gov/hazard/tsu_db.shtml), most events of Mw ≥ 7 in the past few years, and several events of regional importance.

As a measure of tsunami impact we define an approximate measure of tsunami importance, It, for the reference earthquakes based on 0–4 descriptive indices of tsunami effects and maximum water height h in metres from the NOAA/WDC database (see LM2011 for details). It is approximate because it is highly dependent on the instrumentation available, coastal bathymetry, and population density in the event region. It ≥ 2 corresponds approximately to the JMA threshold for issuing a “Tsunami Warning”; the largest or most devastating tsunamis typically have It ≥ 10.

Using the Early-est software and off-line event data from available stations, we determine event locations, mb(Vmax), Mwp, Mwpd(RT), T0, TdT0, and TdT50Ex for the reference earthquakes at OT + 5 min to simulate the information available within 5 min after an earthquake occurs, and at OT + 15 min to simulate the near final values of the event parameters. These delay times do not include real-time data transmission and processing latencies, which are currently typically less than 1 min. Figure 3 shows a comparison of MwCMT, Mwp, and TdT0 with It for the reference earthquakes at OT + 15 min; Table 1 shows summary results for all discriminants at OT + 15 min or later (MwCMT), and for T50Ex and TdT50Ex at OT + 5 min; Table S1 lists complete results for all events at OT + 15 min. Figure 3 also shows that Mwpd(RT) compares well with final MwCMT for the earthquakes studied, justifying the real-time modifications to Mwpd presented above.
Fig. 3

Processing results for the studied events. aMwpd(RT), modified for real-time application and evaluated at OT + 15 min, compared with final MwCMT. bd Comparison of tsunami importance, It, with b final MwCMT, cMwp at OT + 15 min, and dTdT0 at OT + 15 min. Verticalgrey lines show the target It ≥ 2 threshold; horizontal grey lines show the critical values for each discriminant (Table 1). The TdT0 axis uses logarithmic scaling. Event labels show earthquake type for non-interplate-thrust events with It ≥ 2 (I interplate-thrust, T tsunami earthquake, O outer-rise intraplate, B back-arc intraplate, U upper-plate intraplate, So strike-slip oceanic, S strike-slip continental, R reverse-faulting)

Table 1

Tsunami discrimination results

Discriminant

Available (min after OT)

Critical value

Correctly identified

Missed

False

It ≥ 2

%a

It < 2

It ≥ 2

It < 2

MwCMT(final)b

20+

7.45

34

68

41

16

13

Results at OT + 15 minb

 Mwp

3–10

7.45

22

44

45

28

9

 Mwpd (RT)

6–10

7.45

38

76

37

12

17

 T0

6–10

55

34

68

41

16

13

 TdT0

6–10

510

36

72

44

14

10

 T50Ex

4–8

1.0

31

62

48

19

6

 TdT50Ex

4–8

8.0

37

74

45

13

9

Results at OT + 5 minc

 T50Ex

4–8

1.0

17

81

20

4

9

 TdT50Ex

4–8

8.0

17

81

20

4

9

aPercentage of all events with It ≥ 2 that are correctly identified

b104 events classified: 50 have It ≥ 2

c50 events classified; 21 have It ≥ 2

For the reference events, the period–duration discriminant TdT0 shows a better correspondence with It than either MwCMT or Mwp (Fig. 3). At OT + 15 min both TdT0 and TdT50Ex correctly identify 72–74 % of tsunamigenic events with It ≥ 2 (Table 1), more than the Mwp(44 %), MwCMT (68 %), and T0 (68 %) discriminants, with fewer false positive identifications of events with It < 2. All discriminants are poor at identifying the tsunami potential of oceanic, strike-slip events, and give mixed results for back-arc intraplate earthquakes (see LM2011 for details). At OT + 5 min, with fewer available events (because of lack of data, mainly for earlier events), correct classification of events by the TdT50Ex discriminant is approximately the same as at OT + 15 min. These results show that use of the TdT50Ex discriminant for tsunami potential at OT + 5 min and perhaps earlier should be reliable and useful, especially as more real-time stations become available in regions of high tsunami hazard.

In LM2011 the tsunami discriminants are also compared using a physical measure of tsunami size—a tsunami wave amplitude at 100 km distance from the source, At, estimated from water height data in the NOAA/WDC database. As above for tsunami impact, It, comparisons in LM2011 show that TdT0 or TdT50Ex give more information on tsunami size, At, than MwCMT, Mwp and other currently used discriminants (Table S1). Moreover, there is indication of a linear, and thus possibly a physical, relationship between log(TdT0) and At.

The results for the reference events also show that the difference mb − log10(TdT0) < ~3.2 is a useful discriminant for TsEs (Fig. 4; Table S1). This difference relationship is related to the energy-to-moment parameter, Θ, (Newman and Okal1998; Weinstein and Okal2005) and to recent, rapid, energy-to-duration and moment-to-duration discriminants for TsEs (Lomaxet al.2007; LM2009A; Newmanet al. 2011). In addition to previously identified TsEs, preliminary results for this discriminant suggest a strong component of TsE for other events, including the Mw 7.0, 1998.07.17 Papua New Guinea earthquake and the Mw 8.1, 2007.04.01 and Mw 7.0, 2010.01.03 Solomon Islands earthquakes, all of which were strongly tsunamigenic (Table S1). Because the relationship mb − log10(TdT0) can be evaluated rapidly (mb(Vmax) is available a few minutes after OT and TdT0 is typically available before OT + 10 min), further investigation of this difference relationship is warranted, perhaps with mb replaced by a mean of HF amplitude, or an energy integral, from TP to TP + T0.
Fig. 4

mb(Vmax) compared with TdT0, both evaluated at OT + 15 min (see also Table S1). The TdT0 axis uses logarithmic scaling. The horizontal grey line shows the critical value for the TdT0 discriminant (Table 1); the diagonal line shows the constant difference mb − log10(TdT0) = 3.2

4 Example Timelines for Assessment of Tsunami Potential

We simulate and examine detailed timelines for event characterisation and assessment of tsunami potential using the Early-est software and off-line event data for several events: an earthquake that produced a mild tsunami, 2009.03.19, MwCMT 7.6, Tonga (Fig. 5), and two earthquakes that produced large tsunamis, the 2011.03.11, MwCMT 9.1, Tohoku, Honshu, Japan mega-quake (Fig. 6) and the 2010.10.25, MwCMT 7.8, Mentawai, Sumatra, Indonesia TsE (Fig. 7).
Fig. 5

Timeline of event parameter determination for the 2009.03.19, MwCMT 7.6, Tonga earthquake using the Early-est software and off-line event data. The top panel shows (black curve) epicentre location error relative to the final epicentre and (grey curve) the number of stations used for location. The remaining panels show the main magnitude and tsunamigenic discriminants discussed in this paper. Horizontal grey lines show the critical values for the TdT50Ex, T0 and TdT0 discriminants (Table 1)

Fig. 6

Timeline of event parameter determination for the 2011.03.11, MwCMT9.1, Tohoku, Japan earthquake using the Early-est software and off-line event data. Plot elements as in Fig. 5

Fig. 7

Timeline of event parameter determination for the 2010.10.25, MwCMT7.8, Mentawai earthquake using the Early-est software and off-line event data. Plot elements as in Fig. 5

For the 2009 Tonga event the PTWC (2009) issued a tsunami warning at OT + 12 min based on a magnitude measure of M 7.7; the tsunami warning was cancelled 1.5 h later. This earthquake generated a mild, non-destructive tsunami with observed water heights mostly less than 0.1 m (NOAA/WDC database). For this event (Fig. 5), Early-est determines at OT + 4 → 5 min the epicentre with an error of approximately 35 km, along with Mwp 7.6 and at OT + 7 min Mwpd(RT) 7.7, both in good agreement with the final MwCMT. At OT + 5 → 7 min, all three discriminants for tsunami potential, TdT50Ex, T0, and TdT0 are available and indicate a low tsunami potential for this event. All Early-est measurements have stabilized to near-final values within OT + 8 → 10 min.

For the devastating 2011 Tohuku earthquake and tsunami JMA first issued a tsunami warning at approximately OT + 3 min, but a near-final Mw and the true size of this event were not determined until OT + 20 → 30 min (Ozaki2011; Hayeset al. 2011). For this earthquake (Fig. 6), Early-est determines the epicentre at OT + 2 → 3 min with an error of approximately 60 km, Mwp 8.0–8.3 at OT + 4 → 6 min, and Mwpd(RT) 9.0–9.2 at OT + 6 → 8 min. These epicentre and Mwp results compare favourably with the JMA and NEIC response timelines (Ozaki2011; Hayeset al. 2011), and the near-final Mwpd(RT) determination precedes the earliest MwCMT estimation (NEIC Mww) by approximately 12 min. The TdT50Ex discriminant for tsunami potential is available at OT + 4 → 5 min and the T0 and TdT0 measurements are available at OT + 6 → 7 min; all three of these discriminants indicate a very high likelihood that a tsunami was generated. Almost all Early-est measurements have stabilized to near-final values within OT + 7 → 8 min, one exception being the epicentre, which only stabilizes at OT + 9 → 10 min because of the lack of station coverage off-shore to the east of the epicentre. For this event it is notable that the Mwpd(RT) 9.2 magnitude and T0 = 160 s duration estimates provide early information about the true size and extent of the earthquake rupture and tsunami source; these rapid measurements should be useful during future large earthquakes not only for tsunami warning, but also for early shake-map, finite-fault, and tsunami forecast modelling.

The 2010 Mentawai earthquake generated a large and destructive, local tsunami and has been identified as a TsE (Newmanet al. 2011; Madlazim2011). The timeline for Early-est characterization of this event (Fig. 7) is similar to that for 2011 Tohoku, except that the Mwp, Mwpd(RT), and TdT50Ex values are available approximately 1 min earlier, at OT + 3 → 4 min, and the epicentre is better constrained in the first minutes. These differences are mainly because of denser station coverage near this event, though there is still a lack of stations off-shore to the southwest of the epicentre. All three discriminant for tsunami potential, TdT50Ex at OT + 3 → 4 min, and T0 and TdT0 at OT + 5 → 6 min, indicate a high likelihood that a tsunami was generated. All measurements have stabilized to near their final values within OT + 7 → 8 min, although the large, early values for TdT50Ex indicate that the determination of this measure for near stations might be improved. For this event, in addition to early indication of high tsunami potential at OT + 3 → 6 min, Early-est gives at OT + 6 → 9 min a stable Mwpd(RT) 7.8–7.9 that matches final MwCMT, and mb − log10(TdT0) ≈ 3.0, suggesting this event is a TsE.

5 Conclusions

We have presented rapid determination of tsunami potential by use of two direct and simple measurements on P-wave seismograms, the predominant period, Td, and the likelihood, T50Ex, that the T0, exceeds 50–55 s (LM2011). We have also introduced a modified, real-time, Mwpd(RT) magnitude, and special treatment of the signal around the S arrival at close stations to enable early estimation of all event parameters. We find that either of the period–duration products TdT0 or TdT50Ex give more information on tsunami impact and size than MwCMT, Mwp, and other currently used discriminants. This result follows from T0 and Td being most sensitive to rupture length, L, and depth, z, which determine total seafloor uplift and tsunami potential.

We show that the TdT50Ex discriminant can be obtained within 5 min after an earthquake occurs with real-time data currently available in most regions of tsunami hazard. We also show that other critical event parameters can be obtained within 5–10 min, including: the T0 and TdT0 tsunami potential discriminants, an Mwpd(RT) that matches closely final MwCMT, and the difference mb − log10(TdT0) which is a rapid discriminant for slow, tsunami earthquakes. The rapid availability of these direct and simple measurements can aid faster and more reliable tsunami early warning for near to regional distances.

Acknowledgments

We thank two reviewers for critical comments that greatly improved the clarity of the manuscript. This work is supported by INGV—Centro Nazionale Terremoti institutional funds and by the EC n.262330 NERA 2010-2014 project. The IRIS DMC (http://www.iris.edu) and GFZ Data Archive (http://geofon.gfz-potsdam.de) provided access to waveforms used in this study.

Supplementary material

24_2012_512_MOESM1_ESM.xls (63 kb)
Supplementary material 1 (XLS 63 kb)

Copyright information

© Springer Basel AG 2012

Authors and Affiliations

  1. 1.ALomax ScientificMouans-SartouxFrance
  2. 2.Istituto Nazionale di Geofisica e Vulcanologia (INGV)Centro Nazionale TerremotiRomeItaly

Personalised recommendations