Notice: Wiley Online Library will be unavailable on Saturday 27th February from 09:00-14:00 GMT / 04:00-09:00 EST / 17:00-22:00 SGT for essential maintenance. Apologies for the inconvenience.
 At 23:41 UTC on 15 August 2007, an offshore earthquake of magnitude 8.0 severely damaged central Peru and generated a tsunami. Severe shaking by the earthquake collapsed buildings throughout the region and caused 514 fatalities. The tsunami resulted in three casualties and a representative maximum runup height of ∼7 m in the near field. The first real-time tsunami data available came from a deep-ocean tsunami detection buoy within 1 hour of tsunami generation. These tsunami data were used to produce initial experimental forecasts within 2 hours of tsunami generation. The far-field forecasts indicated that the tsunami would not flood any of the 14 U.S. communities. Comparison with real-time tide gage data showed very accurate forecasts.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 The 2004 Indian Ocean tsunami has been a major driving force in accelerating the implementation of a tsunami forecast system at the National Oceanic and Atmospheric Administration's (NOAA) Tsunami Warning Centers (TWCs) [Lautenbacher, 2005; Synolakis et al., 2005; Bernard et al., 2006; Geist et al., 2006; Synolakis and Bernard, 2006]. After the 26 December 2004 tsunami, the U.S. expanded the role of the National Tsunami Hazard Mitigation Program to implement the recommendations of the National Science and Technology Council  to enhance tsunami forecast and warning capabilities along the U.S. coastlines. Toward these goals, the NOAA Center for Tsunami Research (NCTR) is developing a forecasting system to use real-time deep-ocean tsunami measurements [González et al., 2005] ingested into site-specific tsunami inundation models to produce timely, accurate tsunami forecasts for potentially vulnerable U.S. coastal communities.
 Based on Titov et al.'s  methodology, NOAA's tsunami forecast system uses pre-computed models of tsunami propagation that match deep-ocean tsunami measurements to generate offshore tsunami characteristics. The offshore scenario is then used as the initial and boundary conditions for the high-resolution Standby Inundation Models (SIMs). These forecast models are an optimized version of the Method of Splitting Tsunami (MOST) [Titov and González, 1997], which has been extensively validated and verified based on Synolakis et al. . SIMs for specific communities have been carefully developed, tested, and validated by known historical data, and designed to provide at least 4 hours of coastal tsunami simulation in less than 10 minutes of real-time computation (L. Tang, et al., Developing tsunami forecast inundation models for Hawaii: procedures and testing, submitted to NOAA Technical Memorandum, Office of Oceanic and Atmospheric Research, Pacific Marine Environmental Laboratory, 2008). The key component of the forecast system is the data assimilation method that inverts the real-time deep-ocean tsunami measurements from DART buoys to constrain the tsunami source, which is then used to provide the propagation forecast (pre-computed) and the coastal inundation forecast (real-time). The application of this approach in the 17 November 2003 Rat Island tsunami, the debut event of the newly developed DART system, for the first time obtained real-time tsunami model predictions [Titov et al., 2005; Bernard and Titov, 2007], and paved the way for developing the more comprehensive and time-efficient tsunami forecast methodology described here.
 Presently, a system of 29 DART buoys (27 U.S.-, 1 Chilean-, and 1 Australian-owned) is monitoring tsunami activity in the Pacific Ocean (Figure 1). The pre-computed propagation models currently have 884 scenarios to cover Pacific tsunami sources, and the high-resolution forecast models are now set up for 26 U.S. coastal communities. The fully implemented system will use real-time data from the DART network (39 U.S. buoys by 2008) to provide high-resolution tsunami forecasts for at least 75 communities in the U.S. (by 2013) [Titov, 2008].
 Since the 2003 Rat Island tsunami, NOAA's forecast system has produced experimental forecasts of far-field tsunami impact for four more tsunamis (Tonga, 2006; Kuril, 2006; Kuril, 2007; and Solomon, 2007) and proven its accuracy, efficiency, and reliability. The 2007 Peruvian tsunami provided the sixth test.
 The coast of Peru is located at the boundary between the Nazca and South American Plates. Large earthquakes in the region have characteristically triggered destructive tsunamis in both the near- and far-fields. Okal et al.  reviewed the 17 largest tsunamigenic earthquakes along the central and southern coasts of Peru in the past 350 years. Okal et al.  reported 22 fatalities and 52 missing attributed to the 23 June 2001 Camaná tsunami, with runup of 7 m at Camaná. The 21 February 1996 Chimbote tsunami caused by a slow rupture, typical “tsunami earthquake” [Kanamori, 1972], was surveyed by Bourgeois et al. , who reported runup of ∼5 m, 12 deaths, and damage to many houses/beach huts at Chimbote. Okal et al.  evaluated the effect of tsunamis in Pisco, Peru. They concluded that Pisco could expect tsunami runup of a few meters (substantial damage) with a return period of 53 years and runup greater than 10 m (catastrophic destruction of infrastructures) with a return period of 140 years. They also noted that the last such catastrophic event occurred nearly 140 years ago.
 On 15 August 2007, at 23:41 UTC, a massive earthquake of moment magnitude (Mw) 8.0 struck off the Pacific coast of Central Peru. The earthquake (76.509°W, 13.354°S) was offshore about 150 km southeast of Lima at a focal depth of 39 km (USGS). This earthquake caused severe shaking and damage in nearby towns, especially in Pisco and Ica. The most recent estimate of the death toll inferred from news sources is about 514, forcing the government to declare a state of emergency. Coastal flooding induced by tsunami waves as high as 5 m was observed near Paracas (S. Tubbesing, personal communication, 2007). Pictures taken 4 hours after the earthquake showed that the tsunami resulted in minor flooding in La Punta, Callao (W. Power, personal communication, 2007). Fritz et al.  surveyed the area and measured inundation of 2 km at Lagunilla, where three bodies were found about 1.8 km inland. They also measured maximum runup of 10 m at Playa Yumaque, with representative runup of 7 m.
3. Event Timeline
 We illustrate the timeline of the experimental real-time forecast with respect to the Pacific Tsunami Warning Center (PTWC) bulletin series during the 15 August 2007 Peruvian tsunami in Figure 2. Twelve minutes after the earthquake, PTWC disseminated its first information bulletin reporting the earthquake location with a preliminary magnitude Mw 7.5. Potential tsunami threats were advised, but limited to coasts near the epicenter. Twenty six minutes later, at 00:19 UTC on 16 August, PTWC upgraded the earthquake magnitude to Mw 7.9 and issued a regional tsunami warning to the entire Pacific coast of South America, as well as a tsunami watch to the Pacific coast of Central America. Based upon the updated seismic magnitude, a tsunami energy projection map similar to Figure 1 was produced at 00:27 UTC using pre-computed sources. This map reflects the forecast maximum wave amplitude at each computational grid.
 The Chilean-owned DART 32401 was automatically triggered by the seismic waves 4 minutes after the earthquake. The clear evidence of tsunami generation was the first 6-cm-high pulse measured at DART 32401, located approximately 700 km southeast of the epicenter, 56 minutes after the main shock (Figure 3a). In response to this “tsunami signal on the deep-ocean gauge off northern Chile,” PTWC issued the third bulletin at 01:21 UTC to supplement the tsunami warning and watch published earlier, and to advise for potential tsunami risks at distant Pacific coasts. Real-time deep-ocean measurements free of coastal contamination are required as an input to the forecast system [Titov et al., 2005]. At this point, the DART measurements provided the arrival time, amplitude, and half wave period, which is the minimum information required for the data assimilation and inversion to constrain an initial tsunami source. The inversion algorithm assumes the linearity of the tsunami propagation in the open ocean. The tsunami waveforms recorded at DART buoys are approximated as a linear superposition of tsunami source functions computed from each unit source through the least squares method. During an event, a set of unit sources is chosen as the initial condition for the inversion based on earthquake epicenter location and the seismic moment.
 During this experimental forecast, the data were being inverted offline manually, therefore with substantial time delay. The fully automated procedure will shorten the inversion time considerably. At 01:50 UTC, 2 hours 9 minutes after tsunami generation, a preliminary tsunami source was obtained based on scaling a single unit source (A9 in Figure 1 inset) to fit the amplitude, the arrival time, and the half wave period of the measurement at DART 32401. This DART-constrained source was immediately used to produce a new tsunami energy projection map and initial assessment of the tsunami impact, while the formal inversion was being computed. The forecasted tsunami energy projection indicated large tsunami amplitudes off the Hawaiian Islands. Harbors in California were also concerned because they were the first coasts in the U.S. to be affected. Note that a tsunami with small far-field effects, such as the 15 November 2006 Kuril Island tsunami, could cause significant damage at the harbor of Crescent City due to the harbor resonances [Kelly et al., 2006; Uslu et al., 2007]. The final stage of the tsunami forecast was to provide high-resolution assessment of coastal tsunami impact for these U.S. forecast sites while the tsunami threats were hours away. The high-resolution forecasts based on the preliminary tsunami source indicated no tsunami flooding for any of the sites. The propagation snapshots in Figure 2 show that, at the time of the forecast, the influence of the tsunami was still limited to the Pacific coast of Peru. At 02:09 UTC, 2 hours 28 minutes after the initial earthquake, PTWC sent out the final statement to cancel the tsunami warning and watch.
 The tsunami waves would take at least another 8 hours to arrive at distant Pacific coasts. Four hours and 45 minutes after the tsunami generation, a model tsunami source was computed based on the formal data inversion of the measurements on DART 32401. Six unit sources (A8-10 and B8-10 in Figure 1 inset), a 300 km × 100 km fault area, were used as an initial condition for this inversion. The synthetic tsunami waveform was re-sampled at 15 seconds, given that the DART measurements in an event mode are sampled every 15 seconds. Although DART buoys 32411, 46412, and 51406 (Figure 1) also recorded the passing tsunami (Figure 3a), they were not used to constrain the tsunami source during the forecast because of the late tsunami arrival at these buoys. In the 2006 Tonga, 2006 Kuril, 2007 Kuril, and 2007 Solomon events, the inversion based on one DART station led to excellent forecasts at other stations [Titov, 2008], suggesting the inversion using data from just one DART produces robust and meaningful results. Out of six unit sources, the inversion of the Peru event only yielded slips on A9 and B9 (Figure 1 inset), 4.3 m and 4.1 m, respectively, giving a 100 km × 100 km tsunami source area and a seismic moment of 1.68 × 1021 Nm (Mw 8.1). Based on the updated tsunami source, the final forecasts were provided for 14 U.S. coastal communities in Hawaii, the U.S. West Coast, and Alaska (Figure 3b) — including the final tsunami energy projection in the Pacific (Figure 1) — while the tsunami was 6 hours away from arriving at the nearest U.S. coastline.
4. Discussion of Forecast Results
 The comparison of the wave amplitude in Figure 3a indicates that the forecast provided an excellent fit with observations at DART 32401 for the first several waves. The late waves at DART 32401 present better comparisons for waves with long period (∼30 minutes) than those with short period (<15 minutes), suggesting that the poor coastal bathymetry may be responsible for the inaccurate computation for the reflected and scattered waves arriving later at DART stations. The good agreement at the other three DARTs verified that the forecast estimated the tsunami source reasonably well. One noticeable result in Figure 3a is the approximately 12-minute time discrepancies of the arrival of the first wave at DARTs 32411, 46412, and 51406. These time discrepancies are possibly due to multiple causes, i.e., the errors induced by the ocean bathymetry, model approximation, and the location of the pre-computed tsunami source functions, which are offshore of the earthquake epicenter (Figure 1, inset). Extra unit sources are being incorporated into the pre-computed database to account for any future events in this area.
Figure 1 shows the basin-wide tsunami energy projection and the travel-time contours computed from the refined DART-constrained tsunami source. It is clear that the tsunami energy was mostly directed to the south and west of the Pacific coast of South America. Figure 1 indicates that the coasts of Ecuador, Panama, and countries in Central America were in the shadow zone of the tsunami energy and much less affected than Peru and Chile. The islands in the south Pacific, however, faced more serious threats. Figure 1 predicted high tsunami waves in the Chatham Islands, east offshore of New Zealand, and was later confirmed by the peak-to-trough heights of 53 cm and 46 cm observed at tide gages in Waitangi (S. Weinstein, personal communication, 2007) and Kaingaroa (R. Bell, personal communication, 2007), respectively. Figure 1 also indicated higher tsunami energy at several locations on the east and north coasts of New Zealand that were verified by tide-gage measurements, including the peak-to-trough height of 54 cm at Sumner Head (R. Bell, personal communication, 2007). The energy distribution shows minor tsunami impact along the east and north coasts of Australia, where were protected by New Zealand and the island chain to its north. The observed maximum peak-to-trough height there was about 20 cm. The tsunami brought waves of 20–25 cm high to the Japanese coasts after a more than 20-hour propagation, causing Japanese emergency agencies to issue tsunami warnings to the coastal communities. We note that most of the far-field tsunami measurements show leading-elevation N wave [Tadepalli and Synolakis, 1996], while eyewitnesses reported leading depression locally [Fritz et al., 2007].
 Coastal inundation was computed in real time during the event for four harbors in Hawaii, six in the U.S. West Coast, and four in Alaska using SIMs with grid resolution of 50–60 m. Figure 3b shows the forecast results compared with observations for the 14 tide gages. Since all the modeling forecasts were completed before the tsunami arrived, Figure 3b provides additional evidence of the accuracy and efficiency of the forecast system. Among the 14 harbors, Hilo, HI recorded the largest maximum wave height, 67 cm, Kahului, HI recorded 56 cm, and Honolulu, HI recorded 10.5 cm, while Crescent City, CA recorded 30.7 cm. The others had wave heights less than 10 cm that were in the range of the background noise. All 14 forecast results showed earlier arrivals around 12 minutes, which is about the same discrepancy noted at distant DART locations. A 12-minute difference out of 12 hours represents a less than 2% error. After this 12-minute time difference was adjusted in Figure 3b, the model results and observations matched very well in both wave height and period at all tide stations. Most notable are Hilo, HI and Honolulu, HI, where the modeling results and observations are in excellent agreement for up to 24 hours after tsunami generation. The discrepancies of the late waves at the Kahului and Crescent City tide gages were mostly induced by a mixture of causes, such as inter-island trapped waves, edge waves [González et al., 1995], complicated bathymetry near the station, as well as harbor resonance. Higher-resolution computational grids are possibly needed to resolve some of these complexities [Uslu et al., 2007]. Table 1 shows the correlation between forecast and observation for four tide gages where the observed maximum wave height is greater than 10 cm. The noise level is represented by the ratio of the root-mean-square amplitude of noise to that of the tsunami signal over a 4-hour time series, and exhibits dominant interference with the tide gage measurements when tsunami height is small. The estimated errors of the forecast are largely within the noise level and indicated a very accurate forecast of the maximum tsunami height at all four tide gages.
Table 1. Comparisons of Error Estimation of the Real-Time Forecast With Noise Level at Tide Gages During the 15 August 2007 Peruvian Tsunamia
The error of the forecast maximum wave height is estimated as Eforecast = ∣Hforecast − Hobs∣/Hobs × 100%, where Hobs is the observed maximum tsunami wave height, and Hforecast is the DART-constrained forecast maximum tsunami wave height; Hcorr is the cross correlation of forecasts and observations at 12-minute delay; Rnoise = Anoise/Aforecast measures the ratio of the noise level relative to the tsunami signal, where Anoise and Aforecast are respectively the root-mean-square amplitudes of the 4-hour observation prior to the tsunami arrival and the first 4-hour tsunami signal of the forecast.
Crescent City, CA
 The real-time far-field forecast of the 15 August 2007 Peruvian tsunami event was another successful test of NOAA's experimental tsunami forecast system. The accurate, efficient, and reliable forecasts demonstrated the great potential of the experimental system in real-time operation. The real-time tsunami signal recorded by the DART network guaranteed high-quality input to the forecast system. The data assimilation and inversion provided a DART-constrained tsunami source within 4.5 hours. The forecast results indicated that tsunami flooding would not occur in any of the 14 U.S. coastal communities. Comparison of forecasts with the observations at deep-ocean buoys and coastal tide gages showed excellent agreement for up to 24 hours after tsunami generation in both wave height and period when estimated errors are within the noise level. The time discrepancies between forecasts and observations are attributed to errors in source location, bathymetry, and model approximation.
 Including the 15 August 2007 Peru event, NOAA's tsunami forecasting system has produced excellent experimental forecasts for far-field tsunami impact for six tsunamis since its first test in the 2003 Rat Island tsunami. The essential components of the forecast system are deep-ocean measurements and numerical modeling. High-quality forecasts have shown the strength of the DART implementation, even for just one node of the DART network, in obtaining an accurate tsunami source. A full set of DART buoys will produce better constraint of the tsunami source while providing timelier tsunami detection and observation. With improved DART network and automated inversion processes, we expect a substantial decrease in forecasting time when this system is transferred to the TWCs.
 The authors would like to thank the staff of NOAA's TWC for the joint work on real-time analysis of this event and for providing additional data for post-event analysis. We thank N. Merati and M. Watson for data downloading and processing. We also thank R. L. Whitney for comments and editing. We are grateful to H. Fritz, R. Bell, W. Power, S. Tubbesing, and S. A. Weinstein for sharing their data and findings. This publication is partially funded by the Joint Institute for the Study of the Atmosphere and Ocean (JISAO) under NOAA Cooperative Agreement NA17RJ1232, Contribution 1454; PMEL Contribution 3114.