Time history data of sea levels that are recorded at coastal sites provide one source of information that we can use to study tsunami patterns. To investigate the characteristics of the 2006 tsunami, sea level records from tide gauge stations were employed for analysis in the present study. For this purpose, the recorded data from three tide gauge stations (Kaohsiung, Dongkung, and Houbihu) located in southern Taiwan were obtained. These tide gauge stations are operated and maintained by the CWB, R.O.C. All stations recorded sea levels at a sampling interval of 6 min. In this doublet event, the first and second earthquakes occurred at 20:26:21 and 20:34:15 (national standard time), respectively. Hence, 28 h of tide gauge records (from 08:00 on 26 December 2006 to 12:00 on 27 December 2006, national standard time) were adopted for analysis. To approximate the wave components of the tsunami and to remove the low-frequency noise that was attributed to the tidal effect, the sea level records at the tide gauge stations were de-tided by removing the long-period (> 2 h) tidal constituents. The original data recorded at the tide gauge stations in southern Taiwan are shown in Fig. 2a, and the de-tided data are presented in Fig. 2b. The locations of the tide gauge stations are shown in Fig. 3.

The tsunami duration represents the observation time of high-energy tsunami waves persisting at a coastal site. The tsunami durations at all the stations were identified based on a calculation of root mean square (RMS) sea levels, indicating the elapsed time of the wave amplitude above the normal oscillation level before the tsunami wave arrived (Hayashi et al., 2012; Heidarzadeh and Satake, 2014; Heidarzadeh et al., 2019, 2021). The RMS analysis calculated the moving average of the recorded sea level along a moving time window of 24 min. The calculation for RMS sea level is presented in Eq. (1). 1St=1w∫t-w2t+w2hx2dx In this equation, S(t) represents the RMS sea level at time step t, h(t) denotes the recorded sea level at time t, and w stands for the moving time window. In the present study, the length of the tsunami data employed for RMS analysis is 12 h, which includes 120 data points, ranging from 17:00 on 26 December 2006 to 05:00 on 27 December 2006 (national standard time).

To apply spectral analyses to the tsunami data, two types of analyses were included and processed in this study: Fourier analysis and wavelet (time–frequency) analysis. The Fourier analysis is based on the fast Fourier transform (FFT) algorithm and applied based on the updated open-source library, NumPy, in the Python package (Harris et al., 2020). Fourier analysis was performed to estimate the spectral components of the time history data of the tsunami waveform. The entire dataset of the tsunami waveform inputted for Fourier analysis covered 600 min, which included 100 data points ranging from 5 h before to 5 h after the tsunami, as the sampling rate of the data was 6 min. The Fourier analysis was separately applied to the de-tided background (i.e., 5 h data before the tsunami arrival) and the tsunami signals (i.e., 5 h data after tsunami arrival) to identify significant changes in the spectral energy associated with the tsunami. Additionally, the spectral ratio was computed for the tsunami spectra to exclude the local modes of coastal sites from the periodic components. Wavelet analysis was computed based on the Morlet mother function (Torrence and Compo, 1995). Wavelet analysis detects the periodic change in spectral peaks over time. The length of the tsunami data input in the wavelet analysis was 15 h (15:00 on 26 December 2006 to 06:00 on 27 December 2006, national standard time). A similar method has been widely applied to solve time–frequency problems for many tsunami events, such as the 1945 Makran earthquake tsunami and the 2020 Alaska earthquake tsunami (Heidarzadeh and Satake, 2015; Heidarzadeh and Mulia, 2021).

Numerical simulation is a computer-based method that describes equations for the motion of tsunami wave propagation. Tsunami wave propagation can be numerically modeled based on various theories, including shallow water and dispersive wave theories. Among those theories, the shallow water equations are some of the most commonly used methods to model tsunami propagation from the source to nearshore areas. Various computational models have been developed to solve shallow water equations, and the TUNAMI (Tohoku University Numerical Analysis Model for Investigation of tsunamis) code is one of the widely used models to numerically simulate both far-field and near-field tsunamis (Suppasri et al., 2012, 2014). The second version of the TUNAMI code (TUNAMI-N2) was mainly developed to deal with near-field tsunamis by applying the nonlinear theory of shallow water equations, which is solved using a leap-frog scheme (Imamura, 1995). Since the 2006 tsunami presented as a near-field tsunami in Taiwan, the TUNAMI-N2 model was used in this study to simulate the 2006 tsunami with nonlinear shallow water equations. The nonlinear shallow water equations on the Cartesian coordinate system are presented in Eqs. (2)–(4), and the nonlinear equations are solved by applying the finite difference method. 2∂η∂t+∂M∂x+∂N∂y=03∂M∂t+∂∂χM2D+∂∂yMND+gD∂η∂χ+gn2D73MM2+N2=04∂N∂t+∂∂χMND+∂∂yN2D+gD∂η∂y+gn2D73NM2+N2=0 In these equations, η is the water level, M and N are the discharge fluxes in the x and y directions, respectively, D is the total water depth, g is the gravitational acceleration, and n is Manning's roughness coefficient. The bottom friction term was represented by the Manning roughness coefficient, which was set as 0.025 s m-1/3, assuming that the seafloor in the model domain is in perfect condition. The numerical tsunami simulations were conducted with a time interval of 0.1 s and grid intervals of 450 m. The entire model domain covered the source region and southern Taiwan, which comprised mesh numbers of 538 and 631 in the x and y directions, respectively. The time interval and grid intervals were set up to satisfy the Courant–Friedrichs–Lewy (CFL) condition to ensure the stability of the simulation. The CFL condition is presented in Eq. (5): 5Δt≤Δx2ghmax⁡, where Δt is the time interval, Δx is the grid spacing, and hmax⁡ is the maximum water depth in the model domain. As the initial condition inputted for numerical tsunami simulation, the initial water level distribution was calculated from the earthquake fault parameters using Okada's theory (Okada, 1985). In addition, the horizontal deformation contribution to tsunami generation on steep bathymetric slopes was included (Tanioka and Satake, 1996). The calculation conditions for the numerical tsunami simulation are summarized in Table 1.

In addition to the fault parameters of the source models, bathymetry data are needed for simulating tsunami wave propagation. Simulated tsunami propagation results are known to be sensitive to the accuracy and resolution of bathymetry data. Although it can be expected that bathymetry data with a higher accuracy and resolution can produce simulated results that better fit the actual values, such data are not always available and freely accessible. Due to this limitation, open-source datasets have often been utilized for modeling tsunamis in many previous studies (Koshimura et al., 2008; Li et al., 2016; Otake et al., 2020; Wang et al., 2022; Ren et al., 2022).

Unfortunately, open-source datasets are sometimes problematic and insufficient for the accurate simulation of tsunami waves because they lack accurate, quality data (Griffin et al., 2015). A similar issue has been reported by previous studies in simulating the 2018 Sulawesi tsunami and the 2018 Anak Krakatoa tsunami (Heidarzadeh et al., 2019; Zengaffinen et al., 2020). Significant differences in various sources of datasets can also result in modeled results that contrast estimated values from previous studies. Therefore, for the purpose of tsunami hazard assessment, it is important to assess and note different available open-source bathymetries in relation to model performances, using the 2006 tsunami as an example.

For this purpose, a tsunami simulation was separately applied to two different sources of bathymetry data, namely General Bathymetric Chart of the Oceans (GEBCO) data and ETOPO1 data, and the misfit between the modeled results was evaluated. The GEBCO data contain bathymetry data with grid intervals of 15 arcsec, while ETOPO1 data have sea depth data with a resolution of 1 arcmin. To fairly investigate the model performances from different datasets, bathymetry data from the two datasets were converted to 450 m grids and used as the input for the numerical tsunami simulations. Figure 7 shows the bathymetry data of the modeled domain obtained from GEBCO and ETOPO1 data. As the initial condition, the initial water distribution of the tsunami generated by the proposed multiple fault model (LS2) was used for these simulations, in which the asperity locations of the two earthquakes were assumed to be at the center of the entire fault area.

To examine any significant change in tsunami wave transmission that could be attributed to the bathymetry effect during the passage of the 2006 tsunami, numerical experiments (MS, EXP1, EXP2) for tsunami propagation were conducted using actual and manipulated bathymetry data. For the main simulation (MS) numerical experiment, actual GEBCO bathymetry data with a resolution of 450 m derived from sea depth data with grid intervals of 15 arcsec were used. For the manipulated bathymetry data that were used for numerical experiment EXP1, sea depths greater than 500 m were replaced with 500 m depths. For numerical experiment EXP2, the bathymetry data were manipulated by removing sea depth data with a flattened sea bottom at a depth of 500 m. The 500 m depth was specified because the bathymetric slopes are very gentle at sea depths shallower than 500 m near southern Taiwan, and the area is therefore considered a shelf region. Figure 8 shows the map-manipulated bathymetry of the model domain for numerical experiments EXP1 and EXP2. The details of the bathymetry data used for numerical experiments MS, EXP1, and EXP2 are summarized in Table 5.

The results of the numerical experiments were compared to examine how tsunami wave directivity could change due to the bathymetric effect and to evaluate how much tsunami wave energy could be coastally trapped in different bathymetric conditions during the passage of the 2006 tsunami.

The December 2006 earthquake tsunami was observed at several tide gauges located along the southwestern coast of Taiwan. The tsunami observations are plotted in Fig. 9a. The initial wave arrived at all three tide stations in southern Taiwan with an amplitude sign of a trough wave. The travel times of the initial wave to all the stations were recorded: 16 min to Houbihu, 28 min to Dongkung, and 52 min to Kaohsiung. The initial wave was recorded as -0.12 m in Houbihu, -0.09 m in Dongkung, and 0.06 m in Kaohsiung. Following the trough sign of the initial wave, the first wave crest record at Houbihu was 0.3 m, which was approximately 3 times greater than that at Dongkung and 4 times larger than that at Kaohsiung. This was natural because Houbihu was the station closest to the epicentral region and therefore had an earlier arrival time and was relatively sensitive to the surface elevation change in sea level that was induced by the tsunami. The maximum wave heights were recorded as 0.08 m (Kaohsiung), 0.12 m (Dongkung), and 0.3 m (Houbihu). In Kaohsiung and Dongkung, the maximum height was not recorded for the initial wave. The maximum wave height appeared 36 min after the initial wave arrived at Kaohsiung and after 24 min at Dongkung, indicating a pattern of wave amplification at these stations. These results suggest that different propagation effects existed at these coastal sites during the passage of the 2006 tsunami. In addition to significant differences in wave amplitude and arrival time, the tsunami records at each station also varied in terms of visible wave periods. The visible period of the tsunami wave at Kaohsiung was recorded from 30–48 min based on the tsunami waveform, which was approximately 2 times longer than those observed at Dongkung and Houbihu (from 18–24 min). This indicated that wave components with shorter periods were not well recorded in Kaohsiung. The locations and details of the tide gauge observations are summarized in Table 6a for wave amplitude and Table 6b for arrival time and visible period.

Another issue was to determine the tsunami duration at each station because it can help to identify the length of wave oscillations at a coastal site due to the tsunami. Typically, the tsunami duration describes the elapsed time during which a high-energy wave at a tide gauge station exceeds the mean sea level of a normal oscillation. The normal oscillation was defined as the site-specific oscillation at each station before the tsunami arrived. RMS analysis was applied to the recorded sea level data at each station. The results of the RMS analysis are plotted in diagrams shown in Fig. 9b.

The RMS sea level diagram illustrates how long the high-energy wave persisted at each station. Accordingly, the tsunami duration was determined through a comparison of the RMS sea level and the basic oscillation in sea level at each station. The maximum RMS sea level derived at the Houbihu station was estimated to be 2–3 times higher than those at the Dongkung and Kaohsiung stations. The calculated tsunami duration at Dongkung was as much as 3.9 h, while the tsunami continued for more than 6 h in Kaohsiung and Houbihu.

Generally, several oscillation modes are expected to be induced during a tsunami event associated with the tsunami source, propagation path, and topographic effects (Rabinovich, 1997; Rabinovich et al., 2013). An island setting such as Taiwan, where continental shelves and gentle slopes exist, commonly traps waves over the shelf during the passage of tsunamis (Munger and Cheung, 2008; Roeber et al., 2010). The trapped waves propagate along the coastline and normally trigger various oscillation modes in the coastal water due to the interference of wave reflection at the edge of the continental shelves (Yamazaki and Cheung, 2011). The wave resonance of these oscillation modes with the fundamental modes can enhance coastal hazards with amplified amplitudes and long tsunami durations (Tanioka et al., 2019; Yamanaka and Nakamura, 2020; Wang et al., 2021, 2022). The triggered oscillation modes are expected to be mixed with the tsunami source spectrum in the observation records from the coastal sites. To identify these modes from the tsunami source spectrum, spectral analyses were performed on the observation records at all three tide gauge stations in southern Taiwan, as detailed in the next section.

To examine the spectral characteristics of the tsunami waves, Fourier analysis was applied to 10 h of de-tided observed data (i.e., 5 h before and after the tsunami's arrival) that were recorded at all the tide gauge stations in southern Taiwan. The background spectra were calculated in addition to the spectra of the observed tsunami waveform to identify the tsunami effect. The background spectra were the spectral components calculated from observed data 5 h before the tsunami's arrival, and the spectral components of the observed tsunami waveform were computed using the 5 h of data recorded at the tide gauge after the tsunami wave arrived. Figure 10 shows the respective spectra of the observed tsunami waveform and background signals at each tide gauge station.

At all the stations, the spectral peaks of the observed tsunami spectra were estimated to be different from those of the background spectra. A visible gap also appeared in the spectral energy between the observed tsunami and the background spectra, revealing the energy generated by the arrival of tsunami waves. To examine the spectral components induced by the arrival of the tsunami waves, the spectral ratio of the observed tsunami and background spectra was derived using Eq. (11). 11STω=SobsωSbgω In this equation, Sobs(ω) is the spectral component of the observed tsunami waveform, Sbg(ω) is the background spectrum, and ST(ω) is the spectral component induced by the arrival of the tsunami waves. Figure 11 shows the spectral ratios for the tsunami spectra at all the stations. Equation (11) assumes equivalent background spectra before and after the tsunami's arrival, indicating that there was no large change in the coastal topography during the tsunami event. Although earlier studies have reported that coastal topography might be largely changed during a massive event like the 2004 Indian Ocean tsunami (Masaya et al., 2020), this was not the case for the 2006 Hengchun tsunami because the tsunami wave was small. Therefore, the dominant peaks of the spectral ratio were connected to either the tsunami source or perhaps the wave oscillation induced by the non-source phenomenon.

Tsunami source periods are periodic components that primarily appear in coastal observations close to the tsunami source region (Rabinovich, 1997). Accordingly, the tsunami source periods can be estimated from the mean of the spectral ratios calculated from all three stations (i.e., the solid black line shown in Fig. 11). From the analysis result of the spectral ratio, the periods of 13.6, 16.7, and 23.1 min are distinct in comparison to other periodic components. The periods within this band most likely presented the source periods of the 2006 tsunami since the periodic components within this band were mostly visible at all stations.

In general, a larger earthquake can ordinarily generate a larger tsunami wave with a longer period. For instance, the major periods of the 2011 Tohoku-Oki earthquake tsunami were reported to be 37–67 min associated with that magnitude Mw 9.0 earthquake (Heidarzadeh and Satake, 2013), while shorter dominant periods of 10–22 min were found for the 2013 Santa Cruz tsunami, a Mw 8.0 earthquake (Heidarzadeh and Gusman, 2021). According to Rabinovich's theory (Rabinovich, 1997), the approximate dimensions of the fault rupture can be estimated from the source periods using the empirical formula defined in Eq. (12): 12L=T2gh, where g stands for the gravitational acceleration and is set to a constant value of 9.81 m s-2, h represents the seafloor depth around the tsunami source region, L denotes the fault rupture dimensions of length or width, and T is the source period. The approximate source region could be illustrated based on the aftershock distribution 1 d after the first earthquake occurred. Assuming that the sea depths around the tsunami source region range from 0–600 m and the source periods are 13.6, 16.7, 20.0, and 23.1 min, the relationship between the fault rupture dimensions and sea depths can be derived from Eq. (12). The correlation derived from Eq. (12) is plotted in Fig. 12.

Assuming that the mean sea depth around the tsunami source region is 300 m, the fault rupture dimensions for the two earthquakes can be estimated to be 20–40 km. The approximate fault size of these two earthquakes was estimated to be 800 km2, for which a longer dimension of 40 km was considered the fault length and 20 km was considered the fault width. The estimation of fault size was fairly consistent with the results derived from the empirical scaling relations of Papazachos et al. (2004), with a fault area of 794 km2 associated with the Mw 7.0 normal fault earthquake (first earthquake) and 738 km2 associated with the Mw 6.9 strike-slip fault (second earthquake).

In addition to the Fourier analyses, wavelet (time–frequency) analyses were also applied to 15 h of de-tided observed data (i.e., from 15:00 on 26 December 2006 to 06:00 on 27 December 2006, national standard time) at all the stations. Wavelet analyses are commonly employed as a method to examine periodic variations over time series through the distribution of tsunami spectral energy. Figure 13 shows the tsunami wavelets derived from the tsunami records observed at each station. According to the wavelet plots at all the stations, period bands of 13.6–23.1 min were clearly recorded after the first wave arrived at all the stations. This also confirmed that the period bands of 13.6–23.1 min were associated with the source periods. At Kaohsiung, the tsunami energy became apparent with periods of 16 and 36 min approximately 3 h after the arrival of the first wave. In the period channel of 16 min, the oscillation was preserved for approximately 5 h, while the 36 min channel was occupied by a high-energy wave for more than 9 h. At Houbihu, more energy was channeled than at other stations in the period bands of 13.6–23.1 min soon after the first earthquake. This was reasonable because Houbihu was the closest station to the epicentral region and was therefore considered to be more sensitive to the tsunami source than the other stations were. Following the arrival of the first wave, the persistent oscillation (i.e., lasting more than 4 h) was visible approximately 2 h after the first earthquake in the period channels of 16, 16.4, 20, 22.5, 25.7, 30, 36, and 60 min. These periodic components were considered as possible modes of trapped tsunami waves resonating within the shelf since the wave resonance commonly requires some time to be formed (Heidarzadeh et al., 2021). Among these periods, the 16 and 36 min periods most likely presented the resonance mode since that mode is visible at only the Kaohsiung and Houbihu stations, where tsunami durations of more than 6 h were recorded (Fig. 9b). From the wavelet analysis of the observed data recorded at the tide gauge stations, the persisting wave oscillations at the Kaohsiung and Houbihu stations might be attributed to tsunami resonance.

The sensitivity of simulated tsunami waveforms to non-uniform fault slip distribution was evaluated based on the best fitting fault geometry of S7. The fault model with uniform slip was also modeled to identify the significant differences in the modeled results from the uniform and non-uniform slip fault models.

Figure 18 shows the observed and simulated tsunami waveforms at the Dongkung and Houbihu stations using non-uniform slip models (nine cases in total) and a uniform slip model. At the Dongkung station, the simulated tsunami waveforms from multiple fault models were not much different from those of the single fault models. Both models could produce tsunami waveforms in good agreement with the observed values recorded at this station. At the Houbihu station, the non-uniform slip models produced a significantly higher first wave crest than the observations. The simulated wave peaks from the non-uniform slip models produced wave heights approximately twice those simulated using the uniform slip. These results indicated that the near-field station of Houbihu was rather sensitive to the effect of the fault slip distribution, and some high tsunami waves might have been missing from the recorded signals at the Houbihu station during the 2006 tsunami.

To analyze the tsunami sensitivity to different sources of open-source, accessible bathymetry data, numerical simulations were applied using GEBCO and ETOPO1 data. The differences between the modeled results using these different bathymetry data were evaluated to compare the modeled wave peaks and waveforms of the 2006 tsunami.

Figure 19a and b show the spatial distribution of the maximum wave heights simulated using two bathymetric grids, the GEBCO data and ETOPO1 data. To evaluate the differences between the modeled wave peaks, the variation and percent change in the variation were calculated, which can be defined in Eqs. (13) and (14): 13Varpeak=PeakGEBCO-PeakETOPO1,14%Varpeak=PeakGEBCO-PeakETOPO1PeakGEBCO×100, where Varpeak is the variation in the modeled wave peaks calculated at each computational grid with GEBCO and ETOPO1 data, and PeakGEBCO and PeakETOPO1 are defined as the calculated wave peaks of progressive waves in a unit area of the free surface. Figure 19c and d illustrate the spatial distribution of the variation and percent change in the variation of the modeled wave peaks in the model domain, indicating the differences in the modeled results using the two bathymetries. The results suggested that the variation in the modeled wave peaks using the two bathymetries was greater than 0.05 m and that the percent change was greater than 50 % between the modeled results for areas with sea depths of less than 500 m.

Figure 20 shows the modeled tsunami waveforms at the three coastal stations (i.e., black circles in Fig. 19) using the two bathymetric grids. At Kaohsiung, the modeled waveforms from the two bathymetries matched each other well; however, the modeled wave peak from the ETOPO1 data was significantly smaller than that from the GEBCO data. The bathymetries from the GEBCO and ETOPO1 data could produce tsunami waveforms at Dongkung and Houbihu that were similar in both wave periods and peaks. Table 8 summarizes the details of the coastal stations and the peak variation percentage of the modeled results from the two bathymetries.

It is commonly understood that tsunami velocities are mainly governed by seafloor depths. A tsunami propagates at a slower speed when the tsunami wave enters shallow water from deeper water. The significant change in propagation speed allows the tsunami to change its wave direction. To assess the bathymetry effect on tsunami wave directivity during propagation, simulations were applied using actual (MS) and manipulated bathymetry experiments (EXP1 and EXP2).

Simulated snapshots of tsunami wave propagation using actual (MS) bathymetry data are shown in Fig. 21. The continental shelves in front of Hengchun Peninsula have shallow depths compared to the open ocean. Figure 21a and b present how tsunami waves repeatedly changed their directions among the shelves and then refracted into the west coast embayment. The tsunami waves were reflected from the coast after arrival and tended to radiate offshore. However, they did not fully radiate offshore; instead, they were reflected again at the boundary of the shelf and refracted north toward Kaohsiung and Dongkung (Fig. 21c, d). The high-energy waves repeatedly reflected and refracted among the shelves. Only rarely were tsunamis transmitted back to the open ocean or to the east coast. These results indicated that the tsunami waves were trapped over the shelves during their passage in the 2006 tsunami event. Due to this fluctuation, the high-energy tsunami wave remained along the western coast for a long time, which could be clearly seen at 75 and 90 min after the occurrence of the first earthquake (Fig. 21e, f).

Figure 22 shows snapshots of the simulated tsunami wave propagation using manipulated (EXP1) bathymetry. In this situation, the transmission of tsunami waves in the shallow area was similar to those simulated using the actual (MS) bathymetry, in which the tsunami waves were persistent and repeatedly reflected and refracted among the shelves, but more reflected waves from the coast radiated to the open sea (Fig. 22b–f). This is because the tsunami source was located in an area with sea depths over 500 m, and bathymetry data with sea depths over 500 m were replaced with a 500 m depth in this hypothetical situation.

Aside from the numerical experiment EXP1, a rather hypothetical situation (EXP2) was conducted to simulate tsunami wave propagation on a bathymetry with a flat sea bottom and a sea depth of 500 m. Figure 23 shows snapshots of simulated tsunami wave propagation using the manipulated (EXP2) bathymetry. An inspection of the tsunami wave transmission in the shallow area indicated that the reflected tsunami waves from the coast radiated homogeneously offshore, and the wave reflection and refraction could not be clearly seen. In addition, the tsunami waves propagated at a rather fast speed (i.e., in comparison to MS and EXP1) and mostly radiated out of the model domain at 75 and 90 min after the occurrence of the first earthquake (Fig. 23d, e).

While the past section specified that tsunami waves are trapped over shelves due to the wave directivity change associated with the configuration of coastal bathymetry, the question remains of how much wave energy can be trapped over the shelves in front of southern Taiwan during the passage of tsunamis. To quantitatively evaluate the wave energy trapped over the shelves, the trapped ratio was used to indicate the tsunami energy trapped in bathymetric situations, as calculated in Eq. (15): 15RT=EShelfETotal×100, where RT is the ratio of tsunami energy trapped, EShelf is the calculated tsunami potential energy on the shelves (i.e., shallow areas with sea depths under 500 m), and ETotal is the calculated total tsunami potential energy of the model domain at each time step. The tsunami potential energy was determined assuming that the energy flux of the tsunami wave progressed in a unit region of the free sea surface and was determined using Eq. (16) (Nosov et al., 2014):​​​​​​​ 16Ep=∯12ρgη2dxdy, where Ep is the tsunami potential energy, ρ is the water density of the ocean, g is the gravitational acceleration (set as 9.81 m s-2), and η represents the surface integral of the ocean surface disturbance at each time step. The ratio of trapped tsunami energy was calculated from the snapshots of tsunami simulations using actual (MS) and manipulated (EXP1 and EXP2) bathymetry. Figure 24 shows the calculated trapped ratio from simulated tsunami propagation snapshots every 15 min using actual (MS) and manipulated (EXP1 and EXP2) bathymetry. Note that for calculating the trapped ratio from simulations using manipulated bathymetry (EXP1 and EXP2), the shelf region corresponding to the actual bathymetry (MS) was used (i.e., the shallow area illustrated by the solid and dashed black lines shown in Figs. 22 and 23). According to Eqs. (15) and (16), the simulations yielded a ratio of trapped tsunami energy of more than 50 % when using actual bathymetry (MS) and manipulated bathymetry (EXP1) but a smaller trapped ratio of 20 % when using manipulated bathymetry (EXP2). These results quantitatively provided another confirmation that the coastally trapped tsunami wave energy was related to the shape of the bathymetry.

To understand any significant change in tsunami waveforms that can be recognized with and without wave trapping, tsunami waveforms simulated from actual (MS) and manipulated bathymetry (EXP1 and EXP2) were compared. Figure 25 shows the simulated tsunami waveforms at the three coastal stations in southern Taiwan using actual and manipulated bathymetry.

Using the manipulated bathymetry (EXP1), the first few circles of simulated tsunami waveforms at all the stations were consistent with those simulated using actual bathymetry (MS) but produced slightly smaller later-phase amplitudes. An inspection of the simulated waveforms using the manipulated bathymetry (EXP2) indicated an earlier arrival time of the first wave and smaller amplitudes of the later phase than those of the simulation results using actual (MS) bathymetry. These results indicated that the persistent high-energy waves along the south coast of Taiwan were associated with the mechanism of tsunami wave trapping.

As described in the previous sections, the tsunami wave was trapped over the shelves and transmitted along the coast as edge waves during the 2006 tsunami. This section describes how tsunami waves behave as edge waves and to what extent such wave fluctuations influence the amplified and persisting high-energy waves along the south coast of Taiwan. Figure 26 shows the shelves in front of south Taiwan and the simulated tsunami heights of the 2006 tsunami from the main simulation (MS).

To study the behaviors of edge waves along the south coast during the 2006 tsunami, a time–distance diagram of tsunami waves is shown. Figure 27a shows the time–distance diagram of the tsunami wave along the contour of the 20 m sea depth (i.e., dashed black line in Fig. 26a). Based on the phase shift of the tsunami wave, the propagation path and the travel time curve of edge waves are illustrated (i.e., green arrow in Fig. 27a). According to the travel time curve, the edge waves propagated along the coast at a speed of 50 m s-1. The edge waves propagated along the coast and were iteratively reflected at the shelf edge. The coupling of the edge waves and the later-arriving incident waves amplified the tsunami waves and maintained the wave oscillation in the later phase. These were visible from simulated tsunami waveforms at numerical wave gauges C and E, as shown in Fig. 27c.

To understand the persisting high-energy waves along the south coast of Taiwan during the 2006 tsunami, the decreasing tendency of the tsunami wave energy along the 20 m sea depth contour was analyzed. The temporal tsunami wave energy was first determined using Eq. (16) and then normalized according to the maximum temporal tsunami energy in the time series. Figure 27b shows the time–distance diagram of the normalized tsunami energy along the 20 m sea depth contour (i.e., dashed black line in Fig. 26a). Figure 27d shows the normalized tsunami energy at numerical wave gauges C and E. At the numerical wave gauge C, the normalized tsunami energy achieved its greatest value at approximately 40 min after the first earthquake occurred. However, this high-energy channel did not decrease with time after the first wave arrived; instead, a persisting channel of strong energy was visible. This energy channel lasted for more than 60 min, and the wave energy repeatedly reached the maximum value in this channel. Beyond this channel, the energy commenced to decrease with a rate of energy loss of 50 % at 110 min and 20 % at 270 min after the occurrence time of the first earthquake. At the numerical wave gauge E, the normalized tsunami energy achieved its greatest value approximately 30 and 120 min after the first wave arrived. Beyond this channel, the energy commenced to decrease at a rather fast rate of energy loss of 80 % at 150 min and 70 % at 215 min after the occurrence time of the first earthquake. Accordingly, the tsunami decay process in this region was expected to last for more than 300 min. These results indicated that the wave amplification and persistent high-energy waves along the coast during the 2006 tsunami were connected to tsunami wave trapping and the influence of edge waves. According to these behaviors, southern Taiwan could be affected by intensified coastal hazards and severe impacts from tsunamis.

In this article, the characteristics of the consecutive tsunamis on 26 December 2006 and the resulting tsunami behaviors in southern Taiwan were investigated and clarified. The methodology comprised analyses of tide gauge tsunami waveforms, spectral analyses, and numerical tsunami simulations. The main findings are summarized as follows.

The physical characteristics of the tsunami waveforms at all three tide gauge stations in southern Taiwan during the December 2006 tsunami were analyzed. The initial tsunami wave arrived at Kaohsiung, Dongkung, and Houbihu at 21:18, 20:54, and 20:42 national standard time, respectively, with a trough sign of tsunami amplitude. Following the initial wave trough, the initial wave crests were 0.07 m (Kaohsiung), 0.09 m (Dongkung), and 0.3 m (Houbihu). The maximum tsunami wave heights at the three tide gauge stations at Kaohsiung, Dongkung, and Houbihu were 0.08, 0.12, and 0.3 m, respectively, and the maximum tsunami wave heights at Kaohsiung and Dongkung were not recorded with the first waves. The approximate tsunami duration in Dongkung was 3.9 h, while the tsunami lasted for more than 6 h in Kaohsiung and Houbihu.

In this study, the characteristics of the December 2006 tsunami and resulting tsunami behaviors in southern Taiwan were explored using available data from tide gauge tsunami waveforms and numerical tsunami simulations. Nevertheless, the analyses in this article had some limitations. The first limitation was related to the tsunami data recorded at the tide gauge stations, which were employed as input data for the spectral analyses (i.e., Fourier analyses and wavelet analyses) and compared with the numerical results. The sampling interval of the tide gauge data recorded at all CWB tide gauge stations was 6 min, indicating that tsunami wave components with shorter periods might not be well recorded in the tide gauge data. Due to this existing limitation, spectral analyses might cause discrepancies in detecting periodic components of tsunami spectra. This limitation could be improved by including tsunami data with more frequent sampling rates.

Another limitation was related to the simulation grid size (i.e., 450 m) for the tsunami propagation simulation. Although the simulated tsunami waveforms were reasonably consistent with the observed values recorded at the tide gauge stations in terms of wave amplitude and arrival time, the reproducibility of the numerical results for the 2006 tsunami could be further improved by constructing a finer grid of bathymetric data.

All the limitations mentioned above suggest further improvements to research to provide a more detailed investigation of long-lived edge wave and shelf resonance issues, especially in the region of southern Taiwan. In addition, more fundamental studies on the complex wave mechanisms of tsunami reflection and refraction, shoaling effects, and wave trapping by insular shelves are planned for future work.