 Technical report
 Open access
 Published:
An approximate travel time calculation and a robust GNSSacoustic positioning method using an MCMC technique
Earth, Planets and Space volume 74, Article number: 176 (2022)
Abstract
It is important to consider horizontal heterogeneity in an underwater sound speed structure to accurately estimate positions of GNSSacoustic sites. Although large amounts of moving survey data (a seasurface platform moves around when acoustic signals are transmitted) are required to accurately detect a sloping sound speed structure, the actual observational data do not necessarily include sufficient moving survey data due to sea conditions or observational time. To treat these insufficient data, it was assumed that a shallow sound speed gradient was continuously present up to a fixed water depth (gradient depth). However, the validity of this assumption has not been investigated, and the gradient depth has not been optimized. In this study, we developed a new GNSSacoustic array positioning method that optimizes the gradient depth using an MCMC technique. To employ this technique, we also developed an approximate technique for rapidly calculating travel time, because the conventional travel time calculation requires high computational cost for integration into the MCMC technique. We assessed the performance of the approximate travel time calculation technique and demonstrated its sufficient accuracy and precision for estimating array positions. Then, we applied the new GNSSacoustic array positioning method to the actual observational data collected by the Japan Coast Guard and Tohoku University. Using enough amounts of the moving survey data, our method demonstrated the results comparable with the conventional GNSSacoustic positioning method estimating a sloping sound speed structure; thus, the assumption of the sound speed gradient with the fixed water depth was valid. Moreover, due to the physical restriction of this assumption, our method provided robust solutions even when the observational data contained small quantities of moving survey data with a simple seasurface track. Although our method still cannot be used in the scenario, where no moving survey data are available, it can work robustly compared with the conventional methods.
Graphical Abstract
Introduction
Over the last 10 years, seafloor geodetic observational techniques, particularly the GNSSacoustic (GNSSA) technique, have provided important geophysical results and have been technically improved. The GNSSA observational technique was contrived by Spiess (1985): a set of several (generally 3–6) seafloor transponders composes one GNSSA site as a transponder array. GNSSA positioning can be performed as a combination of GNSS positioning at a seasurface platform and acoustic ranging between the seasurface platform and seafloor transponders; seafloor displacement can be obtained as displacement of the transponder array, assuming that all seafloor transponders at each site move as a rigid body (referred to as the array positioning method). A basic GNSSA positioning method was established around 2000, and it has shown geophysical observational results related to fault slips (e.g., coseismic slips and interplate coupling) in subduction zones (Spiess et al. 1998; Fujita et al. 2006; Kido et al. 2006; Ikuta et al. 2008). The importance of seafloor geodetic observations has been widely recognized since the occurrence of the 2011 Tohoku–Oki earthquake. The GNSSA observations detected high coseismic displacements of over 30 m in the 2011 Tohoku–Oki earthquake (Kido et al. 2011; Sato et al. 2011), suggesting that a large coseismic fault slip of over 50 m reached the trench (Iinuma et al. 2012). After the 2011 Tohoku–Oki earthquake, GNSSA observational networks around Japan were reinforced and provided a broad postseismic deformation pattern in the Tohoku region (Tomita et al. 2017; Watanabe et al. 2021) and a spatial heterogeneity pattern of the interplate coupling in the Nankai region (Yokota et al. 2016). Recently, the GNSSA positioning method has been improved by considering spatial heterogeneities in underwater soundspeed structures (Yasuda et al. 2017; Yokota et al. 2018a; Honsho et al. 2019; Watanabe et al. 2020). This accurate positioning method has enabled the discussion of various geophysical phenomena, such as slowslip events, in detail (Honsho et al. 2019; Yokota and Ishikawa 2020).
The GNSSA positioning methods have been performed by a linearized least squares method, because unknown parameters for seafloor transponder positions are nonlinear. Most studies have used a classic iterative nonlinear leastsquares (NLLS) method, such as the Gauss–Newton method, for conventional kinematic estimation (i.e., positioning for each acoustic ping, e.g., Spiess et al. 1998; Kido et al. 2006) and simple static estimation (i.e., positioning using a batch of acoustic pings, e.g., Honsho and Kido 2017). Meanwhile, some studies have introduced hyperparameters for constraining the temporal fluctuation of the sound speed structure and optimized them using the Akaike Bayesian Information criterions (ABIC) (Akaike 1980) framework or trialanderror (Ikuta et al. 2008; Watanabe et al. 2020). This approach is applicable under the assumption that each unknown parameter follows a Gaussian distribution, and it is difficult to consider the influence of the uncertainty of the hyperparameters on the positioning errors.
Considering spatial sound speed heterogeneities, the recent GNSSA positioning methods often assume a gradient structure of the sound speed. The sound speed gradient structure may be expressed by four parameters, which are often referred to as “shallow” and “deep” gradients, and each gradient has two components in the East–West (EW) and North–South (NS) directions (Yokota et al. 2018a; Honsho et al. 2019; Watanabe et al. 2020). Shallow gradients can easily be found as biased travel time residuals following the track of a seasurface platform. The influence of shallow gradients on the travel time is a tradeoff with temporal fluctuation of the sound speed as long as a single seasurface platform is used (Honsho et al. 2019); therefore, their influence on the positions is not significant. However, the deep gradients cause biases of up to dozens of centimeters in the positioning results. The number of acoustic pings with various shot angles is critical for the estimation of the deep gradients (Honsho et al. 2019). Ideally, the deep gradients can be estimated using a transponder array with wellarranged over five transponders, even if a seasurface platform stays at a fixed point (point survey) (Kido 2007; Tomita et al. 2019). However, most GNSSA sites have only four transponders, and the estimation of the deep gradients using only point survey data is significantly affected by shortterm and smallscale perturbations of the sound speed structure (Matsui et al. 2019). Therefore, it is important to use acoustic data when a seasurface platform moves around (moving survey) to estimate deep gradients. Meanwhile, all GNSSA campaign data do not necessarily include enough moving survey data because of sea conditions or observational objectives. Note that our previous studies mainly conducted the point survey for parts of the GNSSA campaigns, because the point survey has numerically better positioning precision for horizontal components than the moving survey (e.g., Tomita et al. 2015, 2017). To analyze these data, Honsho et al. (2019) assumed that shallow sound speed gradient is continuously and constantly present up to a certain depth; in other words, the deep gradients are proportional to the shallow gradients with a proportional coefficient using a certain depth (hereafter referred to as the gradient depth). A similar approach was adopted by Yasuda et al. (2017) and Kinugasa et al. (2020). Although the gradient depth was fixed at approximately 500–1000 m in these studies, this value was not optimized according to the data. The gradient depth should be appropriately set, because this value functions as a type of hyperparameter. However, since the gradient depth has a numerical limit (from zero to seafloor depth), this makes the optimization of gradient depth with a simple least squares method difficult.
In this study, we implement a Markov Chain Monte Carlo (MCMC) technique in the GNSSA array positioning method (hereafter referred to as the MCMCbased array positioning method). The MCMC technique can flexibly provide posterior probability distributions of the unknown parameters without the Gaussian distribution assumption and can also estimate the uncertainties of hyperparameters, similar to the other unknown parameters. Moreover, the MCMC technique easily provides the numerical limits of unknown parameters. These characteristics are expected to work well for the array positioning method including the gradient depth parameter. However, the MCMCbased array positioning method has a critical problem in terms of its computational cost. In GNSSA positioning, we must perform travel time calculations for each acoustic ping. Because a travel time calculation technique requires an iterative acoustic raytracing technique with numbers of the sound speed layers, the computational cost of the MCMC technique is quite high. Thus, in this study, we developed a fast approximate travel time calculation technique and implemented it in the MCMCbased array positioning method. We then applied the new method to actual observational data.
Method: approximate travel time calculation
In this section, we first introduce a conventional travel time calculation method as a reference for the approximate travel time calculation in “Introduction of exact travel time calculation” section. We then introduce two types of approximate travel time calculation methods in “Approximate travel time calculation” section. Their performance and parameter tuning are investigated in “Parameter optimization of the approximate travel time calculation” section.
Introduction of exact travel time calculation
In this section, we introduce a conventional travel time calculation method using the shooting method in a horizontally stratified soundspeed structure. This method considers Snell’s law in a spherical frame. In this method, we approximate the Earth’s shape around an observational site using a spherical Earth model with a local radius. This local Earth radius must naturally consider the acoustic ray azimuth angle of each acoustic ping to express the curvature of the ellipsoidal Earth (Chadwell and Sweeney 2010), and contributions of the azimuth angles require complex functions to our approximation approach. Meanwhile, Chadwell and Sweeney (2010) reported that approximate ellipsoidal Earth radii (such as local mean radius and Gaussian mean radius) generally provided ranging accuracy of < 1 mm for a few kilometers of the horizontal distance between the seasurface and seafloor points without considering the azimuth angle contributions. Thus, we adopted the Gaussian mean radius as the local Earth radius for simplicity. The Gaussian mean radius is expressed as
with
where \(a\), \(e\), and \(\phi\) are the semimajor axis, eccentricity of the ellipsoidal earth, and latitude, respectively.
We then formulated the travel time along path A–B, as shown in Fig. 1. The sound speed structure was given as N layers, and the sound speed and radius of the nth layer are denoted as \({v}_{n}\) and \({r}_{n}\), respectively. Point A was located on the sea surface, whereas point B was located on the seafloor with an angular distance of \(\varDelta\) from point A. From Snell’s law, we can obtain the ray parameter \(p\) as follows:
where \(v(r)\) and \(\xi (r)\) indicate the sound speed and ray angle at the radius of \(r\). The ray parameter is preserved throughout the layers. The angular distance and travel time along path A–B can be calculated as follows:
where \(dl\) is the infinitesimal distance along path A–B; then, \(dl\mathrm{cos}\xi =dr\). Using Eq. 3 and considering \({v}_{n}\) as constant in the corresponding layer, Eqs. 4 and 5 were rewritten using integral formulas as follows:
Since the positions of A and B were given, the angular distance could be obtained as follows:
where \({l}_{{\mathrm{AB}}}\) is the geodetic distance between A and B and is calculated as the horizontal distance in a local Cartesian coordinate by the transverse Mercator method. Thus, we could estimate the appropriate \(p\) using an iterative nonlinear leastsquares method that minimizes \(\Delta {\Delta }_{\mathrm{cal}}\). Note that we assigned the initial value of \(p\) using Eq. 3 with the sound speed at the uppermost underwater layer, and then performed the Gauss–Newton method as the nonlinear least squares method. Using the appropriate \(p\) estimated by the nonlinear least squares method, we could obtain the precisely calculated travel time (hereafter referred to as \({T}_{\mathrm{Exact}}\)) using Eq. 7.
Approximate travel time calculation
Next, we developed a method to obtain an approximate travel time, \({T}_{\mathrm{Appr}}\), reproducing \({T}_{\mathrm{Exact}}\). First, we introduced “rough” travel time, \({T}_{\mathrm{Sphere}}\), in the spherical coordinates without considering Snell’s law. Denoting travel time in the vertical path (A, B in Fig. 1) as \({T}_{\mathrm{vert}}\), it was calculated as follows:
where \({R}_{{\mathrm{A}}}\) and \({R}_{{\mathrm{B}}}\) are the Earth’s radii at points A (sea surface) and B (seafloor), respectively. Assuming the ellipsoidal heights at points A and B to be \({h}_{{\mathrm{A}}}\) and \({h}_{{\mathrm{B}}}\), respectively, \({R}_{{\mathrm{A}}}={R}_{G}\left(\phi \right)+{h}_{{\mathrm{A}}}\) and \({R}_{{\mathrm{B}}}={R}_{G}\left(\phi \right)+{h}_{{\mathrm{B}}}\) are obtained, respectively. Since \({h}_{{\mathrm{A}}}\) fluctuated from a certain height \({\overline{h} }_{{\mathrm{A}}}\) due to sea waves during an observational survey in the actual data, \({T}_{\mathrm{vert}}\) should be calculated for each acoustic ping in nature. However, integral calculation required significant calculation costs. Thus, we approximately calculate \({T}_{\mathrm{vert}}\) using \(\delta h\) which is defined as \(\delta h={h}_{{\mathrm{A}}}{\overline{h} }_{{\mathrm{A}}}\):
where \({\overline{R} }_{{\mathrm{A}}}\) is defined as \({\overline{R} }_{{\mathrm{A}}}={R}_{G}\left(\phi \right)+{\overline{h} }_{{\mathrm{A}}}\). Thus, once we calculated the first term on the righthand side of Eq. (10), the approximate \({T}_{\mathrm{vert}}\) could be easily calculated. Then, travel time in the spherical coordinates could be calculated as follows:
with
As shown above, \({T}_{\mathrm{Sphere}}\) could be obtained by the simple calculation.
Then, we estimate the correction terms for \({T}_{\mathrm{Sphere}}\) to obtain the approximate travel time \({T}_{\mathrm{Appr}}\). When the correction terms are expressed as a function \(f\), \({T}_{\mathrm{Appr}}\) is defined as
As shown in “Parameter optimization of the approximate travel time calculation” section, the correction terms can be obtained as polynomial functions of \(\Delta\) and \(\delta h\). We investigated two models of the correction terms: one includes \(I\)th degrees of \(\Delta\) (Model1: \({f}_{1}\)), and the other includes \(I\)th degrees of \(\Delta\) and \(\delta h\) (Model2: \({f}_{2}\)) as in
By assigning various values to \(I\), we investigated the best pattern for the correction terms. Once the best pattern of the correction terms was determined, we could estimate the correction coefficients (\({a}_{i}\), \({b}_{i}\)) for each seafloor transponder through a numerical test in advance as follows: (1) numbers of the synthetic observational points with various \(\Delta\) and \(\delta h\) were prepared; (2) \({T}_{\mathrm{Exact}}\) and \({T}_{\mathrm{Appr}}\) for each seafloor transponder were calculated using the synthetic points; (3) the correction coefficients for each seafloor transponder were estimated by a simple least square method using \({T}_{\mathrm{Exact}}{T}_{\mathrm{Appr}}\) as a data vector. After this optimization, we could calculate \({T}_{\mathrm{Appr}}\) that precisely reproduced \({T}_{\mathrm{Exact}}\) for a given \(\Delta\) and \(\delta h\) using Eqs. 13–15 and the optimized correction coefficients. Although this method initially required calculation costs to optimize the correction coefficients, it could quickly calculate the approximate travel time after the initial optimization.
Parameter optimization of the approximate travel time calculation
Here, we investigate the optimal correction terms for the approximate travel time (Eqs. 13–15) and the performance of the approximate travel time calculation. Figure 2 shows the exact (\({T}_{\mathrm{Exact}}\)) and rough (\({T}_{\mathrm{Sphere}}\)) spherical travel times in a numerical simulation. We assumed a transponder with a depth of 3000 m and a horizontal position of zero (i.e., \(X=0\) m, \(h=\,3000\) m). Then, we calculated the exact and rough spherical travel times for the transponder from 10,000 seasurface points randomly provided from a horizontal range of \(0\, \mathrm{m}\le X\le 10000\) m and a height of \(5\, \mathrm{m}\le h\le 0\, \mathrm{m}\). The height range imitates shaking of a seasurface platform. A schematic of the numerical simulation is shown in Additional file 1: Fig. S2. A sound speed profile at the G10 site (143.483°E, 38.302°N, depth of 3271 m) produced from the average temperature and salinity models of WOA2018 (World Ocean Atlas 2018; Locarnini et al. 2018; Zweng et al. 2018) was used for the travel time calculation. The residuals of the travel times (\({T}_{\mathrm{Exact}}{T}_{\mathrm{Sphere}}\)) significantly increased depending on the angular distance \(\Delta ,\) as shown in Fig. 2b. Considering this feature, we modeled the residuals using the polynomial functions of \(\Delta\) as shown in Eqs. 13 and 14. The red dots in Fig. 3 denote the residuals using the \(I\) th degrees of polynomial functions (Model1: \({T}_{\mathrm{Exact}}{T}_{\mathrm{Sphere}}{f}_{1}\left(I\right)\)). The polynomial functions of \(\Delta\) largely reduced the residuals; however, systematic residuals up to 0.2 cm remained even using higher degrees, such as in Fig. 3h, assuming \(I=9\). As indicated by the blue dots in Fig. 3, these systematic residuals can be reduced by the height correction terms using \(\delta h\) shown in Eq. (15) (Model2: \({T}_{\mathrm{Exact}}{T}_{\mathrm{Sphere}}{f}_{2}\left(I\right)\)). The RMS (root mean square) in Model2 was approximately 0.01 mm, which was enough precision for the GNSSA positioning.
We performed the same numerical simulations assuming different depths. Additional file 1: Table S1 summarizes the polynomial approximation results at depths of 1500 m, 3000 m, and 5000 m. In all cases, the RMS was approximately 0.01 mm in Model2 with 8d polynomial functions. The Bayesian information criterion (BIC) (Schwarz 1978) did not necessarily show a minimum value in Model2 with the 8d polynomial function, suggesting larger degrees. However, the precision was enough high; we employed Model2 with 8d polynomial functions as the optimal model. To attain much higher ranging accuracy and precision, we need to consider the azimuth contribution to the travel time (discussed in “Introduction of exact travel time calculation” section) and to include its component to the approximation function \(f\), shown in Eqs. 13–15. However, since the currently required accuracy is considered to be below 1 mm, because the other GNSSA observational errors are further evident (e.g., GNSS positioning errors, sound speed structure bias, and bias of relative distance between a GNSS antenna and a seasurface transducer), the performance of our current technique has enough accuracy and precision.
Moreover, to confirm the stability of Model2 with 8d polynomial functions, we additionally calculated the approximate travel times for another set of randomly generated 10,000 seasurface points using the coefficients for the 8d polynomial functions predetermined above; then, we compared the approximate travel times with the exact travel times. This investigation was performed for the above three types of the depth (1500, 3000, 5000 m). As a result, the RMSs are almost same with those shown in Additional file 1: Table S1. Thus, our optimal approach (Model2 with the 8d polynomial functions) can be stably performed without overfitting.
The practical procedure to obtain the approximate travel time \({T}_{\mathrm{Appr}}\) was as follows: (1) calculate \({\overline{h} }_{{\mathrm{A}}}\) by averaging the height component of the timeseries of the sea surface transducer position; (2) generate numerical simulation data for each seafloor transponder, similar to the above numerical simulation in Fig. 2 (note that seasurface points are generated from a horizontal range roughly within the water depth and a height range of ± 2.5 m from \({\overline{h} }_{{\mathrm{A}}}\)); (3) estimate coefficients of the 8d polynomial function in Model2 using a linear leastsquare method, (4) calculate the approximate travel time for the given actual seasurface transducer position data using Eqs. 13 and 15 with the determined coefficients; (5) finally, once the coefficients for the polynomial functions are determined, calculate the approximate travel time for the given actual seasurface transducer position data. Because \({\overline{h} }_{{\mathrm{A}}}\) is different among the observational cruises, the coefficients must be recalculated for each cruise and for each site.
Next, we investigated the performance of the approximate travel time by applying it to the actual observational data. The observational data were collected at the G10 site (water depth of ~ 3270 m) on November 4, 2016. The seasurface vessel’s track and initial sound speed profile are shown in Additional file 1: Fig. S3. Using these data, we performed a conventional GNSSA array positioning by an NLLS method assuming a horizontally stratified sound speed structure in three types of travel time calculation methods: exact travel time, approximate travel time of Model1 with the 4d polynomial function, and the approximate travel time of Model2 with the 8d polynomial function. Details of the conventional GNSSA array positioning method are introduced in “Introduction of the conventional NLLSbased GNSSA positioning methods” section. The approximate travel time of Model1 with the 4d polynomial function was only used for comparison, which is the simplest model among the investigated models with practical ranging precision (< 0.1 cm), as shown in Additional file 1: Table S1. The estimation results are summarized in Additional file 1: Table S2. The estimated position using the approximate travel time of Model2 demonstrated almost the same values as those using the exact travel time. The difference between them was within approximately 0.01 mm. Therefore, our new travel time calculation method could be used for actual observational data with sufficient accuracy. The estimated position using the approximate travel time of Model1 with the 4d polynomial function demonstrated about approximately 0.2 mm in difference in the height component. Although Model1 with the 4d polynomial function was useful for practical use, we adopted Model2 as the optimal model, because it demonstrated superior precision.
Finally, we confirmed the computational cost of the approximate travel time calculation technique. The computational costs vary depending on number of layers in the initial sound speed profile especially by the conventional calculation technique (Exact traveltime calculation technique in “Introduction of exact travel time calculation” section). Here, we measured the computational times for our new technique and for the conventional technique using the above actual observational data at G10 on November 4, 2016. The initial sound speed profile for this site has 658 layers (5 m width for each layer). As a result, our new technique successfully reduced the computation time to 270 times less than the conventional technique. Note that this comparison did not consider the initial configuration of the coefficients for the polynomial functions. Because the initial configuration can be generally conducted for a minute, this initial configuration does not significantly affect for total calculation cost when performing the MCMC process.
Method: MCMCbased GNSSA positioning method
We first introduce conventional GNSSA array positioning methods using an NLLS technique assuming a horizontally stratified sound speed structure and a horizontally sloping sound speed structure in “Introduction of the conventional NLLSbased GNSSA positioning methods” section, and we then present the MCMCbased array positioning method in “MCMCbased GNSSA positioning method” section. The NLLSbased array positioning methods are introduced to validate the proposed method.
Introduction of the conventional NLLSbased GNSSA positioning methods
Here, we introduce NLLSbased array positioning methods before formulating an MCMCbased GNSSA positioning method. We first introduce an NLLSbased array positioning method assuming a horizontally stratified sound speed structure, and then introduce the assumption of a horizontally sloping sound speed structure.
Following Honsho and Kido (2017) and Tomita et al. (2019), the observation equation for the horizontally stratified sound speed structure for the \(n\)th shot to the \(k\)th seafloor transponder is expressed as follows:
with
\({T}_{k,{n}_{k}}^{\mathrm{obs}}\) is the observed roundtrip travel time and \({T}_{k,{n}_{k}}^{\mathrm{cal}}\) is the calculated roundtrip travel time obtained in “Introduction of exact travel time calculation” and “Approximate travel time calculation” sections. The initial position of the \(k\)th seafloor transponder is denoted by \({\mathbf{p}}_{k}\), \(\delta \mathbf{p}\) denotes the array displacement, \({\mathbf{u}}_{{n}_{k}}\) denotes the sea surface transducer position, and \({v}_{0}\) denotes the initial sound speed profile in the vertical direction. \(M\) indicates a normalizing factor which depends on the shot angles \({\xi }_{k,{n}_{k}}\) and corresponds to a mapping function in GNSS positioning. The nadir total delay (NTD), which is indicated by \({C}_{0}\left({t}_{{n}_{k}}\right)\), expresses the temporal fluctuation of the average sound speed at the time of the \(n\)th shot. Similar to Honsho and Kido (2017), we express the temporal fluctuation of the NTD by the superposition of cubic Bspline functions, \(\Phi\). Thus, using \(J\) number of cubic Bspline functions, \({C}_{0}\left(t\right)\) can be expressed as
where \({c}_{j}\) is a coefficient of the \(j\)th component of the cubic Bspline function. Defining \({\tilde{T}}_{k,n}^{\mathrm{cal}}\) using Eqs. 16–18 as
the cost function can be written as
where \({N}_{k}\) and \(K\) indicate the total number of shots for the \(k\)th transponder and total number of seafloor transponders, respectively. Because the calculation of \({T}_{k,{n}_{k}}^{\mathrm{cal}}\) is nonlinear against \(\delta \mathbf{p}\), this estimation of the unknown parameters is performed using the Gauss–Newton method, which is one of the most popular NLLS methods, as in Honsho and Kido (2017). To obtain reasonable solutions, \(J\) must have been optimized in advance. Because the total number of cubic Bspline functions expresses the roughness of temporal fluctuations of the average sound speed, the roughness potentially has a tradeoff relationship with data misfits. Thus, we simply optimized \(J\) using BIC. We assumed that the residuals in Eq. 20 follow the Gaussian distribution and then calculate the likelihood to obtain the BIC.
Following Honsho et al. (2019), the observation equation in the horizontally sloping sound speed structure for the \(n\)th shot to the \(k\)th seafloor transponder is expressed as
with
where \({\mathbf{g}}_{\mathrm{s}}\) and \({\mathbf{g}}_{\mathrm{d}}\) indicate shallow and deep gradients, respectively. These terms contain the EW and NS components of the gradients, respectively. \({\mathbf{u}}^{\mathrm{hor}}\) indicates the horizontal component of \(\mathbf{u}\), and \(\phi\) is the azimuth of the acoustic path. Here, we assumed that the gradients were constant during the survey period (i.e., \({\mathbf{g}}_{\mathrm{s}}\left(t\right)={\tilde{\mathbf{g}}}_{\mathrm{s}}\) and \({\mathbf{g}}_{\mathrm{d}}\left(t\right)={\tilde{\mathbf{g}}}_{\mathrm{d}}\)). Moreover, as long as a single sea surface platform is used, a tradeoff relationship between the contributions of shallow gradients and the temporal fluctuation of NTD appears (Honsho et al. 2019). Therefore, we simultaneously expressed the temporal fluctuation of NTD and the contribution of the shallow gradients by the superposition of the cubic Bspline functions as follows:
Defining \({\tilde{T}}_{k,n}^{\mathrm{cal}}\) using Eqs. 21–23 as
the cost function can be written as
Using the Gauss–Newton method, we could estimate the array position, temporal fluctuation of NTD, and deep gradients. The number of cubic Bspline functions \(J\) was set to be the same value, which is determined by the BIC of the inversion results, assuming a horizontally stratified sound speed structure.
MCMCbased GNSSA positioning method
Using the fast travel time calculation technique described in “Approximate travel time calculation” section, we construct an MCMCbased GNSSA positioning method. Following Honsho et al. (2019), we considered a case in which the deep gradient could not be directly estimated due to the data set quality (i.e., moving survey data were insufficient to directly solve the deep gradient, as described in “Introduction” section). In this case, we assumed that a shallow gradient \({\widetilde{\mathbf{g}}}_{\mathrm{s}}\) was present up to the gradient depth \(D\). According to Honsho et al. (2019), this model can provide the following relationship between the shallow and deep gradients:
Considering Eqs. 21 and 26, we could obtain the following observation equation:
As indicated above, a tradeoff relationship between the contributions of the shallow gradient \({\tilde{\mathbf{g}}}_{\mathrm{s}}{\mathbf{u}}^{\mathrm{hor}}\) and the temporal fluctuation of NTD \({C}_{0}\left(t\right)\) appeared; thus, it was difficult to directly estimate \({\tilde{\mathbf{g}}}_{\mathrm{s}}\) and \(\mathbf{c}\) when \({C}_{0}\left(t\right)\) was modeled by cubic Bspline functions (Eq. 18). To cope with this issue, Honsho et al. (2019) modeled \({C}_{0}\left(t\right)\) as a combination of polynomial functions and cubic Bspline functions, which express the longterm and shortterm fluctuations of NTD, respectively. They then introduced a hyperparameter for suppressing the shortterm fluctuation of NTD and optimized it using the ABIC framework (e.g., Yabuki and Matsu’ura 1992). This approach successfully extracted the contributions of a shallow gradient. However, in this method, it is necessary to introduce a large number of cubic Bspline functions; then, the variation of their coefficients is optimized by ABIC. In our method, number of the cubic Bspline functions is optimally reduced by BIC in advance as noted in “Introduction of the conventional NLLSbased GNSSA positioning methods” section. This optimization is quite useful for fast computation through the MCMC method. Hence, we cannot directly utilize the approach of Honsho et al. (2019) to determine the shallow gradient. Here, we contrived a new approach similar to Honsho et al. (2019). To avoid introducing a hyperparameter that constrains the shortterm fluctuations of NTD in their approach, we propose the following two observation equations:
Equation 28 models the temporal fluctuation of NTD \({C}_{0}\left(t\right)\) and the contribution of the shallow gradient \({\tilde{\mathbf{g}}}_{\mathrm{s}}{\mathbf{u}}^{\mathrm{hor}}\) as cubic Bspline functions following Eq. 23, whereas Eq. 29 models the temporal fluctuation of NTD \({C}_{0}\left(t\right)\) as the longterm fluctuation of NTD by a quartic polynomial function. Therefore, Eq. 28 aims to solve the array displacement similar to Eqs. 24 and 25, and it treats \(\delta \mathbf{p}\) and \(\mathbf{c}\) as unknown parameters. On the other hand, Eq. 29 aims to extract the contribution of the shallow gradient, and it treats \({\widetilde{\mathbf{g}}}_{\mathrm{s}}\) and \({\varvec{\upgamma}}\) as the unknown parameters.
In the Bayesian approach, information on unknowns is expressed using a probability density function (PDF). According to Bayes’ theorem (Bayes 1763), when data vector \(\mathbf{d}\) is given, a posterior PDF for the unknown parameter vector \(\mathbf{x}\) can be formulated as
where \(p\left(\mathbf{d}\mathbf{x}\right)\) indicates the likelihood for a given \(\mathbf{x}\), \(p(\mathbf{x})\) is a priori PDF of \(\mathbf{x}\), and \(p(\mathbf{d})\) is a priori PDF of \(\mathbf{d}\). Since \(p(\mathbf{d})\) indicates evidence and is independent of \(\mathbf{x}\), the posterior PDF is propotional to \(p\left(\mathbf{d}\mathbf{x}\right)p(\mathbf{x})\). In this study, we employed the Metropolis–Hastings algorithm (Metropolis 1953; Hastings 1970). This algorithm samples the posterior PDF from iterative calculations and is one of the most general algorithms for MCMC (Fukuda and Johnson 2008; Kubo et al. 2016). At each iteration step, we produced a candidate unknown parameter vector \(\mathbf{x}\boldsymbol{^{\prime}}\) by adding perturbations to the unknown parameter vector of the previous step \(\mathbf{x}\). In this study, we generated a perturbation for each unknown parameter from a uniform distribution with an individually constant step width. We then calculated the acceptance probability \(\alpha\) as follows:
where \(q\) denotes a proposal PDF. Because the proposal PDFs \(q\left(\mathbf{x}\mathbf{x}\mathbf{^{\prime}}\right)\) and \(q\left({\mathbf{x}}^{\prime}\mathbf{x}\right)\) are symmetric in this case, we could calculate the acceptance probability using the ratio of posterior PDFs. In our approach, we generated a random value \(u\) from a uniform distribution with a range of \(\left[0, 1\right]\) and then accept the candidate when \(\alpha \left({\mathbf{x}}^{\prime}\mathbf{x}\right)>u\).
In this study, we assumed a uniform distribution with a range of \(\left[\infty ,\infty \right]\) for the priori PDF of each unknown parameter except the gradient depth \(D\). For the gradient depth, we assumed a uniform distribution with a range from 0 (seasurface) to an observational site water depth. Thus, when the candidate gradient depth falls within this range, \(p\left(\mathbf{x}\right)=1\); otherwise, \(p\left(\mathbf{x}\right)=0\).
We calculated the likelihood \(p\left(\mathbf{d}\mathbf{x}\right)\) by assuming that the data misfits followed a Gaussian distribution. Here, the data, residual, and unknown vectors are given as \({\mathbf{d}}_{l}\), \({\mathbf{r}}_{l}\), and \({\mathbf{x}}_{l}\), respectively. These vectors for Eq. 28 are provided as \(l=1\), whereas those for Eq. 29 are provided as \(l=2\). The residual vectors were obtained by subtracting the righthand side from the lefthand side of Eqs. 28 and 29. The likelihood PDF for each equation is written as
where \({\mathbf{C}}_{l}\) is a covariance matrix for each case and \(N\) is the total number of data given as \(N=\sum_{k}^{K}{N}_{k}\). We introduced a scaling parameter, \({\Lambda }_{l}\), for each covariance matrix (Tomita et al. 2021) as follows:
where \({\sigma }_{0}\) is the initial observational error, which is given as \({\sigma }_{0}=1.0\times {10}^{4} [s]\). The scaling parameter is given as a power of 10 to promote efficient sampling (e.g., Kubo et al. 2016; Tomita et al. 2021). Including these scaling parameters, the unknown vectors are defined as \({\mathbf{x}}_{1}={\left(\delta \mathbf{p}, \mathbf{c},D, {\Lambda }_{1}\right)}^{\mathrm{T}}\) and \({\mathbf{x}}_{2}={\left({\tilde{\mathbf{g}}}_{\mathrm{s}}, {\varvec{\upgamma}},{\Lambda }_{2}\right)}^{\mathrm{T}}.\)
The MCMC procedure used in this study is summarized in Additional file 1: Fig. S1. We first estimated \(\delta \mathbf{p}\) and \(\mathbf{c}\) by the NLLSbased positioning, assuming a horizontally stratified sound speed structure and also determined \(J\) using BIC, which are treated as initial values for the MCMC iteration (\(\delta {\mathbf{p}}^{\mathrm{ini}}\), \({\mathbf{c}}^{\mathrm{ini}}\)). Then, fixing the obtained \(\delta {\mathbf{p}}^{\mathrm{ini}}\), we calculated the travel times for all data and obtained the following observation equation:
Because \({\tilde{\mathbf{g}}}_{\mathrm{s}}\) and \({\varvec{\upgamma}}\) are linear parameters, this equation could be easily solved using the linear leastsquares method. The obtained solutions were used as the initial values for the MCMC iteration (\({\tilde{\mathbf{g}}}_{\mathrm{s}}^{\mathrm{ini}}\), \({{\varvec{\upgamma}}}^{\mathrm{ini}}\)). We set the scaling parameters' initial values to zero, \({\Lambda }_{1}^{\mathrm{ini}}= {\Lambda }_{2}^{\mathrm{ini}}=0\). Using these initial values for \({\mathbf{x}}_{1}\) and \({\mathbf{x}}_{2}\), we ran an MCMC iteration. If the number of iterations was odd, we provided perturbation in \({\mathbf{x}}_{1}\) and evaluated it based on the acceptance ratio. If the number of iterations was even, we provided perturbation in \({\mathbf{x}}_{2}\) and evaluate it based on the acceptance ratio. We performed \(5\times {10}^{5}\) iterations and collected samples when the number of iterations was even and over the burnin period (we set it to be \(1\times {10}^{5}\)).
Application of the MCMCbased array positioning method
In this section, the performance of the MCMCbased array positioning method using the approximate travel time of Model2 with an 8d polynomial function by applying it to actual observational data. As shown in “MCMCbased GNSSA positioning method” section, our method estimates the gradient depth, and moving survey data are essential (e.g., Honsho et al. 2019). Therefore, we employed observational data collected by the Japan Coast Guard (Watanabe et al. 2021) and investigated the performance of our method. We also applied our method to the observational data collected by Tohoku University (Tomita et al. 2017; Honsho et al. 2019, 2021). The data from Tohoku University included some campaigns without the moving survey data. We investigated the performance of our method on these observational data, which were considered unsuitable for solving underwater sound speed gradient structures. The spatial distributions of the GNSSA observational sites employed in this study are shown in Additional file 1: Fig. S4. Moreover, we used conventional array positioning methods for comparison. Hereafter, we refer to our method and the conventional methods as the following: Method1: the NLLSbased method assuming a horizontally stratified sound speed structure; Method2: the NLLSbased method assuming a horizontally sloping sound speed structure; Method3: the MCMCbased method estimating the gradient depth; Method4: the MCMCbased method fixing the gradient depth of 600 m, similar to previous studies (Yasuda et al. 2017; Honsho et al. 2019; Kinugasa et al. 2020).
Application to GNSSA data collected by Japan Coast Guard
Figure 4 summarizes the GNSSA positioning results of the campaign on August 1, 2014, at KAMN site estimated by Methods 1–3. Figure 4a shows the entire track of this campaign, and Fig. 4b shows the travel time residuals in the nadir direction. In Fig. 4b, purple plots demonstrate the travel time residuals (\(\frac{1}{M\left({\xi }_{k,{n}_{k}}\right)}{T}_{k,{n}_{k}}^{\mathrm{obs}}\frac{1}{M\left({\xi }_{k,{n}_{k}}\right)}{T}_{k,{n}_{k}}^{\mathrm{cal}}\), as shown in Eqs. 28 and 29); the green curve shows the temporal fluctuation of NTD fitted by 3d Bspline functions (\(\sum_{j=1}^{J}{c}_{j}{\Phi }_{j}\left({t}_{{n}_{k}}\right)\), as shown in Eq. 28); the red curve shows the longterm fluctuation of NTD fitted by the 4d polynomial function (\(\sum_{m=0}^{4}{\gamma }_{m}{t}^{m}\), as shown in Eq. 29), and the black curve shows a combination of the longterm fluctuation of NTD and the shallow gradient (\(\sum_{m=0}^{4}{\gamma }_{m}{t}^{m}+{\tilde{\mathbf{g}}}_{\mathrm{s}}{\mathbf{u}}^{\mathrm{hor}}\), as shown in Eq. 29). The shallow gradient \({\tilde{\mathbf{g}}}_{\mathrm{s}}\) used in this figure was estimated using Method3. As indicated in the “Method: MCMCbased GNSSA positioning method” section, the shallow gradient was temporally constant, and Fig. 4b shows that the temporally constant shallow gradient can roughly express the spatial bias of the observed travel times. Since we employed the simple polynomial functions for modeling temporal fluctuation in Eq. (29), the fitness is relatively worse than Eq. (28) (see the black and green curves in Fig. 4b); however, we successfully extracted the spatial change depending on the shallow gradient in Eq. (29). This shallow gradient is also shown in Fig. 4a, where the color shows the onthespot NTDs subtracting the longterm NTD fluctuation, and the direction of the shallow gradients is roughly northwest. Although the introduction of a temporally varying shallow gradient might improve data fitting, these two factors show a tradeoff relationship (e.g., Honsho et al. 2019). We considered the current technique to be adequate to stably obtain a temporally constant shallow gradient; however, a more sophisticated technique must be investigated in the future.
Figure 4c–h shows histograms of the MCMC sampling results for the major unknown parameters: the EW, NS, and UD (updown) components of the array positions, the EW and NS components of the shallow gradients, and the gradient depth. All histograms are expressed as Gaussian distributions (equivalent Gaussian distributions are shown by dashed red curves), because absolute values of the skewness and kurtosis for these parameters are below 0.5; therefore, the assumption that an unknown parameter follows a Gaussian distribution in the conventional NLLSbased array positioning methods is suitable. However, note that the gradient depth occasionally shows a truncated Gaussian distribution when estimated close to the limits of the apriori parameter range (i.e., depths of sea surface or of seafloor). In Figure 4c–e, we also describe Gaussian distributions for the array positions estimated by the NLLSbased array positioning methods (black and blue distributions show the results of Method1 and Method2, respectively). In this campaign cruise, the large offsets were caused in the both EW and NS components between Method1 and Methods2–3 (Figure 4c, d), while the UD positions were almost the same among these three positioning methods (Figure 4e). This suggests the sound speed gradient structure potentially affected the horizontal array positions. Method2 directly estimated the deep gradient without any constraints, whereas Method3 considered contributions of the gradient assuming that the deep gradient was linked to the shallow gradient. In spite of the difference among these methods, PDFs of the estimated positions in Method2 and those in Method3 were similar. Thus, the assumption that the deep gradient was linked to the shallow gradient was appropriate in this campaign.
To investigate the correlation relationship among the estimated parameters by Method3, we described pairplot of the major unknown parameters in Additional file 1: Fig. S5. This figure clearly shows strong correlation between the horizontal array positions and the gradient depth, whereas the updown position and the shallow gradients were determined independently of the other parameters. This indicates that the direction of the sound speed gradient (the shallow gradients) is independently determined and that the intensity of the sound speed gradient is controlled by the gradient depth and affects the horizontal array positions.
To investigate the performance of our new method, we applied the above four array positioning methods to other campaign data collected by the Japan Coast Guard. Figure 5 shows a summary of the estimation results for all campaign data at KAMN distributed by Watanabe et al. (2021), and Additional file 1: Figs. S6–S10 show the estimation results for other sites of the Japan Coast Guard in the Tohoku region (KAMS, MYGI, MYGW, FUKU, and CHOS). Additional file 1: Fig. S11 shows a comparison of our results with the estimation results considering the sound speed gradients provided by the Japan Coast Guard (Yokota et al. 2018b). The estimation errors of the array positions by Method1 and Method2 are summarized in Additional file 1: Table S3, and the estimation errors, skewness, and kurtosis of the array positions by Method3 and Method4 are summarized in Additional file 1: Tables S4 and S5, respectively.
Figure 5a and Additional file 1: Figs. S6a–S10a show timeseries of the array positions estimated by Methods 1–4. We then estimated a fitting function of the array positions for each component and each method \(d(t)\), where \(t\) is the time after the 2011 Tohoku–Oki earthquake, using the following equation:
where \({a}_{0}\), \({a}_{1}\), and \({a}_{2}\) are estimation parameters, and \(\tau\) is the relaxation time (\(\tau =0.787\) year, which was obtained by Fujiwara et al. 2022). The GNSSA sites used in this study have showed significant contribution of the postseismic deformation of the 2011 Tohoku–Oki earthquake. Fujiwara et al. (2022) demonstrated that postseismic displacements of the onshore GNSS sites can be modeled using duallogarithmic functions, and Watanabe et al. (2021) demonstrated that those of the offshore GNSSA sites can be modeled using the single logarithmic function. Here, we estimated the fitting function using the single logarithmic function and assessed the robustness of the estimated array positions from the RMSs. The estimated fitting curves are shown in Fig. 5a and Additional file 1: Figs. S6a–10a, and the RMSs are summarized in Additional file 1: Table S6.
Figure 5a, Additional file 1: Figs. S6a–S10a and Additional file 1: Table S6 clearly demonstrate that Method2 and Method3 provided temporally stable solutions (the RMSs are ~3 cm) compared to Method1 (the RMS is ~5 cm) and Method4 (the RMS is ~4 cm) in the horizontal components. Moreover, there is no clear difference for the updown component among the estimation methods. These results suggest that the accuracy of our new method (Method3) is compatible with that of Method2, suggesting that appropriate consideration of the sound speed gradients is important for obtaining robust solutions. Note that Method3 is also compatible with Yokota et al. (2018b) (Additional file 1: Fig. S11). Although the estimation errors in Method1 (~0.4 cm in Additional file 1: Table S3) are generally smaller than those in Method2 (~0.9 cm in Additional file 1: Table S3) and Method3 (~0.6 cm in Additional file 1: Table S4), Method1 provided relatively lower accuracy than the other methods (Additional file 1: Table S6). This suggests that the solutions of Method1 contain the systematic errors due to the horizontal heterogeneity of the sound speed structure and failed to evaluate them in the estimation errors shown in Additional file 1: Table S3. Since Method2 and Method3 consider the sound speed gradients, their estimation errors are generally larger than Method1. Although Method4 considers the sound speed gradients, the estimation errors of Method4 are comparable with Method1 (~0.4 cm in Additional file 1: Table S5). This is because the intensity of the sound speed gradients (i.e., the gradient depth) was fixed.
Here, we examined the assumption of Method3 that the deep gradient is linked to the shallow gradient. Figure 5b shows the relationship between the shallow gradients obtained as the initial values for Method3 \({\tilde{\mathbf{g}}}_{\mathrm{s}}^{\mathrm{ini}}\) and the deep gradients estimated by Method2 for all campaigns at KAMN. Figure 6 shows a compilation of the shallow and deep gradient scatter plots of for all sites (Figure 5b and Additional file 1: Figs. S6b–S10b). Note that the shallow gradients \({\tilde{\mathbf{g}}}_{\mathrm{s}}^{\mathrm{ini}}\) are obtained independently of the deep gradients. If the assumption was appropriate, these parameters should show positive correlation, as expressed by Eq. 26 (Honsho et al. 2019). Then, we calculated correlation coefficients for these scatter plots. As a result, the positive correlation was found for all sites (correlation coefficients of ~0.36–0.86). Moreover, we also calculated a linear regression line for each scatter plot without including an intercept. Then, we obtained significant positive proportional relationship between the shallow and the deep gradients (proportional coefficients of ~0.11–0.46, which corresponds to the gradient depths of ~0.22–0.92 km as derived from Eq. 26). Since the gradient depth is not a constant value among the campaigns, the scatter plots do not necessarily demonstrate strict linear relationship. Therefore, each proportional coefficient might represent an average gradient depth for the corresponding site. These findings show the positive correlation between the shallow and the deep gradient and support the validity of the assumption of Method3 on actual observational data. Furthermore, we confirmed that the shallow gradients estimated by Method3 were almost same with their initial values \({\tilde{\mathbf{g}}}_{\mathrm{s}}^{\mathrm{ini}}\); thus, the shallow gradients can be constrained by Eq. (29) independently of the assumption.
We then examined the degree of the gradient depth in the actual ocean from the estimation results. Figure 5c and Additional file 1: Figs. S6c–S10c show the gradient depths estimated by Method3 and the equivalent gradient depths calculated from the deep gradients of Method2. The equivalent gradient depth can be obtained by the following process: we first calculated a gradient depth for each horizontal component using Eq. 26 with the shallow gradient of Method3 and the deep gradient of Method2; we then calculated the mean of these two gradient depths weighted by the intensity of the shallow gradients (i.e., absolute values of the EW and NS shallow gradients) as the equivalent gradient depth. In these figures, the equivalent gradient depths generally coincide with the gradient depths of Method3, suggesting the validity of the linear relationship between shallow and deep gradients. The equivalent gradient depths for most campaigns fall within the estimation range of the gradient depth in Method3 which was introduced as a physical restriction (i.e., from zero to a water depth). Meanwhile, the equivalent gradient depths for several campaigns were outside this physically restricted range. These campaigns are indicated by the cyan and orange vertical lines in Figure 5a and Additional file 1: Figs. S6a–S10a, which indicate campaigns with an equivalent gradient depth lower than zero and larger than the water depth, respectively. The equivalent gradient depth presupposes a simplified sound speed structure (uniform sound speed gradient from the seasurface to the gradient depth). Therefore, it would occasionally demonstrate an unphysical depth when a complex sound speed structure exists. If a strong sound speed heterogeneity exists only in a deep portion, the equivalent gradient depth would exceed the water depth. If directions of the shallow and deep sound speed heterogeneities are remarkably different, the equivalent gradient depth would be negative. When the equivalent gradient depth was lower than zero, the absolute values of the gradient depth of Method3 and the equivalent gradient depth were generally small, that is, the deep gradient was estimated to be small. In this case, the influence of the sound speed gradient on the positions was small, and the difference between Method2 and Method3 was also small. However, when the equivalent gradient depth was larger than the water depth, differences of up to 10 cm between Method2 and Method3 were occasionally found. Because the intensity of shallow gradients among these campaigns was generally low (color of Figure 5c and Additional file 1: Figs. S6c–S10c), the sound speed structures of these campaigns might be exceptional, which the deep gradients were much larger than the shallow gradients. Although such differences were found between Method2 and Method3, it was difficult to determine whether Method2 or Method3 was more plausible based on the stability of the time series of the array displacements.
From the above findings, our assumption used in Method3 was appropriate; accuracy of Method3 was compatible with that of Method2, which directly estimates the deep gradients. Figure 5a and Additional file 1: Figs. S6a–S10a show Method4, which considered the sound speed gradient with a fixed gradient depth and corresponds to the methods of some previous studies (e.g., Yasuda et al. 2017; Honsho et al. 2019; Kinugasa et al. 2020). Method4 generally demonstrated better solutions than Method1 did; however, these solutions are much more unstable than the solutions of Method2 and Method3. This suggested that the gradient depths should be optimized for each campaign. However, Method4 showed improved solutions as along with Method2 and Method3 at the CHOS site (Additional file 1: Fig. S10a). As shown in Additional file 1: Figs. S10b and S10c, the sound speed sloping structure at this site was almost constant among the campaigns, and the gradient depth was also constantly determined as approximately 0.5–1 km. Moreover, number of the equivalent gradient depth under unphysical range is quite smaller than the other site. Accordingly, the fixed gradient depth (0.6 km) in Method4 worked well for this site. This constant soundspeed sloping structure might have been caused by the Kuroshio Current, which is constantly present over this site, and a complex sound speed structure might not tend to be caused compared with the other sites. Therefore, Method4 was effective for such exceptional sites or when a gradient depth could be derived from other underwater measurements (Kinugasa et al. 2020), but it is plausible to optimize the gradient depth for diverse cases.
Application to campaign data collected by Tohoku University
Figure 7 summarizes the GNSSA positioning results of the campaign on November 13, 2015, at the G20 site estimated by Methods1–3. The pairplots of the posterior PDFs estimated by Method3 are shown in Additional file 1: Fig. S12. Including this campaign survey, a survey track for each site of Tohoku University was composed of a shortterm moving survey (e.g., during approximately 0–2 h in Fig. 7b) and a longterm point survey (e.g., during approximately 2–14 h in Fig. 7b). As shown in Figure 7a, b, f, g, the shallow gradient was determined from the shortterm moving survey data, with a direction of northwest. In this campaign, the gradient depth was roughly determined as a Gaussian distribution (Fig. 7h), and the differences in the array displacements between Method1 and Method3 reached several centimeters. Method2 also provided similar solutions as Method3; thus, such small amounts of moving data had the potential to detect a soundspeed sloping structure.
Figures 8, 9, and 10 show the time series of the array displacements and seasurface tracks for campaigns at G20, G17, and G10, respectively. We conducted moving surveys for all campaigns at G20 (Fig. 8). Although the moving survey data were expected to be useful for estimating a soundspeed sloping structure and vertical motions, the solution of TB1201 showed a clear bias in the vertical component using all methods and a large bias in the horizontal component using Method2. These anomalies of TB1201 at G20 were considered to have occurred because of the low signaltonoise ratio of the acoustic data in this campaign. Aside from these anomalies, Method2 demonstrated temporally irregular motions compared with Method3 in Figure 8, possibly meaning that Method2 failed to provide stable solutions for most campaign data of Tohoku University. As Honsho et al. (2019) pointed out, it is difficult to robustly solve a sound speed structure using Method2. In contrast, Method3 successfully provided robust solutions because of the following two factors: (1) Method3 practically reduced unknown parameters. Although Method2 estimated two gradient parameters for the soundspeed sloping structure: the deep gradients for the EW and NS components, Method3 estimated a single gradient parameter: the gradient depth (note that the shallow gradients were almost independently solved in Method3). (2) Method3 employed a constraint on the range of the gradient depth, which prevented physically unnatural solutions. However, it was difficult to estimate the gradient parameters even with Method3 when a moving survey was lacking. No moving surveys were conducted during campaigns of D3K1501, YK1602, KS1614, and KS1713 at G17 (campaigns with vertical gray lines in Fig. 9a). Among these campaigns, the solutions assuming a soundspeed sloping structure (Methods2, 3, and 4) were temporally violated; thus, these methods were not appropriate for data application without a moving survey. For these cases, the conventional positioning method assuming a horizontally stratified sound speed structure (Method1) was appropriate for obtaining stable positions.
Finally, we presented a special case to detect a soundspeed sloping structure at a site with multiangled transponders, for example, at G10 (Fig. 10). Tohoku University has adopted multiangled transponder geometry at some GNSSA sites, allowing us to obtain acoustic data with multiple shot angles even through a point survey (Tomita et al. 2019). The GNSSA site with multiangled transponders could theoretically solve vertical motions and gradient parameters without moving survey data (e.g., Kido 2007; Tomita et al. 2019; Matsui et al. 2019). Because of this benefit, Methods2 and 3 successfully provided robust solutions, even for campaigns without moving survey data (campaigns with vertical gray lines in Fig. 10a). Moreover, unlike G17 and G20, Method2 and Method3 provided temporally stabilized motions for most campaigns. This also reflected the utility of multiangled transponders. Considering that it was naturally impossible to properly estimate shallow gradients without moving survey data, Methods3 and 4 did not work well among these campaigns, even if these methods numerically solved the gradient parameters. Therefore, Method2 was the most appropriate method for processing the observation data of multiangled transponder sites. However, as claimed by Matsui et al. (2019), a long observational period is required to eliminate the contributions of a shallow soundspeed gradient using only point survey data at a multiangled transponder site; therefore, if the observational time was not sufficiently long, Method1 was preferred.
Conclusions and discussion
In this study, we developed an approximate calculation technique for rapidly calculating travel times, and then, using this technique, we developed a new GNSSA array positioning method using an MCMC technique. In the approximate travel time calculation technique, we first calculated the approximate travel time by a simple geometric computation and then corrected it using the 8th degrees of polynomial functions of an angular distance and the height of a seasurface platform. The correction terms must be optimized in advance suing numerical simulations. Conventional travel time calculation techniques in a horizontally stratified sound speed structure which consider Snell’s law were exact but required too long a calculation time to perform the MCMC technique. The approximate calculation technique can rapidly provide travel times (the new technique reduces the computation time to ~ 270 times less than the conventional technique) with enough accuracy and precision in array position estimation. The computational cost of the MCMCbased array positioning method depends on the data quantity; however, for example, it took 30 min to perform our MCMC method with \(5\times {10}^{5}\) iterations for the actual observational data on August 1, 2014 at KAMN (the results are shown in Fig. 4).
The MCMCbased array positioning method estimated the gradient depth and array positions, assuming a horizontally sloping sound speed structure. In a horizontally sloping sound speed structure, shallow and deep gradients could be defined, and the deep gradient provided a significant bias in the horizontal array position. In our method, we assumed that a sound speed gradient was constantly present from the sea surface to the gradient depth, that is, the deep gradient was proportional to the shallow gradient, and the gradient depth corresponded to a proportionality coefficient. It was plausible to directly estimate the deep gradient using the amount of moving survey data (e.g., Yokota et al. 2018a; Watanabe et al. 2021); however, when the moving survey data were insufficient, the pseudo deep gradient was obtained from the shallow gradient through this assumption (e.g., Yasuda et al. 2017; Honsho et al. 2019; Kinugasa et al. 2020). In previous studies, the generality of this assumption was not wellinvestigated, and optimization of the gradient depth was not performed.
We applied the MCMCbased array positioning and the conventional array positioning methods on the actual observational data collected by the Japan Coast Guard, which contain sufficient moving survey data. The following findings were obtained:

1.
Our new method generally demonstrated Gaussian distributions for PDFs of unknown parameters, whereas the PDF of gradient depth occasionally showed a truncated Gaussian distribution. Namely, the array positioning could be performed using the framework of the leastsquares method; however, it was plausible to perform an MCMC technique when the range of an unknown parameter could be physically restricted, as was the case of the gradient depth.

2.
Our new method provided robust solutions for the array positions compared with those estimated by a conventional array positioning method by assuming a horizontally stratified sound speed structure; it was comparable with the conventional method that directly estimated the deep gradient assuming a horizontally sloping soundspeed structure.

3.
The deep gradients generally showed a proportional relationship with shallow gradients. Thus, our assumption that a sound speed gradient exists from the sea surface to the gradient depth was valid.
Next, we applied our method to actual observational data collected by Tohoku University, which contained small amounts of moving survey data or no moving survey data. The following findings were found:

1.
When compared to the traditional array positioning method assuming a horizontally stratified sound structure, our new method stably obtained a truncated Gaussian distribution as the PDF of the gradient depth and demonstrated that a temporally robust timeseries of the array positions was obtained in the case of observational data containing small amounts of moving survey data. Therefore, small amounts of moving survey data had the potential to detect horizontal gradients to a certain degree.

2.
It was difficult to directly estimate the deep gradient using the conventional method, assuming a horizontally sloping sound speed structure from small amounts of moving survey data. In contrast, our new method made more robust estimations than the conventional method did, because it practically reduced the numbers of unknown parameters modeling the horizontally sloping sound speed structure, and also due to the introduction of physical restriction on the gradient depth.

3.
Even with our new method, we could not detect a horizontally sloping sound speed structure from actual observational data without moving survey data. In this case, the conventional array positioning method is preferred, assuming a horizontally stratified sound structure. However, it should be noted that deep gradients could be directly estimated for a multiangled transponder site if the observational time is sufficiently long.
As the above findings show, our new method successfully optimizes the gradient depth, and is, therefore, an effective technique to model an underwater sloping sound speed structure and robustly obtains array positions, except in the case in which no moving survey data is available.
Availability of data and materials
The GNSSacoustic observational data collected by the Japan Coast Guard at KAMN, KAMS, MYGI, MYGW, FUKU, and CHOS have been distributed in Zenodo, as shown by Watanabe et al. (2021). Array displacements at KAMN, KAMS, MYGI, MYGW, FUKU, and CHOS were distributed on the homepage of the Japan Coast Guard by Yokota et al. (2018b). The GNSSacoustic observational data collected by Tohoku University for G10, G17, and G20 are available from the authors upon request. The GNSSA positioning programs (Methods 1–3 in “Application of the MCMCbased array positioning method” section; they are written in Julia) are available from the authors upon request.
Abbreviations
 ABIC:

Akaike Bayesian information criterion
 BIC:

Bayesian information criterion
 GNSS:

Global Navigation Satellite System
 GNSSA:

GNSSacoustic
 MCMC:

Markov Chain Monte Carlo
 NLLS:

Nonlinear leastsquares
 NTD:

Nadir total delay
 PDF:

Probability density function
References
Akaike H (1980) Likelihood and the Bayes procedure. In: Bernardo JM, DeGroot MH, Lindley DV, Smith AFM (eds) Bayesian statistics. University Press, London, pp 143–166
Bayes T (1763) An essay towards solving a problem in the doctrine of chances. Phil Trans R Soc 53:370–418. https://doi.org/10.1098/rstl.1763.0053
Chadwell CD, Sweeney AD (2010) Acoustic raytrace equations for seafloor geodesy. Mar Geod 33:164–186. https://doi.org/10.1080/01490419.2010.492283
Fujita M, Ishikawa T, Mochizuki M, Sato M, Toyama S, Katayama M (2006) GPS/acoustic seafloor geodetic observation: method of data analysis and its application. Earth Planets Space 58:265–275. https://doi.org/10.1186/BF03351923
Fujiwara S, Tobita M, Ozawa S (2022) Spatiotemporal functional modeling of postseismic deformations after the 2011 TohokuOki earthquake. Earth Planets Space 74:13. https://doi.org/10.1186/s40623021015680
Fukuda J, Johnson KM (2008) A fully Bayesian inversion for spatial distribution of fault slip with objective smoothing. Bull Seis Soc Am 98(3):1128–1146. https://doi.org/10.1785/0120070194
Hastings WK (1970) Monte Carlo sampling methods using Markov chains and their applications. Biometrika 57:97–109. https://doi.org/10.1093/biomet/57.1.97
Honsho C, Kido M (2017) Comprehensive analysis of traveltime data collected through GPSacoustic observation of seafloor crustal movements. J Geophys Res 122:8583–8599. https://doi.org/10.1002/2017JB014733
Honsho C, Kido M, Tomita F, Uchida N (2019) Offshore postseismic deformation of the 2011 Tohoku earthquake revisited: application of an improved GPSacoustic positioning method considering horizontal gradient of sound speed structure. J Geophys Res 124:5990–6009. https://doi.org/10.1029/2018JB017135
Honsho C, Kido M, Ichikawa T, Ohashi T, Kawakami T, Fujimoto H (2021) Application of phaseonly correlation to traveltime determination in GNSSacoustic positioning. Front Earth Sci 9:600732. https://doi.org/10.3389/feart.2021.600732
Iinuma T, Hino R, Kido M, Inazu D, Osada Y, Ito Y, Ohzono M, Tsushima H, Suzuki S, Fujimoto H, Miura S (2012) Coseismic slip distribution of the 2011 off the Pacific Coast of Tohoku Earthquake (M9.0) refined by means of seafloor geodetic data. J Geophys Res 117:B07409. https://doi.org/10.1029/2012JB009186
Ikuta R, Tadokoro K, Ando M, Okuda T, Sugimoto S, Takatami K, Yada K, Besana GM (2008) A new GPSacoustic method for measuring ocean floor crustal deformation: application to the Nankai Trough. J Geophys Res 113:B02401. https://doi.org/10.1029/2006JB004875
Kido M (2007) Detecting horizontal gradient of sound speed in ocean. Earth Planets Space 59:33–36. https://doi.org/10.1186/BF03352027
Kido M, Fujimoto H, Miura S, Osada Y, Tsuka K, Tabei T (2006) Seafloor displacement at Kumanonada caused by the 2004 off Kii Peninsula earthquakes, detected through repeated GPS/Acoustic surveys. Earth Planets Space 58:911–915. https://doi.org/10.1186/BF03351996
Kido M, Osada Y, Fujimoto H, Hino R, Ito Y (2011) Trenchnormal variation in observed seafloor displacements associated with the 2011 TohokuOki earthquake. Geophys Res Lett 38:L24303. https://doi.org/10.1029/2011GL050057
Kinugasa N, Tadokoro K, Kato T, Terada Y (2020) Estimation of temporal and spatial variation of sound speed in ocean from GNSSA measurements for observation using moored buoy. Prog Earth Planet Sci 7:21. https://doi.org/10.1186/s40645020003315
Kubo H, Asano K, Iwata T, Aoi S (2016) Development of fully Bayesian multipletimewindow source inversion. Geophys J Int 204(3):1601–1619. https://doi.org/10.1093/gji/ggv540
Locarnini RA, Mishonov AV, Baranova OK, Boyer TP, Zweng MM, Garcia HE, Reagan JR, Seidov D, Weathers K, Paver CR, Smolyar I (2018) World Ocean Atlas 2018, volume 1: temperature. A. Mishonov Technical Ed., NOAA Atlas NESDIS, vol 81, p 52
Matsui R, Kido M, Niwa Y, Honsho C (2019) Effects of disturbance of seawater excited by internal wave on GNSSacoustic positioning. Marine Geophys Res 40:541–555. https://doi.org/10.1007/s11001019093946
Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller E (1953) Equation of state calculations by fast computing machines. J Chem Phys 21:1087–1092. https://doi.org/10.1063/1.1699114
Sato M, Ishikawa T, Ujihara N, Yoshida S, Fujita M, Mochizuki M, Asada A (2011) Displacement above the hypocenter of the 2011 TohokuOki earthquake. Science 332(6036):1395. https://doi.org/10.1126/science.1207401
Schwartz G (1978) Estimating the dimension of a model. Ann Statist 6(2):461–464. https://doi.org/10.1214/aos/1176344136
Spiess FN (1985) Suboceanic geodetic measurements. IEEE Trans Geosci Rem Sens GE23(4):502–510. https://doi.org/10.1109/TGRS.1985.289441
Spiess FN, Chadwell CD, Hildebrand JA, Young LE, Purcell GH Jr, Dragert H (1998) Precise GPS/acoustic positioning of seafloor reference points for tectonic studies. Phys Earth Planet Int 108(2):101–112. https://doi.org/10.1016/S00319201(98)000892
Tomita F, Kido M, Osada Y, Hino R, Ohta Y, Iinuma T (2015) First measurement of the displacement rate of the Pacific Plate near the Japan Trench after the 2011 TohokuOki earthquake using GPS/acoustic technique. Geophys Res Lett 42:8391–8397. https://doi.org/10.1002/2015GL065746
Tomita F, Kido M, Ohta Y, Iinuma T, Hino R (2017) Alongtrench variation in seafloor displacements after the 2011 Tohoku earthquake. Sci Adv 3(7):e1700113. https://doi.org/10.1126/sciadv.1700113
Tomita F, Kido M, Honsho C, Matsui R (2019) Development of a kinematic GNSSacoustic positioning method based on a statespace model. Earth Planets Space 71:102. https://doi.org/10.1186/s406230191082y
Tomita F, Iinuma T, Agata R, Hori T (2021) Development of a transdimensional fault slip inversion for geodetic data. J Geophys Res 126:e2020JB020991. https://doi.org/10.1029/2020jb020991
Watanabe S, Ishikawa T, Yokota Y, Nakamura Y (2020) GARPOS: analysis software for the GNSSA seafloor positioning with simultaneous estimation of sound speed structure. Front Earth Sci 8:597532. https://doi.org/10.3389/feart.2020.597532
Watanabe S, Ishikawa T, Nakamura Y, Yokota Y (2021) Co and postseismic slip behaviors extracted from decadal seafloor geodesy after the 2011 Tohokuoki earthquake. Earth Planets Space 73:162. https://doi.org/10.1186/s40623021014870
Wessel P, Smith WHF (1998) New, improved version of generic mapping tools released. Eos Trans AGU 79(47):579. https://doi.org/10.1029/98eo00426
Yabuki T, Matsu’ura M (1992) Geodetic data inversion using a Bayesian information criterion for spatial distribution of fault slip. Geophys J Int 109(2):363–375. https://doi.org/10.1111/j.1365246X.1992.tb00102.x
Yasuda K, Tadokoro K, Taniguchi S, Kimura H, Matsuhiro K (2017) Interplate locking condition derived from seafloor geodetic observation in the shallowest subduction segment at the Central Nankai Trough, Japan. Geophys Res Lett 44:3572–3579. https://doi.org/10.1002/2017GL072918
Yokota Y, Ishikawa T (2020) Shallow slow slip events along the Nankai Trough detected by GNSSA. Sci Adv 6(3):eaay5786. https://doi.org/10.1126/sciadv.aay5786
Yokota Y, Ishikawa T, Watanabe S, Tashiro T, Asada A (2016) Seafloor geodetic constraints on interplate coupling of the Nankai Trough megathrust zone. Nature 534:374–377. https://doi.org/10.1038/nature17632
Yokota Y, Ishikawa T, Watanabe S (2018a) Gradient field of undersea sound speed structure extracted from the GNSSA oceanography. Mar Geophys Res 40:493–504. https://doi.org/10.1007/s1100101893627
Yokota Y, Ishikawa T, Watanabe S (2018b) Seafloor crustal deformation data along the subduction zones around Japan obtained by GNSSA observations. Sci Data 5:180182. https://doi.org/10.1038/sdata.2018.182
Zweng MM, Reagan JR, Seidov D, Boyer TP, Locarnini RA, Garcia HE, Mishonov AV, Baranova OK, Weathers K, Paver CR, Smolyar I (2018) World Ocean Atlas 2018, volume 2: salinity. A. Mishonov Technical Ed., NOAA Atlas NESDIS, vol 82, p 50
Acknowledgements
The figures were generated using Generic Mapping Tools software (Wessel and Smith 1998). In addition, we thank Editage (www.editage.jp) for English language editing.
Funding
This study was supported by Japan Society for the Promotion of Science (JSPS) KAKENHI (Grant No. 20K14588).
Author information
Authors and Affiliations
Contributions
FT carried out all data analysis and wrote the manuscript. MK provided basic concept of the approximate traveltime calculation, contributed to discussion and interpretation, and improved the developed technique and the manuscript. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
No applicable.
Consent for publication
No applicable.
Competing interests
The author declares that they have no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Additional file 1. Fig. S1
: Flowchart of an MCMC procedure in this study. Fig. S2: Schematic image of the numerical simulation in “Method: MCMCbased GNSSA positioning method” section. Fig. S3: Track of the observational vessel and the sound speed profile used in “Method: MCMCbased GNSSA positioning method” section. Fig. S4: Distribution of the GNSSAcoustic sites employed in this study. Fig. S5: Pairplots of the posterior PDFs at KAMN on August 1, 2014. Fig. S6: Array positioning results at KAMS. Fig. S7: Array positioning results at MYGI. Fig. S8: Array positioning results at MYGW. Fig. S9: Array positioning results at FUKU. Fig. S10: Array positioning results at CHOS. Fig. S11: Time series of the array position at Japan Coast Guard sites. Fig. S12: Pairplots of the posterior PDFs at G20 on November 13, 2015. Table S1: Numerical simulation results for the approximate traveltime calculation model. Table S2: Positioning results in “Method: MCMCbased GNSSA positioning method” section. Table S3: Standard deviation of the array positions estimated by Method1 and Method2. Table S4: Standard deviation, skewness, and kurtosis of the array positions estimated by Method3. Table S5: Standard deviation, skewness, and kurtosis of the array positions estimated by Method4. Table S6: RMSs between the estimated array positions and the logarithmic fitting function.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Tomita, F., Kido, M. An approximate travel time calculation and a robust GNSSacoustic positioning method using an MCMC technique. Earth Planets Space 74, 176 (2022). https://doi.org/10.1186/s40623022017400
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s40623022017400