Instruments and Methods for Acoustic and Visual Survey of Manganese Crusts | IEEE Journals & Magazine | IEEE Xplore

Instruments and Methods for Acoustic and Visual Survey of Manganese Crusts


Abstract:

This paper describes acoustic and visual instruments developed to perform high-resolution surveys of the volumetric distribution of manganese crusts from an underwater ve...Show More

Abstract:

This paper describes acoustic and visual instruments developed to perform high-resolution surveys of the volumetric distribution of manganese crusts from an underwater vehicle. The instruments consist of an acoustic device, developed to perform in situ measurements of manganese crust thickness at depths of up to 3000 m, and a vision-based mapping system that generates 3-D color reconstructions of the seafloor. Methods to process the information obtained by these sensors to automatically identify areas of exposed crust using the 3-D reconstructions, and subsequently determine the thickness of the crusts based on the acoustic measurements, are described. Sea trials were performed at #5 Takuyo seamount with the systems mounted onboard the remotely operated vehicle Hyper-Dolphin during the NT10-11 cruise of the R/V Natsushima. The results are that the first time in situ measurements of manganese crust thickness have been performed, and it is demonstrated that, for the types of substrate dominant in the surveyed area, continuous acoustic measurement of manganese crust thickness is possible. The work described in this paper indicates that the proposed instruments and data processing algorithms can form useful tools to enable more efficient survey of manganese crusts.
Published in: IEEE Journal of Oceanic Engineering ( Volume: 38, Issue: 1, January 2013)
Page(s): 186 - 203
Date of Publication: 19 October 2012

ISSN Information:


SECTION I.

Introduction

Manganese crusts occur throughout the Pacific on guyots and seamounts in areas that have been kept free of sedimentation for several millions of years [1]. The crusts are generally found at depths of between 800 and 2500 m, and are formed by oxidized minerals that precipitate out of the ambient sea water to form a thin layer of deposit that can extend over several tens of kilometers. The crusts form a layer of black deposit on top of a hard-rock substrate, as seen in Fig. 1, with growth rates ranging from 4 to 7 mm/Ma depending on gravity processes, pH, sediment cover, and ocean currents [2]. The crusts consist primarily of manganese and iron oxides, but are rich in Co, Ni, and Cu and contain traces of Pt, making them the focus of attention as a potential target for mining [3], [4]. However, manganese crust coverage can vary from full to none, often with abrupt transitions that can occur on the scale of tens of meters, with intermittent regions of sediment and cluttering by debris rock and to a certain extent marine life. Even in areas with continuous exposed crusts, the thickness of the crust layers can vary anywhere between 10 and 250 mm, and to date, the only way to measure their thickness has been through sampling. The objective of this study is to provide a solution for instrumentation to enable more efficient survey of manganese crusts. For this, we use an acoustic probe [5] that has been developed specifically to perform in situ measurements of manganese crust thickness, and employ a visual approach to mapping the seafloor [6] to identify areas of exposed crust, since one of the contrasting features of the crusts is their color. It is proposed that these systems can be operated at low altitudes, using underwater platforms such as autonomous underwater vehicles (AUVs) or remotely operated vehicles (ROVs) to maintain a close range to the seafloor, to continuously map the distribution and thickness of exposed manganese crusts on regional scales extending over several tens of kilometers.

Fig. 1. - Photo of crust deposits taken at #5 Takuyo seamount during NT10-11. The image was taken at a depth of 1426 m on the shoulder of the seamount, and shows a layer of crust on top of a rock substrate.
Fig. 1.

Photo of crust deposits taken at #5 Takuyo seamount during NT10-11. The image was taken at a depth of 1426 m on the shoulder of the seamount, and shows a layer of crust on top of a rock substrate.

This paper is organized as follows: in Sections II and III, we discuss the feasibility of acoustic measurement of crust thickness for different types of rock substrate by measuring the acoustic properties of samples obtained from #5 Takuyo seamount, a large oceanic guyot located in the northwest Pacific, as shown in Fig. 2. In Section IV, we describe a 3000-m depth rated acoustic probe developed to continuously measure the thickness of the crusts, together with an algorithm that processes the data it obtains. However, acoustic measurements are only useful where crusts are known to be present. To determine this, we introduce a 3-D seafloor mapping device, described in Section V, which generates bathymetric reconstructions of the seafloor in actual colors, together with an algorithm to classify regions of the seafloor based on signatures extracted from their visual and bathymetric features, which is described in Section VI. This enables the acoustic measurements to be interpreted appropriately based on the type of seafloor where they were made, and ultimately determine the volumetric distribution of exposed manganese crusts in the surveyed area. The developed instruments were mounted onboard the ROV Hyper-Dolphin of the Japan Agency for Marine-Earth Science and Technology (JAMSTEC), and were deployed at #5 Takuyo seamount during the NT10-11 cruise of the R/V Natsushima. During the survey, the ROV was operated at depths between 1000 and 3000 m at low altitudes of <1.0 m, with surge velocities between 0.2 and 0.3 m/s to survey the area both acoustically and visually. The results obtained during the cruise demonstrate that for the types of substrate rock dominant in the surveyed area, in situ acoustic measurement of manganese crust thickness is possible. The results indicate that the proposed instruments and the algorithms implemented to process their measurements are useful tools to enable more efficient surveys of manganese crusts.

Fig. 2. - #5 Takuyo seamount is an oceanic guyot located at 23° 00′ N, 153° 20′ E in the northwest Pacific. It has a plateau that extends over 2000 km2 at a depth of 1000 m, more than 4000 m above the abyssal plain.
Fig. 2.

#5 Takuyo seamount is an oceanic guyot located at 23° 00′ N, 153° 20′ E in the northwest Pacific. It has a plateau that extends over 2000 km2 at a depth of 1000 m, more than 4000 m above the abyssal plain.

SECTION II.

Problem Formulation

To date, the only way to determine the thickness of manganese crusts has been through sampling. Although sampling is reliable and enables analysis of the chemical composition of the crusts, for the purpose of measuring crust thickness, it is time consuming and the obtained information is discrete. Though several techniques have been applied, including bathymetric multibeam, sidescan sonar imaging, and acoustic sub-bottom mapping, in an effort to increase the efficiency of crust surveys, these methods do not have sufficient resolution to determine the distribution or resolve the thickness of exposed crusts [7]–​[9]. Subsurface acoustic measurements with resolutions of <1 cm could potentially allow for continuous information concerning crust thickness to be obtained and enable wider areas of crust deposit to be more efficiently surveyed. However, acoustic measurements are only useful where crusts are present. For this, visual images can be helpful in determining whether an area of the seafloor has exposed crusts [8]–​[10], since one of the contrasting features of manganese deposits is their color. By combining acoustic and visual methods, we aim to make measurements that relate directly to the survey objective, which is to determine the distribution and thickness of the crusts, and so obtain data that lends itself more directly to meaningful scientific interpretation. In this section, we discuss some of the considerations behind the development of acoustic and visual instruments to realize this goal.

An acoustic reflection occurs at any boundary where there is a change in acoustic impedance. Therefore, assuming that the compressional wave velocity within an object is known, and that there is some contrast in acoustic impedance between the target and the media on either side of it, it is possible to compute the target's thickness by measuring the time of flight between the acoustic reflections that take place at its top and bottom surfaces. However, manganese crusts grow on a variety of substrates, and, subsequently, the level of contrast in impedance can vary significantly between different regions [11], [12]. For certain types of substrate, there may be insufficient contrast in impedance for detectable reflections to occur at the crust–substrate interface, and acoustic measurement of crust thickness may not be possible. However, for other types of substrate, sufficient contrast can be expected for acoustic measurements to be possible. Even in these cases, measurements are still complicated by seafloor roughness and the fact that the crust layers being measured are typically only a few centimeters thick. With regard to the physical geometry of the crusts, the acoustic devices used in most sub-bottom applications are designed to detect features several meters beneath the seafloor, typically operating at low frequencies of around 2–20 kHz to achieve deep penetration [7] [9] [8]. As such, they do not have sufficient resolution to discretize layers only a few centimeters thick. Further complications arise due to the roughness of the seafloor, which exists on various scales, with bumps of similar or greater magnitude than the thickness of the crusts themselves. For beams with a large acoustic footprint, the high energy reflections from the undulating crust surface may coincide in time with the weaker reflections from the crust–substrate interface. When subject to these conditions, low-frequency signals do not have the necessary spatial or temporal resolution to resolve the thickness of thin, bumpy crusts. The authors propose to overcome these issues by operating a high-frequency transducer with a small acoustic footprint at low altitudes of <1.0 m off the seafloor using an underwater vehicle to maintain a close range to the target.

Manganese deposits can occur in the form of crusts and nodules, in environments that can be complex with intermittent regions of sediment, cluttering by debris rock and although sparse, be populated by marine life. While visual images can be useful to determine the presence of exposed crusts, 2-D images suffer from geometric distortions due to perspective effects, and in the presence of pronounced reliefs or on steep slopes, produce artifacts that result in inherently inaccurate estimates of spatial distribution [13]. In such cases, a method that handles the images in a fully 3-D way from input to output is required to produce accurate results. Approaches respecting this condition include stereo vision [14], structure from motion (SFM) [15], [16], and structured light [6]. In this work, we implement a simple method based on structured light that uses laser profiling and images of the seafloor to generate bathymetric reconstructions in actual colors by assigning red, green, and blue (RGB) values on a pixel-based level. The 3-D reconstructions generated using this technique can be used to confirm the presence of exposed crusts and accurately map their distribution on the seafloor. The visual instruments required are mounted on an underwater vehicle, and measurements are made in parallel with those of the acoustic probe. This is not limiting since the proposed acoustic system already requires the vehicle to operate at ranges suitable for visual survey.

While it is possible to map the measurements made by the acoustic and visual instruments just described, further analysis of the data is necessary to interpret these measurements in a scientifically meaningful way. This stage of analysis can be complex and require the flexibility of human judgement. However, in many cases, the analysis is repetitive and time consuming, and so we seek to increase efficiency by automating some of the more simple tasks associated with interpretation of the data. In [17]–​[19], automatic classification was implemented on images and 3-D reconstructions taken at coral reefs to automatically highlight areas of interest for more detailed analysis by humans. In our application, the aim of the analysis can be clearly defined. We look to identify areas of the seafloor that have exposed crusts so that the measurements of the acoustic system can be processed to determine crust thickness, and equally importantly, prevent false conclusions being drawn from acoustic measurements made over areas that do not consist of exposed crusts, for example, areas covered by sediments. Even though acoustic measurements may not be valid in areas with nodular deposits, or areas cluttered by debris rock, these areas are still of interest and identification of such areas may be useful for further scientific analysis by humans. In cases where the cluttering is by pavement-like slabs of crust, acoustic measurements may still be valid. These situations require flexibility in judgement that is difficult to automate, and it is reasonable to highlight these areas, together with their acoustic measurements, and allow humans to judge whether these measurements should be taken into account. In this work, we develop a method to classify different regions of the seafloor using Gaussian mixture models (GMMs) generated based on signatures extracted from visual and bathymetric features, and use the results of the classification to load a specific set of instructions to interpret measurements made over the different types of seafloor in an appropriate way.

SECTION III.

Acoustic Characterization

To assess the feasibility of acoustic measurement, it is necessary to consider the acoustic properties of manganese crusts and their substrates. The acoustic properties required for assessment are the acoustic impedances of both the crusts and their substrates, and the attenuation coefficient of the crust layer. In [20], the acoustic properties of manganese nodules and crusts were measured, but no measurements of their substrates were performed. In [21], the acoustic impedances of both manganese crusts and their substrates were described. However, parameters such as the acoustic attenuation of the crusts were not measured. To determine the properties necessary for this study, manganese crusts and their substrates were sampled from #5 Takuyo seamount using the ROV Hyper-Dolphin, during the NT09-02 Leg 2 and NT10-11 cruises of the R/V Natsushima. #5 Takuyo seamount is a large oceanic guyot located in the northwest Pacific at 23° 00′ N, 153° 20′ E on the Pacific plate. The seamount is old, with a lithosphere dating back 150 Ma [22], and has a flat top that extends over 2000 km2 at a depth of 1000 m, more than 4000 m above the abyssal plain. During the cruises, samples were obtained from various locations, including the flat top, shoulder, and slope of the guyot between depths of 1000 and 3000 m. Of these, 11 crust and 13 substrate samples, some examples of which are shown in Fig. 3, were chosen for acoustic investigation. These were chosen to be representative of the manganese crust and substrate types observed during the dives with the ROV. The samples were cut and polished into rectangular blocks of either crust or substrate with flat, parallel surfaces for measurement of their acoustic properties. The compressional wave velocities $c_{i}$ and wet densities $\rho_{i}$ of both crust and substrate samples were measured so that their acoustic impedances $z_{i}$ could be computed [23]. In the case of the crusts, the acoustic attenuation coefficients $\alpha_{i}$ were also determined [24]. Fig. 4 shows the measured compressional wave velocities and densities for each sample, and Fig. 5 shows the corresponding acoustic impedances, where the substrates have been grouped into the following categories based on their geological characteristics and the factors that have been found to influence their acoustic properties: relatively unaltered basalt, moderately altered basalt, highly altered basalt, lightly phosphatized carbonate or carbonate matrix, and highly phosphatized carbonate or carbonate matrix. While a detailed geological assessment is beyond the scope of this work, it is clear that the geological characteristics of the substrates have a large influence on their acoustic properties and that the local geology of the region must be considered when performing an acoustic survey.

Fig. 3. - Images of some of the samples used in this investigation. The sample shown in (a) has a crust thickness of 60 mm, with a relatively unaltered basalt substrate, (b) has a highly altered basalt substrate with a 25-mm-thick layer of crust, (c) has a phosphatized carbonate matrix substrate with a 45-mm-thick layer of crust, and (d) has a lightly phosphatized carbonate matrix with a 45-mm-thick layer of crust.
Fig. 3.

Images of some of the samples used in this investigation. The sample shown in (a) has a crust thickness of 60 mm, with a relatively unaltered basalt substrate, (b) has a highly altered basalt substrate with a 25-mm-thick layer of crust, (c) has a phosphatized carbonate matrix substrate with a 45-mm-thick layer of crust, and (d) has a lightly phosphatized carbonate matrix with a 45-mm-thick layer of crust.

Fig. 4. - Compressional wave velocities and wet densities plotted for crust and substrate samples obtained between depths of 1000 and 3000 m from #5 Takuyo seamount.
Fig. 4.

Compressional wave velocities and wet densities plotted for crust and substrate samples obtained between depths of 1000 and 3000 m from #5 Takuyo seamount.

Fig. 5. - Acoustic impedances of crust and substrate samples obtained between depths of 1000 and 3000 m from #5 Takuyo seamount.
Fig. 5.

Acoustic impedances of crust and substrate samples obtained between depths of 1000 and 3000 m from #5 Takuyo seamount.

The properties of the crusts sampled were found to be relatively uniform, with an average density of 1920 kg/m3 with a standard deviation of 36 kg/m3. The average compressional wave velocity in the crusts was measured to be 2932 m/s with a standard deviation of 179 m/s. These translate to a mean acoustic impedance of 5.64×106 kg/m2s with a standard deviation of 0.42×106 kg/m2s. Fig. 6 shows the Rayleigh reflections [23] for a normally incident wave at the crust–substrate interface for substrate impedances between 1.485×106 kg/m2s, the impedance of water, and 20×106 kg/m2s. The figure illustrates that for substrates with an impedance identical to the crusts, no reflection would occur. However, it can be seen that the Rayleigh reflection increases sharply for a relatively small contrast in impedance, which is favorable for acoustic measurement.

Fig. 6. - Rayleigh reflections at the crust–substrate interface for various substrate impedances, assuming an average impedance of 5.64×106 kg/m2s for the crusts.
Fig. 6.

Rayleigh reflections at the crust–substrate interface for various substrate impedances, assuming an average impedance of 5.64×106 kg/m2s for the crusts.

To assess the feasibility of acoustic measurement, it is also necessary to consider attenuation within the crust layer. This was determined by placing a pair of parallel beam 1-MHz transducers a fixed distance apart in water, and measuring the strength of an acoustic signal with and without crust samples of known thicknesses placed between them. The average attenuation coefficient measured for the crusts was 6.37 dB/(MHz cm) with a 1.8-dB/(MHz cm) standard deviation, which is in close agreement with values determined in [20]. The acoustic properties of the crusts and substrates determined in this study are summarized in Table I.

Table I Acoustic Properties Of Manganese Crust And Substrate Samples Characterized In This Study
Table - Acoustic Properties Of Manganese Crust And Substrate Samples Characterized In This Study

Although there is an insufficient number of samples to draw any statistical conclusions, the data seem to suggest that the crusts are generally less dense and have lower compressional wave velocities than the substrates found in the surveyed area, which is in agreement with previous reports [11], [12]. The level of attenuation in the crusts is higher than in most water-saturated marine sediments [25], and this limits the range of frequencies that can be used for acoustic measurement, even for the relatively shallow depths of penetration required to measure crust thickness. The average impedance of the substrates was found to be 11.1×106 kg/m2s, suggesting that there is a strong contrast in impedance between the crusts and their substrates at the site, which is favorable for acoustic measurement. However, since the value of substrate impedance depends mainly on the type of rock, it is more meaningful to think in terms of the average impedance for each of the substrate rock types. The most common type of substrate sampled from the shoulder of #5 Takuyo seamount consists of carbonate or a carbonate matrix with varying degrees of phosphatization, which is consistent with the fact that guyots, such as #5 Takuyo seamount, were at one point in their history a reef supporting a large volume of marine life. For one of the substrate samples obtained during the cruises, a highly altered basalt with an impedance of 5.33×106 kg/m2s obtained from the slope of the seamount at a depth of 2700 m, the Rayleigh reflection would correspond to an energy loss of −31.1 dB, translating to <3% of the amplitude of the incident wave being reflected at the boundary. Under these conditions, it is highly unlikely that a detectable reflection would occur. However, the relatively unaltered basalt and phosphatized carbonate and carbonate matrix substrates have a strong contrast in impedance compared to the crusts, with impedances of 16.63×106 kg/m2s and 15.79×106 kg/m2s corresponding to Rayleigh reflections of −6.13 and −6.48 dB, respectively. These translate to strong acoustic reflections, with almost half (49% and 47%, respectively) of the amplitude of the incident wave being reflected. Furthermore, since a relatively small contrast in impedance yields a sharp increase in the Rayleigh reflection, strong reflections can also be expected for moderately altered basalts and lightly phosphatized carbonate and carbonate matrix substrates obtained in this study, which were found to have average impedances of 9.61×106 kg/m2s and 9.20 kg/m2s, respectively, corresponding to Rayleigh reflections of −11.67 and −12.38 dB, which in turn translates to about a quarter (26% and 24%, respectively) of the amplitude of the incident wave being reflected. While the attenuation in the crusts and difference in impedance between the crusts and the substrates is expected to cause fluctuations in the strength of the signals used for thickness measurements, the crust samples obtained at this site were found to have relatively uniform compressional wave velocities, which suggests that the time of flight between the acoustic reflections at the top and bottom surfaces of the crusts can be accurately translated into values for thickness at this site.

SECTION IV.

Acoustic Instrumentation

Fig. 7. - A 3000-m depth rated spherically focused transducer array developed to make high-resolution subsurface measurements of manganese encrusted seafloors.
Fig. 7.

A 3000-m depth rated spherically focused transducer array developed to make high-resolution subsurface measurements of manganese encrusted seafloors.

A 3000-m depth rated probe has been developed to perform remote acoustic measurements of manganese crust thickness. The probe, pictured in Fig. 7, has a spherically focused transducer of diameter 160 mm and a focal length of 750 mm. The transducer consists of an annular array of 1-MHz 1–3 piezoelectric composite elements for transmission, with a 100-kHz receiver array built-in along the same axis. The probe transmits a 1-MHz, high-frequency amplitude modulated signal to generate a narrow 100-kHz beam that penetrates the target. Fig. 8 shows the beam pattern of the probe measured using a broadband hydrophone (RESON TC-4038). The hydrophone was placed at the focal length of the transducer and was scanned laterally in steps of 1 mm. The high-frequency 1-MHz and low-frequency 100-kHz components of the transmitted amplitude modulated signal were separated using a band-pass filter, and the corresponding beam patterns are plotted by the dotted line and crosses, respectively. It can be seen that the 100-kHz component that penetrates the target, denoted as 100 kHz secondary, is narrow with a −3-dB acoustic footprint of diameter <20 mm on a target 750 mm away. Since the probing wave is generated using a high-frequency primary, it has no sidelobes and does not suffer mechanical or electrical reverberation at the measurement frequency [26]. These points are advantageous for discretizing thin, undulating surface layers such as those of manganese crusts, since the energy of the wave is confined to only a small projected area on the target surface and has a short duration in time. The drawback of this approach is the inefficient transfer of energy from the primary to the secondary frequency. To overcome this point, during operation, the probing wave is generated using a high primary pulse power of 15 kW using a linear amplifier. The beam pattern of the 100-kHz receiver used to detect the echoes of the transmitted wave was recorded by scanning the hydrophone laterally in steps of 5 mm. Although the receiver has a broad beam pattern, with a −3-dB acoustic footprint of diameter 50 mm at a distance of 750 mm, this is not limiting, since the original transmitted wave is narrow. The echoes of the transmitted wave are recorded with a sampling rate of 500 kHz, which translates to a resolution of about 3 mm in the thickness measurements of the manganese crusts.
Fig. 8. - One-way beam pattern measured at the focal length of the probe.
Fig. 8.

One-way beam pattern measured at the focal length of the probe.

Fig. 9. - Block diagram of the peak detection and crust thickness measurement algorithm.
Fig. 9.

Block diagram of the peak detection and crust thickness measurement algorithm.

A simple peak detection algorithm has been implemented to automatically measure crust thickness based on the time delay between reflections that occur at the top and bottom surfaces of the crusts, and the acoustic properties of the manganese crust samples determined in Section III. The algorithm, illustrated in Fig. 9, recognizes peaks in the acoustic data based on envelope detection of the reflected signal $S(t)$, where each signal represents a set of $N$ discrete points containing the acoustic reflections that occur within a fixed period after a single transmission. The algorithm first applies a band-pass filter to isolate the probing frequency component $S_{f}(t)$, and applies the Hilbert transform to extract its low-frequency envelope $H(S_{f}(t))$. Next, the algorithm uses the following condition to check that the signal has a sufficient signal-to-noise ratio (SNR) for reliable peak detection: $${\overline{H\left(S_{f}(t)\right)}\over\max H\left(S_{f}(t)\right)}<k_{S/N}\eqno{\hbox{(1)}}$$View SourceRight-click on figure for MathML and additional features.where $k_{S/N}$ is a constant less than 1. If the above condition is satisfied, the following thresholds are determined for the top and bottom peaks, respectively: $$\eqalignno{\sigma_{\rm top}=&\,k_{\rm top}\max H\left(S_{f}(t)\right)&\hbox{(2)}\cr\sigma_{\rm bot}=&\,k_{\rm bot}\max H\left(S_{f}(t)\right)&\hbox{(3)}}$$View SourceRight-click on figure for MathML and additional features.where $k_{\rm top}$ and $k_{\rm bot}$ are constants with values less than 1. A window function is used to search for the first peak using the condition $$\eqalignno{&{\hbox {if}}\ {t_{k}:\left\{\max H\left(S_{f}(t_{i}\ldots t_{i+n})\right)>\sigma_{\rm top}\right\}\choose H\left(S_{f}(t_{k})\right)>H\left(S_{f}(t_{k\pm 1})\right)+\delta}\cr&\qquad\qquad\qquad t_{\rm top}=t_{k}\cr&{\hbox {else}}\cr&\qquad\qquad\qquad t_{i}=t_{i+{n\over 2}}&\hbox{(4)}}$$View SourceRight-click on figure for MathML and additional features.where $n$ is the width of the search window, and $\delta$ is a threshold defining the minimum height of a peak with respect to its neighbors. If condition (4) is satisfied, the peak is labeled as the top surface reflection and the time at which it occurred is stored as $t_{\rm top}$. The location of the reflection of the top surface is converted to the inertial frame by interpolating the position and pose of the vehicle when the acoustic measurement was made assuming the speed of sound in seawater up to $t_{\rm top}$. The following notation is adopted for the envelope of the subsurface reflections that occur after $t_{\rm top}$:$$\gamma^{r}(t_{\min}\ldots t_{\max})=H\left(S_{f}(t_{\rm top}+t_{\min}\ldots t_{\rm top}+t_{\max})\right)$$View SourceRight-click on figure for MathML and additional features.where $t_{\min}$ and $t_{\max}$ correspond to the two-way travel time through the minimum and maximum thicknesses of the crusts considered, and $r=[x,y]$ is the 2-D projection of the top surface of the seafloor in the inertial frame. We obtain the following expression for the subsurface reflections compensated for attenuation in the crust layer: $$\gamma_{\alpha}^{r}(t_{\min}\ldots t_{\max})=\gamma^{r}(t_{\min}\ldots t_{\max})\times 10^{(2\alpha_{f}d_{z})^{1\over 2}}\eqno{\hbox{(5)}}$$View SourceRight-click on figure for MathML and additional features.where $$2d_{z}=(t_{\min}\ldots t_{\max})\bar{c}_{c}$$View SourceRight-click on figure for MathML and additional features.and $\alpha_{f}$ is the attenuation coefficient of the manganese crusts at the measuring frequency and $\bar{c}_{c}$ is the acoustic velocity in the crusts. Since manganese crusts consist of a single layer [21], in the ideal case, it should be sufficient to assume that the largest value of $\gamma_{\alpha}^{r}(t_{\min}\ldots t_{\max})$ corresponds to the crust–substrate interface. However, such an approach would be sensitive to local inclusions within the crust layer, which may result in incorrect identification of the interface. To overcome this, we make use of the fact that the thickness of the crust layer changes gradually, and for each acoustic measurement $\gamma_{\alpha}^{r}$, we consider the set of subsurface reflections of measurements made within some distance $r_{d}$ of $r$ as follows: $${\mmb\gamma}_{\alpha}^{r}(t_{\min}\ldots t_{\max})={\sum_{r^{\prime}}\kappa(r^{\prime})\gamma_{\alpha}^{r^{\prime}}(t_{\min}\ldots t_{\max})\over\Sigma_{r^{\prime}}\kappa(r^{\prime})}\eqno{\hbox{(6)}}$$View SourceRight-click on figure for MathML and additional features.where $r-r_{d}<r^{\prime}<r+r_{d}$, and $$\kappa(r^{\prime})=1-{\vert r^{\prime}-r\vert\over r_{d}}.$$View SourceRight-click on figure for MathML and additional features.The crust–substrate interface is determined as follows: $$\eqalignno{&{\hbox {if}}\left(p_{\rm bot}:\left\{\max{\mmb\gamma}_{\alpha}^{r}(t_{\min}\ldots t_{\max})>\sigma_{\rm bot}\right\}\right)\cr&\qquad\qquad\quad t_{\rm bot}:\left\{\max{\mmb\gamma}_{\alpha}^{r}(t_{\min}\ldots t_{\max})\right\}.&\hbox{(7)}}$$View SourceRight-click on figure for MathML and additional features.If condition (7) is satisfied, the distance to the crust–substrate interface is measured based on the time of flight between the reflections as follows: $$d=\bar{c}_{c}{t_{\rm top}-t_{\rm bot}\over 2}\eqno{\hbox{(8)}}$$View SourceRight-click on figure for MathML and additional features.where, for the purposes of this study, the effect of refraction at the top surface of the crusts is ignored, and the signal is assumed to travel in a straight line once it enters the seafloor.

It should be noted that the algorithm makes the following assumptions implicitly:

  • the crust deposit consists of a single layer;

  • the target's uppermost surface is manganese crust.

The first condition is reasonable, considering the mechanisms understood to be behind the formation of manganese crusts [21]. The latter condition can be confirmed visually using images of the seafloor taken by the underwater vehicle on which the acoustic device is mounted. Sections V and VI describe the mapping system used to obtain the visual images and the algorithms implemented to automatically identify areas of exposed crust.

SECTION V.

Visual Instrumentation

A vision-based mapping device has been developed to create accurate 3-D color reconstructions of the seafloor. The device consists of a single camera, a sheet laser, a light, and a shade, arranged as shown in Fig. 10. Seafloor topography is measured by laser profiling using the sheet laser to project a line vertically downward onto the seafloor. The laser line projection is captured by a color camera mounted with a translational and angular offset, so that its field of view extends to the projection of the sheet laser while also covering the area vertically beneath the camera, which is illuminated by a light-emitting diode (LED) array. A shade is used to ensure that the area surrounding the laser line projection remains dark to allow for more robust laser line detection, as can be seen in Fig. 11. As the vehicle moves forward, the laser line projection scans the shape of the terrain, which allows a bathymetric model to be generated based on triangulation after identification of the laser line in the images taken by the camera. Color information is retrieved from the portion of the image vertically beneath the camera, which is illuminated by the LEDs, as illustrated in Fig. 12. The vehicle's navigation data provide the necessary position and orientation information to correlate the bathymetry data with its corresponding color using a pixel-based approach, as described in [6]. For most vehicles, this consists of a Doppler velocity log (DVL), depth sensor, and a three-axis magnetic compass, gyro and acceleration sensor or, if available, a fiber optic or ring laser gyro-based inertial navigation system (INS). These sensors are used to determine the vehicle's $x$-, $y$-, $z$-coordinates and the roll, pitch, and yaw orientations in the inertial frame. Although navigation data are important for accurate reconstruction, since the time between the measurement of bathymetry and color is typically only a few seconds, sensor drift does not have a significant adverse effect on the color matching and a standard DVL and compass-based INS has been found to provide sufficient accuracy to generate 3-D maps that are consistent.

Fig. 10. - Photo showing the components of the seaXerocks 3-D mapping device.
Fig. 10.

Photo showing the components of the seaXerocks 3-D mapping device.

Fig. 11. - Image of crust deposits obtained by the mapping device. The sheet laser projection used for generating bathymetry is visible in the top part of the image, which is shaded to increase contrast to allow for more robust laser line detection, while the bottom part of the image is illuminated by white LEDs to record the color of the seafloor.
Fig. 11.

Image of crust deposits obtained by the mapping device. The sheet laser projection used for generating bathymetry is visible in the top part of the image, which is shaded to increase contrast to allow for more robust laser line detection, while the bottom part of the image is illuminated by white LEDs to record the color of the seafloor.

Fig. 12. - Illustration of the mapping algorithm. As the vehicle moves forward, both the laser projection and LED illuminated regions scan the seafloor, and this information is used to generate 3-D color reconstructions of the seafloor.
Fig. 12.

Illustration of the mapping algorithm. As the vehicle moves forward, both the laser projection and LED illuminated regions scan the seafloor, and this information is used to generate 3-D color reconstructions of the seafloor.

The camera used in this study has a resolution of 640 × 480 pixels with images recorded at 29.97 frames/s. While higher resolution cameras are widely available, a fast frame rate is also necessary for this application to achieve comparable resolutions along all three axes. For the setup used during the cruise, at an altitude of 1 m, the system had a swath of roughly 1.36 m about the direction of travel, with a vertical resolution of 4.4 mm, an across-track resolution of 2.1 mm, and, for a surge velocity of 0.2 m/s, an along-track resolution of 6.6 mm. The main advantages of this approach to 3-D reconstruction is that the long baseline introduced by the laser profiling provides both consistent and high vertical resolution, uniform lighting conditions since color information is retrieved only from the area directly beneath the camera, and computational efficiency since all calculations necessary for reconstruction are feedforward, and no feature matching is required between images.

SECTION VI.

Classification and Data Fusion

To identify areas of exposed crust, a multiparameter feature matching technique is used to classify continuous segments of the seafloor into one of the following groups: exposed crust, sediment, or transition. The classification is performed in three stages. First, the seafloor is segmented into continuous regions that are similar in texture and do not have any sudden changes in their bathymetry. Once segmented, a signature is extracted based on the color, texture, and 3-D bathymetric information contained within each segmented region. In the final stage of the classification, the extracted signatures are compared to a set of prelabeled regions that have been chosen by a human expert to represent each class. While there exist a number of methods that could be applied to the task of identifying regions of exposed crust, including unsupervised methods such as those described in [17]–​[19], the goal of the classification in this work is clearly defined and so supervised methods can be used. Supervised parametric classification using GMMs has emerged as an effective method for classification of unstructured natural terrains in the fields of aerial imagery [27] and mobile ground robotics [28], and it has been demonstrated that using both geometry and imagery data can improve the results of classification [29]. In this application, we apply GMMs to classify regions of the seafloor into one of three predefined categories based on features extracted from underwater bathymetric and visual measurements. Each region adopts the label of the class it is most likely to belong to. Once the regions are classified and labeled, it is possible to process the data measured in each region using a set of instructions specific to the type of seafloor. In this application, we process acoustic measurements made in regions of exposed crust. For regions classified as sediment, the acoustic measurements are ignored. Measurements made in transition regions are processed and flagged as requiring further human attention to judge whether the acoustic measurements should be taken into account. The planar surface area is determined for each type of seafloor to allow for statistical analysis.

A. Segmentation and Signature Extraction

Each continuous section of the seafloor should be similar in terms of its texture and its bathymetric features. Thus, to segment the seafloor into continuous regions, we use criteria based on both visual and bathymetric information contained in the 3-D reconstructions. First, the texture of the seafloor over a given spatial range is determined by calculating its entropy, i.e., the level of variation of color information. The entropy is a statistical measure of randomness, or local complexity, in the data over a particular region, that can be defined as $$\psi_{\zeta,\eta}=-\sum_{r,g,b}p_{\zeta,\eta}(\zeta_{i})\log_{2}p_{\zeta,\eta}(\zeta_{i}),$$View SourceRight-click on figure for MathML and additional features.where $p_{\zeta,\eta}$ is the probability that a pixel in the region $\eta$ has a color intensity value of $\zeta_{i}$. In our application, we use the 2-D projections of the 3-D reconstructions obtained by the mapping system as the input, which are resampled at uniform resolution using a distance weighted average of the nearest measurements. Since the information is measured in the inertial frame, the region defined by $\eta$ corresponds to dimensions in space, where in this work, we choose square windows of size 0.1 × 0.1 m2, 0.3 × 0.3 m2, and 0.5 × 0.5 m2, after performing Gaussian filtering at each scale to capture details that occur over the different spatial scales. The entropy is calculated for each of the color channels by letting $\zeta_{i}$ equal the intensity of the red, green, and blue information in the images, respectively. Although for this particular application a grayscale description of color would be sufficient since the crusts are black, we maintain generality as there are a number of applications where different color channels contain useful information. We take the sum of the entropies calculated for each of the color channels and at each spatial scale as our representation of the texture of the seafloor as follows: $$\Psi=\sum_{\zeta}\sum_{\eta}\psi_{\zeta,\eta}.$$View SourceRight-click on figure for MathML and additional features.Since our aim is to identify the underlying nature of the seafloor, the first stage in the segmentation is to remove small local features from the texture, to prevent oversegmentation. Features that display high values of local entropy are removed by assigning each point in our representation the minimum value of intensity that occurs within the neighborhood defined by $R_{X}$, as follows: $$\Psi_{-}(x,y)=\min_{k,l\in\eta}\Psi(x+k,y+l).$$View SourceRight-click on figure for MathML and additional features.Next, we apply the grayscale reconstruction method outlined in [30], by assigning each point the maximum intensity value of its nearest neighbors $I$ up to the limit of the original representation $\Psi$ $$\delta^{(1)}_{\Psi}(\Psi_{-})=\max_{k,l\in I}\Psi_{-}(x+k,y+l)\wedge\Psi(x,y)$$View SourceRight-click on figure for MathML and additional features.where $\wedge$ is the pointwise minimum operator. This step is repeated iteratively until it converges on a stable representation $$\Psi_{\delta}=\bigvee_{n\geq 1}\delta_{\Psi}^{(n)}(\Psi_{-}).$$View SourceRight-click on figure for MathML and additional features.Next, features that display low values of local entropy are removed by assigning each point in $H_{r}$ the maximum value of intensity that occurs within $\eta$ as follows: $$\Psi_{\delta+}(x,y)=\max_{k,l\in\eta}\Psi_{\delta}(x+k,y+l).$$View SourceRight-click on figure for MathML and additional features.Then, we assign each point the minimum intensity value of its nearest neighbors down to the limit of $\Psi_{\delta}$$$\epsilon_{\Psi_{\delta}}^{1}(\Psi_{\delta+})=\min_{k,l\in I}\Psi_{\delta+}(x+k,y+l)\vee\Psi_{\delta}(x,y)$$View SourceRight-click on figure for MathML and additional features.where $\vee$ is the pointwise maximum operator. This step is iterated until a stable representation is obtained $$\Psi_{\epsilon}=\bigwedge_{n\geq 1}\epsilon_{\Psi}^{(n)}(\Psi_{\delta+})$$View SourceRight-click on figure for MathML and additional features.where $\Psi_{\epsilon}$ represents the texture of the underlying seafloor.

To segment the data, we apply the marker controlled watershed technique described in [31], which is a region growing approach to segmentation, and use a modified gradient operator that considers both the texture and bathymetry of the data to control segmentation. First, we consider the function defined by the gradient of the texture $\Psi_{\epsilon}$, so that regions of uniform texture become regional minima surrounded by edges, or peaks, that correspond to changes in texture. The gradient of the texture is obtained through convolution with the Sobel operator as follows:$$\eqalign{G_{x}(x,y)=&\,\sum_{k=-1}^{1}\sum_{l=-1}^{1}\Omega_{x}(k,l) \Psi_{\epsilon}(x+k,y+l)\cr G_{y}(x,y)=&\,\sum_{k=-1}^{1}\sum_{l=-1}^{1}\Omega_{y}(k,l)\Psi_{\epsilon}(x+k,y+l)\cr{\bf G}(x,y)=&\,\sqrt{G_{x}^{2}(x,y)+G_{y}^{2}(x,y)}\cr{\bf G}^{\prime}(x,y)=&\,{{\bf G}(x,y)\over\max{\bf G}(x,y)}}$$View SourceRight-click on figure for MathML and additional features.where $\Omega_{x}$ and $\Omega_{y}$ are Sobel operators. To account for changes in bathymetry, we calculate the gradient of the seafloor through convolution of the bathymetry with the Sobel operator over the region defined by $\eta$ as follows: $$\eqalign{F_{x}(x,y)=&\,\sum_{k=-{\eta\over 2}}^{\eta\over 2}\sum_{l=-{\eta\over 2}}^{\eta\over 2}\Omega_{x}(k,l)z(x+k,y+l)\cr F_{y}(x,y)=&\,\sum_{k=-{\eta\over 2}}^{\eta\over 2}\sum_{l=-{\eta\over 2}}^{\eta\over 2}\Omega_{y}(k,l)z(x+k,y+l)\cr{\bf F}(x,y)=&\,\sqrt{F_{x}^{2}(x,y)+F_{y}^{2}(x,y)}\cr{\bf F}^{\prime}(x,y)=&\,{{\bf F}(x,y)\over\max{\bf F}(x,y)}.}$$View SourceRight-click on figure for MathML and additional features.By combining the gradient of the texture and bathymetry, we formulate the following modified gradient operator: $$\eqalign{{\bf G}^{\prime\prime}(x,y,n)=&\,0,\qquad{\hbox {for}}\ x,y\in M\vert m\cr{\bf G}^{\prime\prime}(x,y,\infty)=&\,{{\bf G}^{\prime}(x,y){\bf F}^{\prime}(x,y)\over 2},\qquad{\hbox {elsewhere}}}$$View SourceRight-click on figure for MathML and additional features.where $M$ and $m$ represent the regional maxima and minima of $H_{\epsilon}$, and $n$ is a discrete label assigned to each continuous region ${\bf G}^{\prime\prime}=0$, and the value $n=\infty$ indicates that the area has not yet been assigned to a region. In the modified gradient operator, regions of uniform texture become regional minima surrounded by edges and peaks that correspond to changes in texture or bathymetry. The watershed transformation finds the set of points $x,y$ covered by each region labeled $n$ as the value of ${\bf G}^{\prime\prime}$ is increased. The new set of points that have not previously been assigned, i.e., those with $n=\infty$, inherit the label $n$ of the region that was grown. For points that have previously been assigned a label, the label remains unchanged. This step is repeated iteratively until every point in the data is assigned a label. The terms ${\bf G}^{\prime}(x,y)$ and ${\bf F}^{\prime}(x,y)$ do not influence the number of regions into which the data are segmented, but have the effect of drawing region boundaries to locations with changes in texture and bathymetry to realize more natural boundaries.

Once the seafloor has been separated into continuous labeled regions, a set of signatures is extracted from each region, which is used as the basis for classification. In most cases, an individual feature cannot contain enough information to discriminate complex, natural scenes such as the surface of the seafloor. However, through the combination of multiple features, a unique description may be possible. The choice of features used to build up the signature plays a key role in the analysis, since their combination must contain enough information and flexibility to discriminate between different types of seafloor, while remaining concise to achieve scalability for large data sets.

We choose a signature $\tau_{i}$ that consists of 15 parameters that relate directly to recognizable physical features of the seafloor. The first three variables $\tau_{1\ldots 3}$ represent the average values of the red, green, and blue color channels in each segment. The variables $\tau_{4\ldots 12}$ represent the average entropy, or texture of the seafloor for the RGB channels over three different spatial scales. As mentioned, the individual color channels are not expected to play an important role for this application, however, the different levels of gray and their textures over different spatial scales are expected to be important. The variables $\tau_{13\ldots 15}$ represent the normalized slope, roughness, and the size of the protrusions in each segment of the seafloor. The normalized slope $\theta_{ev}^{n}$ is calculated by determining the elevation $\theta_{ev}$ of the plane, which gives the least squares fit of the bathymetry using singular value decomposition [32], which is normalized by dividing by $\pi/2$. The roughness, or complexity of the seafloor, is represented by the inverse of its rugosity, where rugosity is defined as the ratio between the draped surface area, i.e., the sum of the surface areas of the individual triangles that form the bathymetry mesh, and its projection onto the best fit plane, as follows: $$\xi={\lambda_{d}\over\lambda_{p}}.$$View SourceRight-click on figure for MathML and additional features.A rugosity value of 1 corresponds to a perfectly flat bathymetry, and as the roughness of the surface increases, the rugosity increases. For our application, we choose the inverse of the rugosity as our representation of roughness since it is bounded, and so the smaller the value, the rougher the surface. While this gives a measure of surface roughness, it does not discriminate surfaces with large protrusions from surfaces with small protrusions. To discriminate these, we compute the volume-to-surface-area ratio of the protrusions. The volume between the surface and its planar projection is determined as $$V=\int\int\lambda_{d}(x,y)-\lambda_{p}(x,y)dA$$View SourceRight-click on figure for MathML and additional features.where we represent the volume-to-surface-area ratio as follows: $$V_{R}=k_{v}{V\over\lambda_{d}}$$View SourceRight-click on figure for MathML and additional features.where $k_{v}$ is a scaling constant with units m−1. A large value of $V_{R}$ indicates that the surface has large protrusions, and a small value indicates that the protrusions are small. All features used, summarized in Table II, are dimensionless with values between 0 and 1.

Table II Signature Used To Describe Visual And Bathymetric Features Of Seafloor Segments
Table - Signature Used To Describe Visual And Bathymetric Features Of Seafloor Segments

B. Classification and Data Processing

The final step in classification involves grouping the different regions of the seafloor into meaningful categories using GMMs. A GMM is a probabilistic model that represents a discrete variable that cannot be observed directly, but can be expressed through a continuous distribution of observable variables $\tau_{i}$. In our application, the discrete variable is the type of seafloor we wish to classify, and the observable descriptors consist of the signature of features given in Table II. For $n_{\tau}$ observable variables, where $\tau=(\tau_{1}\ldots\tau_{n_{\tau}})^{T}$, the $j$th component Gaussian density can be defined $${\cal N}(\tau\vert\mu_{j},\Sigma_{j})\!\!=\!\!{1\over(2\pi)^{n_{\tau\over 2}}\sqrt{\vert\Sigma_{j}\vert}}\exp\!\!\left(-{1\over 2}(\tau\!-\!\mu_{j})^{T}\Sigma_{j}^{-1}\!(\tau\!-\!\mu_{j})\right)$$View SourceRight-click on figure for MathML and additional features.where $\mu_{j}$ is the mean and $\Sigma_{j}$ is the covariance matrix. The probability given a mixture of $K$ component Gaussians densities is $$p(\tau)=\sum_{j=1}^{K}\omega_{j}\cdot{\cal N}(\tau\vert\mu_{j},\Sigma_{j})$$View SourceRight-click on figure for MathML and additional features.where $\omega_{j}$ is the weight of the $j$th Gaussian, subject to the condition $\sum_{j}^{K}\omega_{j}=1$ and $0\leq\omega_{j}\leq 1$.

The values for $\mu_{j}$, $\Sigma_{j}$, and $\omega_{j}$ are determined using the expectation–maximisation (EM) algorithm [33], applied to a set of prelabeled training data that represent the specific type of seafloor. Each segment of the seafloor is classified into one of the following groups: exposed crust, sediment, and transition, where to classify each segment in the data, we calculate the likelihood that the observed variables belong to each of the types of seafloor in the classifier, and the segment is assigned the label of the class that gives the maximum likelihood. For regions of exposed crust, the acoustic measurements made within the crust region are processed to determine the thickness of the crusts, as discussed in Section IV. For regions classed as sediment, the acoustic data are not processed, and for transition regions, the acoustic data are processed but are flagged for further human judgement, since the results may or may not be meaningful, depending on the nature of the transition region. The planar surface area is recorded for all types of seafloor to provide data for statistical analysis and enable computation of crust volume. While other information can also be extracted for analysis, in this initial study, we use only the planar surface area of the different seafloors and the measurements of thickness in areas of exposed crusts and, where appropriate, transition regions. Fig. 13 illustrates the steps employed to process the information.

Fig. 13. - Classification and processing of information based on visual and acoustic measurements.
Fig. 13.

Classification and processing of information based on visual and acoustic measurements.

SECTION VII.

Experiments and Results

Sea trials of the proposed instruments were performed at #5 Takuyo seamount to demonstrate in situ measurement of manganese crust thickness and to generate surface and subsurface 3-D reconstructions of manganese encrusted seafloors. The acoustic probe and the 3-D mapping system were mounted onboard the ROV Hyper-Dolphin of JAMSTEC, as shown in Fig. 14. The acoustic probe was mounted at the front of the vehicle along with its power amplifier, and the 3-D mapping device was mounted at the rear of the vehicle along with a DVL (RDI WHN1200), compass-based INS, and a depth sensor to determine the vehicle's position and pose in the inertial frame. The system was deployed a total of three times during the NT10-11 cruise of the R/V Natsushima. During the survey, the ROV was operated at depths of between 1000 and 3000 m at low altitudes of <1.0 m, with surge velocities of between 0.2 and 0.3 m/s to survey the area both acoustically and visually. In addition to the payloads for the subsurface acoustic measurements and visual mapping, sampling was performed using the ROV manipulators and a rotary blade to verify the measurements of the acoustic system. In this section, we present three short sections of data measured during the cruise at various locations of the seamount.

Fig. 14. - The ROV Hyper-Dolphin about to be deployed during NT10-11, with (1a) the developed acoustic probe, (1b) power amplifier, and (2) a seafloor mapping device mounted as payload.
Fig. 14.

The ROV Hyper-Dolphin about to be deployed during NT10-11, with (1a) the developed acoustic probe, (1b) power amplifier, and (2) a seafloor mapping device mounted as payload.

The classification algorithm was trained using 100 prelabeled segments of 3-D data obtained during the cruise. Each class was assigned 30 segments of data chosen by a human expert to represent each type of seafloor. The signatures extracted from these segments were used to build representative sets to describe each type of seafloor. A fourth class was also implemented to identify poor quality data. This class consisted of ten segments of data where, for example, the vehicle agitated sediments during acquisition, or where only partial laser line detection was possible and so the data are sparse. The visual and acoustic data measured in these areas are not processed further. Fig. 15 shows some examples of segments of 3-D data for exposed crust [Fig. 15(a)], sediment [Fig. 15(b)], and transition [Fig. 15(c)], obtained during the cruise. Fig. 16 shows the mean and standard deviation of the signatures extracted from the data used to train the GMMs. It can be seen that there is significant overlap in the values for the individual features between the different classes, and some features appear to be more discriminative than others. It is also noticeable that some features have the largest contrast between crusts and sediments, while other have a larger contrast between transition and sediments. For example, the first three features, which represent the average color of each segment, have the highest values for sediments and the lowest values for exposed crusts, with the value for transition regions somewhere between. This is intuitive as the sediments are light in color (i.e., high red, green, and blue intensities), and the crusts are dark (i.e., low red, green, and blue intensities) and the transition regions are a mixture of exposed manganese deposits and sediments. The textures, or entropies, on the other hand, have the largest values for transition regions (i.e., a large color variation) and the smallest values for sediments (i.e., relatively uniform color). On the fine spatial scale (features 4–6) both the crusts and the transition regions have larger values for entropy compared to the sediments. However, on this scale, it would be difficult to discriminate the crusts from the transition regions due to the large overlap in their values. On the larger spatial scales (features 7–12), the crusts appear more uniform in color, with lower values for entropy while the entropy for the transition regions remains high, giving more contrast between the two types of seafloor, which is favorable for classification. Since most of the data obtained during the cruise were from the flat shoulder of the seamount, the slope (feature 13) is similar for all types of seafloor. However, the values for the different classes are expected to diverge for data obtained on more steeply sloped regions of the seamount, since sediments and debris are less likely to settle in steeper areas. The inverse rugosity (feature 14) indicates that both the crusts and the transition regions are, on average, rougher than the sediments. The volume-to-planar-surface-area ratio (feature 15) contrasts between transition regions and crusts since the transition regions tend to have larger protrusions. It is not clear at this stage if the patterns observed are general, and it is necessary to analyze data from a wider variety of terrains to draw any statistical conclusions regarding these features. However, for the data available, it is clear that the individual parameters cannot discriminate the different types of seafloor alone, highlighting the point that classification of complex natural terrains requires a combination of different features to be processed together to determine the class of seafloor that each segment is most likely to belong to.

Fig. 15. - Segments of 3-D reconstruction data from NT10-11 showing (a) exposed crust, (b) sediment, and (c) transition regions.
Fig. 15.

Segments of 3-D reconstruction data from NT10-11 showing (a) exposed crust, (b) sediment, and (c) transition regions.

Fig. 16. - Signatures extracted from prelabeled segments of data chosen by a human expert to represent each type of seafloor. These signatures are used to train the GMM in the classification algorithm. Each class consists of 30 segments of data, where the circles show the mean value for each of the features (see Table II) and the bars represent the standard deviation for each feature within each class.
Fig. 16.

Signatures extracted from prelabeled segments of data chosen by a human expert to represent each type of seafloor. These signatures are used to train the GMM in the classification algorithm. Each class consists of 30 segments of data, where the circles show the mean value for each of the features (see Table II) and the bars represent the standard deviation for each feature within each class.

A. Seafloor A

Figs. 17 and 18 show a section of visual and acoustic data measured continuously over a distance of 37 m on the shoulder of the seamount during the cruise. Fig. 17(a) and (b) shows the 2-D projection and bathymetry data extracted from the 3-D reconstruction of this area. Fig. 17(c) and (d) shows the results of segmentation and classification, respectively, where the pie charts show the probabilities calculated by the GMMs that the indicated segments belong to each class. It can be seen that the algorithm is capable of successfully identifying areas of exposed crust, sediment, and transition. The section has a planar surface area of 44.57 m2, with 36.06 m2 (80.8%) of exposed crust, 2.57 m2 (5.8%) of sediment, and 4.29 m2 (9.6%) classified as transition. Fig. 18(a) shows the acoustic measurements made in this area. In Fig. 18(b), the same data are plotted with respect to the surface of the seafloor and the crust thickness computed using the peak detection algorithm is plotted in Fig. 18(c). The thickness measurements in black are those made in regions classified as exposed crust. The measurements in red are those made over transition regions. Fig. 19(a) and (b) shows the actual scale thickness measurements together with 3-D color reconstructions of the regions highlighted with the green and blue borders in Fig. 18(a). The acoustic thickness measurements have been offset vertically by 1 m in the figure for clarity of presentation.

Fig. 17. - Three-dimensional reconstruction of seafloor A, a 37-m-long region that consists mainly of exposed crusts. (a) The top view of the reconstruction, (b) the laser bathymetry, (c) the results of segmentation, and (d) the labels assigned to each segment by the classification algorithm. The pie charts show the probability that the indicated segments belong to each class of seafloor. Feature A is a fault in the crust which is filled with sediment; B is a loose slab of pavement like crust; and C is a small terrace with loose slabs of crust at its base.
Fig. 17.

Three-dimensional reconstruction of seafloor A, a 37-m-long region that consists mainly of exposed crusts. (a) The top view of the reconstruction, (b) the laser bathymetry, (c) the results of segmentation, and (d) the labels assigned to each segment by the classification algorithm. The pie charts show the probability that the indicated segments belong to each class of seafloor. Feature A is a fault in the crust which is filled with sediment; B is a loose slab of pavement like crust; and C is a small terrace with loose slabs of crust at its base.

Fig. 18. - Acoustic data for seafloor A showing (a) acoustic data compensated for vehicle motion, (b) acoustic data with respect to the surface of the seafloor, and (c) crust thickness measurements made using the peak detection algorithm, where the measurements in black are made over exposed crusts and those in red are made over transition regions, as determined by the classification algorithm. In this case, the measurements in red were made over pavement like slabs of crust, and it may be judged by a human that these are valid measurements of thickness. However, for the purpose of this study, these values are not used in the computation of the volume of crusts in this area. The labels A, B, and C in (a) correspond to the features of the seafloor shown in Figs. 17 and 19.
Fig. 18.

Acoustic data for seafloor A showing (a) acoustic data compensated for vehicle motion, (b) acoustic data with respect to the surface of the seafloor, and (c) crust thickness measurements made using the peak detection algorithm, where the measurements in black are made over exposed crusts and those in red are made over transition regions, as determined by the classification algorithm. In this case, the measurements in red were made over pavement like slabs of crust, and it may be judged by a human that these are valid measurements of thickness. However, for the purpose of this study, these values are not used in the computation of the volume of crusts in this area. The labels A, B, and C in (a) correspond to the features of the seafloor shown in Figs. 17 and 19.

The measurements made over the exposed crusts indicate that the crusts in this area have an average thickness of 103.0 mm with a standard deviation of 9.5 mm. This is consistent with crusts sampled in this area, which have thicknesses between 80 and 110 mm, as seen in Fig. 20. The discontinuity in the acoustic data in Fig. 18(a), labeled A, corresponds to a fault line where there is no crust, as can be seen in Figs. 17(a) and 19(a). This region was correctly classified as sediment by the algorithm [see Fig. 17(d)], and so the acoustic data obtained in this area were not processed for thickness measurements. This feature can also be recognized in Fig. 18(b) since there is only a single surface reflection in the acoustic data in this region, as opposed to the two reflections, surface, and subsurface, that can be seen in the measurements over the adjacent crust covered regions. The features labeled B and C in Figs. 17(a) and 19(b) correspond to a loose slab of pavement like crust resting on top of the seafloor, and a vertical drop of 0.72 m at the edge of a small terrace, at the foot of which are several loose slabs of pavement like crust. Both regions were classified as transition by the algorithm.

Fig. 19. - Surface 3-D reconstructions plotted together with subsurface acoustic thickness measurements for seafloor A, where (a) is the region of Fig. 18(a) marked with a green boundary, and (b) is the region with a blue boundary. Features A, B, and C correspond to those marked in Figs. 17(a) and 18(a). The thickness measurements are shown to scale, offset by 1 m from the surface of the seafloor for clarity of presentation. The measurements in black were made over regions of exposed crust, and the measurements in red were made over transition regions, which in this case correspond to loose slabs of pavement-like crust.
Fig. 19.

Surface 3-D reconstructions plotted together with subsurface acoustic thickness measurements for seafloor A, where (a) is the region of Fig. 18(a) marked with a green boundary, and (b) is the region with a blue boundary. Features A, B, and C correspond to those marked in Figs. 17(a) and 18(a). The thickness measurements are shown to scale, offset by 1 m from the surface of the seafloor for clarity of presentation. The measurements in black were made over regions of exposed crust, and the measurements in red were made over transition regions, which in this case correspond to loose slabs of pavement-like crust.

The acoustic data were processed to give measurements of thickness, however, these measurements, shown in red in Fig. 18(c), are flagged as requiring further human interpretation. In this case, since the transitions correspond to loose slabs of pavement-like crust, it is reasonable to consider that the measurements made correspond to the thicknesses of the slabs. The average thickness of the slab labeled B was 62.1 mm, which is thinner than the surrounding crust outcrop. The thickness measurements for the slabs at the base of the terrace had an average value of 100.0 mm, indicating these slabs may have originated from the surrounding crust outcrop.

Based on the average thickness measurements and the surface area of the exposed crusts (and the average density of the crust samples), we can compute the mass of crusts scanned in this area to be 7.12 T. This corresponds to 197.8 kg per unit area of exposed crust, and 159.8 kg per unit area of seafloor in the region shown. The latter value provides an indication of how rich this particular scanned region of the seafloor is in terms of manganese crust deposit.

B. Seafloor B

The second example seafloor is a short region where the acoustic measurements alone would potentially lead to incorrect conclusions regarding the thickness of the crusts in the area. The data were obtained from the shoulder of the seamount, about 60 m from the data shown in seafloor A, where the exposed crusts in these examples both belong to the same outcrop. Fig. 21(a) and (b) shows the 2-D projection and bathymetry extracted from a 21-m-long section of 3-D reconstruction data. The results of segmentation and classification [Fig. 21(c) and (d)] show that the algorithm can successfully identify regions of exposed crusts, such as feature D, and transition regions, which in this case correspond to nodular deposits, feature E, where the pie charts in Fig. 21(c) show the probability that the indicated segments belong to each class of seafloor.

Fig. 20. - Samples obtained in the surveyed area. The sample shown in (a) has a thickness that varies between 80 and 110 mm. The sample shown in (b) has an average thickness of 93 mm. Both samples were obtained from the outcrop near seafloors A and B. The scale in the photos is 100 mm in length.
Fig. 20.

Samples obtained in the surveyed area. The sample shown in (a) has a thickness that varies between 80 and 110 mm. The sample shown in (b) has an average thickness of 93 mm. Both samples were obtained from the outcrop near seafloors A and B. The scale in the photos is 100 mm in length.

Fig. 21. - Three-dimensional reconstruction of seafloor B, a 21-m-long region that consists mainly of exposed crusts and nodular deposits. The image in (a) shows the 2-D projection of the color information and (b) shows the bathymetry information of the same area. The results of (c) segmentation and (d) classification show that the algorithm can correctly identify regions of exposed crust and transition regions. The pie charts show the probability that the indicated segments belong to each class of seafloor. The features labeled D and E show details of an area of crust and a transition region, respectively.
Fig. 21.

Three-dimensional reconstruction of seafloor B, a 21-m-long region that consists mainly of exposed crusts and nodular deposits. The image in (a) shows the 2-D projection of the color information and (b) shows the bathymetry information of the same area. The results of (c) segmentation and (d) classification show that the algorithm can correctly identify regions of exposed crust and transition regions. The pie charts show the probability that the indicated segments belong to each class of seafloor. The features labeled D and E show details of an area of crust and a transition region, respectively.

Fig. 22. - Acoustic data for seafloor B showing (a) acoustic data compensated for vehicle motion, (b) acoustic data with respect to the surface of the seafloor, and (c) crust thickness measurements made using the peak detection algorithm, where the measurements in black are made over exposed crusts and those in red are made over transition regions as determined by the classification algorithm. In this case, the measurements in red were made over nodular deposits, and these measurements are rejected as they do not correspond to measurements of crust thickness. The labels D and E in (a) correspond to the features of the seafloor shown in Fig. 21(a).
Fig. 22.

Acoustic data for seafloor B showing (a) acoustic data compensated for vehicle motion, (b) acoustic data with respect to the surface of the seafloor, and (c) crust thickness measurements made using the peak detection algorithm, where the measurements in black are made over exposed crusts and those in red are made over transition regions as determined by the classification algorithm. In this case, the measurements in red were made over nodular deposits, and these measurements are rejected as they do not correspond to measurements of crust thickness. The labels D and E in (a) correspond to the features of the seafloor shown in Fig. 21(a).

Fig. 22 shows the acoustic data obtained in the same area. From the acoustic data, it appears as though the thickness of the crusts is initially constant, but starts to reduce after some distance. However, the results of classification indicate that the measurements showing the apparent change in crust thickness were made over the nodular deposits in regions classified as transition. These measurements, plotted in red in Fig. 22(c), are highlighted by the algorithm as requiring further human interpretation. It is noticeable that the reflections from the nodular deposits are stronger than for the crusts, and that the reflections occurring after the top surface are also strong. At this point, it is not clear how to physically interpret these acoustic measurements. One hypothesis for the increase in the strength of the signals observed is that the rough surfaces of the nodular deposits may increase scattering, which for nonzero angles of incidence, may result in a stronger signal being received than for less rough surfaces. The second reflection is thought to be the signal reflected directly from the seafloor on which the nodular deposits lie. However, in both cases, further investigations are necessary to verify if these interpretation are appropriate, and for the purpose of this study, the measurements made over the nodular deposits are rejected.

For measurements made over exposed crusts, the average thickness can be computed as 104.8 mm with a 7.0-mm standard deviation. This is similar to the thickness of the crusts in seafloor A, which is as expected, since both examples are part of the same crust outcrop. The section has a planar surface area of 30.00 m2, with 12.38 m2 (41.3%) of exposed crust, 0.12 m2 (0.4%) of sediment, and 12.15 m2 (40.5%) classified as transition. We can compute the mass of crusts scanned in this area to be 2.49 T. This corresponds to 201.2 kg per unit area of exposed crust, and 83.0 kg per unit area of seafloor in the region shown. The latter value indicates that although the crust is marginally thicker in this region than compared to seafloor A, since a smaller percentage of the seafloor consists of exposed crusts, the region as a whole is not as rich in terms of exposed manganese crust deposit.

C. Seafloor C

The final example seafloor shows a 70-m-long region that has intermittent areas of exposed crust, sediment, and transition regions. The data were measured near the center of the flat top of the seamount, about 30 km from the measurements of seafloors A and B. Fig. 23(a) and (b) shows the 2-D projection and laser bathymetry of the 3-D reconstruction of this area. The results of segmentation and classification [Fig. 23(c) and (d)] show that the algorithm successfully identified regions of sediment such as feature F, transition regions such as feature G, and areas of exposed crust such as feature H, where the pie charts show the probability that the indicated segments belong to each class. The scanned section has an area of 97.53 m2, with 21.27 m2 (21.8%) of exposed crust, 36.16 m2 (37.1%) of sediment, and 28.85 m2 (29.6%) classified as transition.

Fig. 23. - Three-dimensional reconstruction of seafloor C, a 70-m-long region that has intermittent regions of exposed crust, sediment, and transition. The image in (a) shows the 2-D projection of the color information and (b) shows the bathymetry information of the same area. The results of segmentation and classification are shown in (c) and (d), respectively. The pie charts show the probability that the indicated segments belong to each class of seafloor. The features labeled F, G, and H show details of an area of sediment, a transition region, and an area of exposed crust, respectively.
Fig. 23.

Three-dimensional reconstruction of seafloor C, a 70-m-long region that has intermittent regions of exposed crust, sediment, and transition. The image in (a) shows the 2-D projection of the color information and (b) shows the bathymetry information of the same area. The results of segmentation and classification are shown in (c) and (d), respectively. The pie charts show the probability that the indicated segments belong to each class of seafloor. The features labeled F, G, and H show details of an area of sediment, a transition region, and an area of exposed crust, respectively.

Fig. 24 shows the acoustic data obtained in the area. The results of thickness measurements in Fig. 24(c) made in the regions of exposed crusts indicate that the crusts have an average thickness of 68.1 mm with a standard deviation of 19.6 mm. Although several attempts were made to sample crusts from the outcrops, the attempts were unsuccessful and only nodular deposits could be obtained. The maximum thickness of the manganese deposits on the nodules was 50 mm, however, this is not representative of the thickness of the crust outcrop, and so it is not possible to verify the accuracy of the measurements made by the acoustic system in this area. Based on the measurements shown, the amount of crust in this area of seafloor can be calculated as 2.78 T, corresponding to 130.8 kg per unit area of exposed crust, and 28.5 kg per unit area of seafloor in the region shown. It is interesting to note that measurements made over several of the regions classified as transition gave similar values of thickness to measurements made over exposed crusts. It is possible that some of the transition regions have only a very thin layer of sediment on top of the crust. A thin layer of sediment would not significantly affect the subsurface measurements of the acoustic system, but would alter the appearance of the seafloor enough to change the results of classification. At this stage, even with human interpretation there is insufficient evidence to conclude whether this is actually the case, and so the measurements are rejected with only regions labeled as exposed crusts used to compute the amount of manganese deposit in the scanned area. Furthermore, although acoustic measurements made over sediments are not processed in this study, several of the signals recorded in regions of sediment, for example, between the features labeled F and G, have subsurface reflections. It is thought that these reflections are of the basement rock layer under the sediments, where the thickness of the sediment layer in this region is no more than a few tens of centimeters. A naive interpretation of the acoustic data may again lead to incorrect conclusions being made regarding the thickness of the crusts.

Table III Distribution of Crusts Together With Error Margins for the Three Example Seafloors
Table - Distribution of Crusts Together With Error Margins for the Three Example Seafloors
Fig. 24. - Acoustic data for seafloor C showing (a) acoustic data compensated for vehicle motion, (b) acoustic data with respect to the surface of the seafloor, and (c) crust thickness measurements made using the peak detection algorithm, where the measurements in black are made over exposed crusts and those in red are made over transition regions as determined by the classification algorithm. The labels F, G, and H in (a) correspond to the features of the seafloor shown in Fig. 23(a).
Fig. 24.

Acoustic data for seafloor C showing (a) acoustic data compensated for vehicle motion, (b) acoustic data with respect to the surface of the seafloor, and (c) crust thickness measurements made using the peak detection algorithm, where the measurements in black are made over exposed crusts and those in red are made over transition regions as determined by the classification algorithm. The labels F, G, and H in (a) correspond to the features of the seafloor shown in Fig. 23(a).

D. Discussion

The instruments and methods described in this paper can provide quantitative estimates of the distribution of manganese crust deposits based on in situ measurements. The accuracy of the measurements depends primarily on the resolution and accuracy of the sensors used, and variations in the properties of the manganese crusts. The accuracy of the 3-D mapping system has been determined experimentally in [34] in pool trials scanning a submerged model of known dimensions. The errors were determined to be between 6.1% and 9.0% across the swath, about 5.7% in the vertical direction, and between 1.0% and 1.7% along the direction of travel. The main source of error across the swath and in the vertical direction is due to lens distortions that are not completely corrected even after lens calibration. Possible methods to reduce these errors include using a lens with a smaller opening angle, which would come at the cost of swath, or introducing a lens model with higher dimensions than the standard rectilinear lens model used currently. The error along the direction of travel is determined by vehicle navigation. In the present application, vehicle position is estimated through dead-reckoning using DVL and compass-based INS data. There exist a number of methods that can improve vehicle position estimates, including methods such as simultaneous localization and mapping (SLAM) using bathymetric measurements [35] and visual features [36], [37], which could potentially be applied to improve mapping accuracy and overcome time-dependent error growth by planning partially overlapped track lines during the survey. The accuracy of the acoustic measurements is determined principally by variations in the properties of the manganese crusts, where the relative standard deviation of the compressional wave velocity is 6.1%. Furthermore, when calculating the mass of the deposits, it is necessary to account for the 1.9% relative standard deviation in the density of the manganese crusts. By combining the geometric mapping errors with the relative standard deviations of the relevant properties of the manganese crusts, a conservative estimate of the standard error in the distribution of manganese crust deposits can be made. The results of the measurements and the standard error margins for each of the three seafloors described are summarized in Table III.

SECTION VIII.

Conclusion

In situ measurements of manganese crust thickness have been successfully performed for the first time, and it has been demonstrated that the acoustic and visual instruments developed in this study can provide a solution to the problem of noncontact, continuous measurement of the distribution and thickness of manganese crusts. The acoustically measured values for crust thickness were consistent with the thickness of crust samples obtained in the surveyed areas, demonstrating that continuous remote measurement of manganese crust thickness is possible for the substrate rock types found in the surveyed region. While there is an insufficient number of samples analyzed to draw any statistical conclusions, the acoustic properties of the samples obtained at #5 Takuyo seamount suggest that highly phosphatized carbonate or carbonate matrix, relatively unaltered basalt substrates, and even lightly phosphatized carbonate and carbonate matrix and moderately altered basalt substrates have sufficient contrast in acoustic impedance to expect detectable acoustic reflections at the crust–substrate interface. However, it is also evident that the acoustic properties of the substrates are highly dependent on the rock type, and it is important to consider the substrate geology when surveying manganese crusts in this manner. The visual instrument described in this study is capable of performing high-resolution measurements of the seafloor in color, and this information can be used to determine the nature of the seafloor where the measurements were made. An algorithm to process the 3-D data was developed and implemented, and it has been demonstrated that the system is capable of identifying regions of exposed crust using data obtained at #5 Takuyo seamount. The mapping system plays an important role in interpreting the acoustic measurements since the distribution and nature of manganese deposits can change abruptly, and in at least two of the data sets analyzed in this work, a naive interpretation of the acoustic data on its own would lead to incorrect conclusions being drawn regarding the thickness of the crusts. In addition to confirmation of exposed crusts, the measurements of the mapping system can provide statistical information concerning the distribution of manganese crusts, and through combination with acoustic measurements of crust thickness, enable quantitative analysis of the volumetric distribution of manganese crusts. Through combination with existing sampling survey methods, it is expected that the proposed instruments and data processing techniques will allow for more efficient survey of manganese crusts, and enable more accurate estimation of the volume of manganese deposits on regional scales.

ACKNOWLEDGMENT

The authors would like to thank the Hyper-Dolphin team and R/V Natsushima crew, Japan Agency for Marine-Earth Science and Technology (JAMSTEC, Yokosuka, Japan), during the NT09-02 Leg 2 and NT10-11 cruises. They would also like to thank Dr. A. Usui of Kochi University, Kochi, Japan; Dr. T. Urabe and A. Tokumaru of the University of Tokyo, Tokyo, Japan; Dr. Y. Yano of Hakuyodo Inc., Tokyo, Japan; Dr. R. Bahl of the Indian Institute of Technology, New Delhi, India; and Dr. T. Nakatani of JAMSTEC for stimulating discussions and for their advice concerning this work.

References

References is not available for this document.