The objective of this project was to measure sediment properties and their uncertainties that affect munitions mobility, burial, and detectability in marine, estuarine, lacustrine and riverine environments. The long-term goal was to develop a rigorous geoacoustic sediment survey (GeoSS) method which would be carried out before a detection survey. The short-term goal of this project was to extend/modify the existing sediment measurement, processing, and inversion techniques to treat survey data in the shallow-water environments relevant to munitions detection.
The research team has previously demonstrated the ability to accurately infer complex sediment structures that occur in marine environments using acoustic sediment reflection data. The vast majority of the results to date are at single locations on the mid- to outer-shelf (80–180 m water depth). For munitions applications, however, the data analysis/inversion procedures and system design must be modified to consider large sequential data sets collected in very shallow water (less than 35 m). Thus, two major extensions were required to apply the acoustic remote-sensing approach to sediment assessment for underwater munitions detection: (1) address the theoretical and computational demands of sequential Bayesian inversion for many data sets along a survey track, and (2) design a prototype geoacoustic sediment survey system for the shallow water depths of interest.
The research team significantly improved several theoretical and computational aspects of sequential Bayesian inversion. Sequential algorithms, or particle filters, quantify the information content of consecutive data sets by considering results from previous data sets along a survey track to inform the importance sampling at the current point. Therefore, efficient transition between data sets was of paramount importance. The most significant new theoretical advance under this project was a novel particle filter that bridges between consecutive data sets via trans-dimensional parallel tempering. The success of this implementation is significant since it upgraded the particle filter to a unified trans-dimensional algorithm. In addition, the approach was intrinsically parallel which leads to efficient use of new computer hardware; in particular, the new algorithm led to full occupancy of high-performance graphics processing units (GPUs). Finally, a more informative prior for seabed complexity was introduced based on even numbered order statistics and a Poisson distribution. This prior prefered simple stratifications and avoided spurious layers that previously occurred due to data noise.
The new advanced sequential Monte Carlo algorithm was tested against simulated reflection data along a track. The advantage of using simulated data was that the seabed properties were completely known and thus performance of the algorithm could be quantified. The noise imposed on the simulated data as well as the source-receiver geometry were similar to measured data. The seabed environment was purposefully designed with challenging geologic features, including: change in geoacoustic properties within a layer along the track, change in the number of layers, an erosional channel, and two abrupt changes simulating a geological fault. These were designed to test the limits of the algorithm and in some cases exceed it. For example, researchers included a layer that pinches out gradually which has a thickness less than the reflection system was able resolve over a section of the track. In summary, the algorithm performed very well. The inversion was not able to resolve pinch-out near its apex, but the inversion was reasonably stable. The worst results were near the fault boundary where the inversion of several pings produced a poor representation of the true seabed. However, within three pings, the algorithm recovered to produce high-accuracy estimates of the seabed. Also, the reductions in runtime exceeded the Go/NoGo criteria by 50%, achieving 1500 data sets/week.
The new advanced sequential Monte Carlo algorithm was also applied to a measured data set. The importance of the application to measured data was that despite the care taken to make the simulation challenging, the physics used to generate the simulated data were the same as used in the inversion. That the physics/assumptions made in the inversion process were adequate for field surveys can only be tested with measured data. Researchers have made numerous such tests to date using single site data, but this was the first time they had used sequential data along a long (14 km) survey track. In summary, only 13% of the track could not be inverted due to physics limitations. The results showed an unprecedented degree of geoacoustic detail and quantified spatial variability along the track in both vertical and horizontal directions. Researchers mapped a low-velocity, low-density sediment wedge (in which a munition would likely bury) 1 m thick near the beginning of the track thinning out to about 0.1 m at the end of the track. Underneath the wedge researchers mapped a more competent layer which likely would have resisted burial. Researchers were also able to map a high velocity, high density sub-bottom erosional layer as it changed depth across the area. The quantitative mapping results were compared with two kinds of independent data: cores and seismic imaging. They had core measurements of velocity (at 400 kHz) and density at the start and end of the track, which showed generally close agreement. The greatest differences were in velocity where the core data were known to be biased by multiple scattering (due to the high frequencies used and the presence of shell material in the core of order or larger than the wavelength). The quantitative mapping results were also compared with independent seismic data for layer positions. The layer structure from the inversion (inferred from the frequency domain reflection coefficient) compared well with the time domain seismic data.
A fundamental aspect of this approach was the requirement of considerable care in the design and execution of the measurements. The information content for quantitative sediment properties in wide-angle reflection data was extremely high; however, exploiting this information content required achieving small measurement uncertainties in angle and amplitude. To achieve these small uncertainties for support of munitions detection/classification required a system adapted to very shallow water depths (less than 35 m). Researchers developed a high-level system design for a prototype geoacoustic sediment system for the SERDP munitions mission, that features a surface-towed 0.2-10 kHz source and a 35 m towed receiver array with some of the array populated by tetrahedral elements.
Fundamental physics and empirical observations indicate that the nature of the sediment structure plays an important role in the ability to detect/classify proud or buried munitions. Thus, foreknowledge of the sediment structure can substantively improve munitions detection and classification. The anticipated benefits of this research are aimed at providing that foreknowledge.
This project's success in obtaining accurate sequential inversions of geoacoustic properties in a complex sedimentary environment along a survey track sets the stage for testing the inversion with data sets collected in water depths less than 35 m. The development of high-level specifications for a prototype geoacoustic sediment system is a key step for construction and testing of such a system. The successful development testing and validation of a robust sequential inversion algorithm opens the door to analyzing survey data from that system.
In addition to SERDP-specific benefits, the progress made in sequential Bayesian inversion and experiment design can be used in other applications including scientific and applied problems, e.g., geohazard surveys.