FINAL REPORT Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO SERDP Project MM-1533

This effort has examined the problem of detection and classification of buried munitions in underwater environments. We have focused on the use of low frequency sonar since high frequency acoustic waves are strongly attenuated by sediments. The focus of this effort has been to process low-frequency data collected from the Buried Object Scanning Sonar (BOSS) into 3D imagery using beamforming, and to develop target/clutter classifiers that use 3D features extracted from this imagery. The principal sonar data sources are BOSS deployments at various shallow water sites. Morphological processing was applied to the derived imagery for feature input into a relevance vector machine classifier. Since ground truth was available, it was possible to compute performance metrics in the form of ROC curves. To enable a systematic understanding of the influence of the environment on target responses, we have developed a poroelastic spectral element method for BOSS data simulations using 2D and 3D models. The classification results establish that buried targets have a high probability of detection with the Buried Object Scanning Sonar. However, features from target imagery responses are easily confused with those of clutter and munitions debris due to their incomplete separation. Small subsets of possible imagery features show the best performance, and various examples are shown. We provide a theoretical development for the estimation of structural acoustic resonance features from BOSS-like data. Future classification performance gains with the sonar modality will likely rely on the combined use of imageryand resonance-based features. Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 1 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 Executive Summary The principal objectives of this research effort are to evaluate the effectiveness of synthetic aperture sonar (SAS1) for detection and classification of underwater munitions, characterize the influence of sediment heterogeneity on buried target detection and discimination, identify processing methods to improve performance of detection systems, and provide recommendations for future algorithm and system development. In the remainder of this summary we provide a description of the environmental problem addressed, the scientific questions that we have explored, the cumulative results received to date, and potential future applications of the research. Environmental problems addressed Military training and weapons testing activities in the past have left the legacy of UXO (Unexploded Ordnance) at a number of sites. This problem is of even greater concern for those sites designated for base realignment and closure and at Formerly Used Defense Sites. Particularly difficult is the characterization and remediation of those sites where UXO is found in underwater environments. Many active and former military installations have ordnance ranges and training areas that include adjacent water environments. (e.g., ponds, lakes, rivers, estuaries, and coastal ocean areas). Wartime activities, dumping, and accidents have also generated significant UXO contamination in coastal and inland waters. Dredging projects frequently encounter UXO, and potential hazards to the public from underwater ordnance encounters are beginning to arise. Much of the U.S. underwater contamination has occurred near military practice and test ranges, and potential hazards to the public from underwater ordnance encounters are beginning to arise. Presently, there exists no effective capability to survey these underwater areas and map the location of UXO for site characterization, and little understanding of the UXO or clutter characteristics from which to establish performance requirements. Factors such as small target size, target burial, natural and man-made clutter and target signature modification due to target-environment acoustic coupling all impact sensor performance. See the recent underwater UXO workshop2 for additional information. This SERDP research program is designed to determine the potential effectiveness of SAS for detection and classification of UXO in complex, underwater environments. A major goal is to acquire a detailed understanding of the influence of sediment heterogeneity on SAS data and the corresponding influence on the performance of detection and classification algorithms. The approach we adopt is three-fold using (i) processing of real SAS data collected for buried and surface target fields in underwater environments, (ii) high fidelity simulations of SAS data for targets embedded in 3D realizations of sediment properties and (iii) post-processing of these data sets for design and performance evaluation of future automatic target recognition (ATR) algorithms, discovery of target features that are robust and repeatable, and insights for successful survey design. These data sets provide will provide rich coverage over a variety of experimental operating 1The term synthetic aperture is taken to include any style of coherent processing of data streams collected from moving sensor platforms. Hence, this encompasses (i) the traditional synthetic aperture sonar (SAS) processing most commonly applied for collection trajectories along a linear path (with suitable motion compensation for small departures from the path) and (ii) generalized beamforming in which data is coherently processed for arbitrary collection geometries e.g., as in backprojection, but for which the source and receiver positions are known. 2SERDP and ESTCP. Final report, SERDP and ESTCP workshop on technology needs for the characterization, management, and remediation of military munitions in underwater environments. Technical report, October, 2007. Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 2 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 conditions and environments, and therefore, can be analyzed to yield a detailed physical and statistical understanding of SAS effectiveness for detection and classification. An additional goal is to quantify the performance of near-optimal classification algorithms using real data acquired under other programs with the Buried Object Scanning Sonar (BOSS), and features derived from 3D SAS beamforming products generated using this data. A final goal is to optimize the design of these algorithms to maximize the ROC performance metric for a selected operating point. Scientific questions explored Sonar is a natural candidate for UXO detection in shallow water due to its wide-area surveillance capability and target sensitivity. However, sonar signature interpretation is complicated by a number of factors including (i) natural and man-made clutter, (ii) data dependence on viewing geometry and target state, (iii) environmental heterogeneity and wave propagation complexity, (iii) coupling of target response with the environment, and (iv) sensor positioning and motion compensation requirements. This study addresses the above factors with a special concentration on items (i) – (iii), and uses these findings for improved design of classification algorithms. The evaluation for sonar-based buried target detection and discrimination was achieved by creating and analyzing a comprehensive catalogue of SAS processing results for shallow water environments. The questions explored included the detectability of buried targets with low frequency sonar, the identification of features useful for image-based discrimination, the design of classifiers using these features, and simulation methods useful for understanding target scattering phenomenology in complex environments. We used BOSS SAS data collected in AUV measurement campaigns (separately funded from this effort), phase histories from high-fidelity simulations (based on the Spectral Element Method and on the T-Matrix method) and environment models derived from ONR-sponsored environment characterization efforts (e.g. the Seismic Acoustic Experiments in 1999 – SAX99, which were specifically designed to improve understanding required for detection and classification of objects buried in sediments). Cumulative results achieved under this effort Our technical approach has yielded advances in sonar modeling capability, SAS data products derived from processing of BOSS data from various collections, a feature database for target and clutter derived from BOSS data and classification tools that exploit this data to achieve discrimination capability. In summary form, these results include • Development of the theory for poroelastic wave propagation in a form suitable for implementation with ‘weak’ forms of the governing equations of motion including the spectral element method (SEM). This effort led to several publications e.g., [27, 29, 4]. • Numerical implementation of the 2D and 3D poroelastic formulations of the SEM. Comparison of the simulation to various benchmarks demonstrated a correspondence of these simulations with the exact analytical results for the selected end-member cases for which such results were available. • A new SPECFEM2D package was released in 2009 and incorporated the poroelastic and adjoint capabilities developed under this effort. The source code has been made publically Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 3 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 available and it can be downloaded from the following web site: http://www.geodynamics.org/cig/software/packages/seismo/specfem2d/ • A 3D version of the code with the poroelastic upgrade is expected to be released in the future. It will be known as SPECFEM3D Sesame, which includes CUBIT compatibility. • Generalized derivation of the T-matrix scattering formalism for free-field targets and for targets buried in a multi-layered medium. In addition, an inverse theory formalism for estimation of T-matrix coefficients was posed, and selected numerical simulations for the BOSS source/receiver geometry were performed. • Processing of data from three separate BOSS data collections (AUV FESTs 2006, 2007 and 2008). Processing results included 3D beamformed imagery and features extracted from target and clutter detections in the imagery. Fifty feature types were extracted including geometric, intensity and statistical descriptors. • Development of a Relevance Vector Machine classifier using extracted features from the BOSS beamformed data products. The classifier used the features to label detections in the generic binary categories of target versus clutter. Access to ground truth knowledge of the object detections enabled the statistical training of the classifier as well as computation of performance metrics. In this case classifiers and ROC curves were developed for various subsets of the available feature classes. Restricted feature set sizes were used to improve generalization performance. Potential future applications of the research The research results achieved to date have numerous potential applications. The poroelastic extension to the spectral element method has been published in various forums [27, 29, 4]. The publically available code can be used to investigate influence of background propagation models and heterogeneity on wave propagation and scattering from targets within these environments. Investigators can use this data, for example, to design classifiers and identify optimal data collection strategies. The drawback to the method is the computation time, so questions or applications that depend on it should be designed accordingly. The SEM is well–suited to investigate scientific questions that require the high-level of fidelity supported by the code, and for which the issue of computation time is not a paramount concern. For example, it can be used to assess the validity of approximating poroelastic media with acoustic or acoustic/elastic models, and to characterize how scattering physics will differ for targets embedded in such different types of media. Examples of this were shown in [27] for the case of line array recordings for a target embedded in a two layer medium with differing layer types and material parameters. A second appropriate SEM application would be to compute off-line scattering properties for use in an on-line data exploitation system, as discussed in more detail below. Our combined work in the SEM and the T–matrix methods provides the opportunity to use the T–matrix approach for on-line applications, but with enhanced fidelity by computing key inputs off-line using the SEM. The SEM code can be used to simulate high fidelity scattered fields for a detailed CAD model of a target of interest. The developed T–matrix theory is developed under the assumption that the target is entirely confined to a single layer of a multi-layered medium, Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 4 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 but is otherwise quite general. Using the inverse theory that we developed in Sec. 3.2.11, T– matrix coefficients can be estimated from the scattered data (given sufficient receiver coverage), and these coefficients can then be used on-line for very rapid synthesis of sonar time series for use in either a automatic target recognition algorithm, optimal survey design exercise, assessment of system designs or parameter selection, etc. Further, the synthetic time series may be further processed to yield 3D beamforming products from which features may be extracted. The utility of these features for classification or discrimination can then be assessed in combination with features from real data for improved classifier design. The power of this approach is that the T–matrix coefficients are intrinsic to the target. Additional factors that affect observed recordings such as target/sensor geometry, source waveform content, medium layering, etc., are extrinsic. Once the T–matrix coefficients are available from the off-line computation, synthetic series can be rapidly simulated on-line for a wide range of extrinsic parameters e.g., number and material properties of layers, source and receiver positions, and so forth. The theory for this is defined in detail in Sec. 3.2, and simple numerical results are shown in Sec. 4.2.1. We have created a database of features derived from target and clutter objects using the BOSS acquisition system. This data set can be used by others for development of advanced classifier algorithms, and the data will be provided to SERDP with a specified format for this purpose. In addition, we have generated a database of 3D beamformed image products for the BOSS data collections. The features we derived were obtained from processing of these image products (see Table 7 for a summary description of these features). However, additional feature types that we have not considered may be derived from the image products, and these may help improve classifier performance. Additional features may be derived from 3D tomographic estimates of the sediment properties (see Sec. 3.4 and [29, 28]). Features derived from the canonical beamformed products may be used to spatially cue where tomographic inferences should be spatially culled for feature construction. Morphological processing and tomographic estimation both represent promising additional research possibilities for the future. The kernels for tomography have been derived and numerically implemented, but the approach has not yet been applied to real data. We have developed and demonstrated a classifier algorithm for buried targets using the feature database mentioned above. We developed the algorithm based on the relevance vector machine and characterized its performance for numerous feature subsets (20 in total). These results are shown in Sec. 4.6. There is considerable additional work that can be pursued including the application of a wrapper around the training algorithm itself that adaptively optimizes the feature subset. This is basically a combinatorial optimization problem involving the discrete index selection from a large superset. Various algorithms are applicable to this including greedy methods such as sequential forward method, the sequential backward method, and global methods such as genetic algorithms. The key here is training of the RVM is very fast (less than one minute using a single 2.1 GHz chip), and therefore, the model space can be adequately explored. To generate the feature classes described above, we applied morphological processing to the 3D beamformed image products. The latter were generated using the image formation algorithms described in Sec. 3.3. However, there were a number of simplifying assumptions used in the described algorithm, and there is the opportunity to enhance the quality and sharpen the focus of the derived data products. This can follow from the inclusion of any number of approaches including exploitation of available environmental information (e.g., by inverting the BOSS data itself for the best fitting 1D sediment model, and to use this in the specification of the beamforming weighting coefficients). An additional strategy to improve image product data is to apply more advanced Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 5 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 beamformers than the simple weighted delay and sum technique that we have used. Candidates for more advanced beamformers include the Capon and Minimum Variance Distortionless Response (MVDR) methods [36]. Finally, the results achieved in this program and the future research directions suggested above should be of interest to Mine Counter-measures (MCM) programs at the ONR. The work is physicsbased, and new methodologies have been introduced making application to mine targets relatively straightforward. Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 6 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 1 Objective Our objective has been to generate a large catalogue of SAS images using real data from multiple data collections acquired elsewhere with the BOSS system, and synthetic data from sonar phase history calculations. Forward simulation methods that are commonly used include Kirchhoff-based approaches and various implementations of T-matrix theories. These methods also are the basis of the simulation approaches supported in PC SWAT [34]. While very useful for insight to target responses, the methods require either homogeneous or layered environment models and therefore do not readily capture target-environment interactions arising from natural clutter. Thus, an objective of this effort is to extend sonar modeling capability using a state-of-the-art geophysical simulation method known as the spectral element method (SEM) (see, e.g. [20]). The SEM captures arbitrary target/environment complexity and accurately models all wave phenomena (resonant modes, surface waves, diffractions, specular scattering, target-environment coupling) over the low and high frequency regimes of interest. This enables careful characterization of realistic environment imprints on target signatures and signature variability in SAS imagery. Shallow-water environment models can be derived from the outcomes of the extensive SAX99 experiments ([24, 9]. Our research focus is SAS-based target recognition of buried munition objects in shallow water environments. The latter consists of the fluid column, the water-sediment interface, and the embedding sediments. The stratification of these components and the heterogeneity within them originating from depositional, biological, oceanographic and other processes strongly control wave behavior and the corresponding sonar observations. Target scattering processes are coupled to propagation phenomena in the environment, and will therefore obscure and/or modify target signatures predicted for targets emplaced in idealized media (e.g. a homogeneous two-layer halfspace). Further, target responses and wave-field interactions with the environment can change significantly depending on the frequency content of the active source signal. Generalized synthetic aperture sonar processing approaches are applied to SAS phase history observations to yield spatial reflectivity maps with cross-range resolution far greater than can be achieved with real aperture processing alone. In addition, SAS processors attempt to compensate for effects that would otherwise degrade the image resolution (e.g. platform position perturbations from assumed straight-line trajectories). Nonetheless, there will always be a data imprint of shallow water complexity on target signatures due the intrinsic acoustic/elastic coupling of targets and the environment. This will vary widely according to the target type and emplacement, environment conditions, sensor array design and source bandwidth, collection conditions, measurement campaign protocol and type of SAS-processor applied. Natural and man-made clutter objects (both buried and proud) pose an additional challenge to the discrimination problem. The success of sonar-based discrimination algorithms will depend on the distinctness of UXO target and clutter object signatures (e.g. their separation in feature space), their stability/measurability in the complex shallow water environment and the suitability of detection/discrimination algorithm design given the above. In summary, our objectives are to provide • a knowledge-base consisting of SAS imaging data products, features and discrimination measures derived from the SAS products, and shallow-water environment models and UXO target models represented with hexahedral elements (required by the SEM below) • a high-fidelity spectral element numerical method for modeling wave propagation and target Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 7 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 scattering in the complex shallow-water environment (and linked to the knowledge-base); the purpose of this is to permit a thorough understanding of the effect of realistic environment complexity on target signatures, to enable numerous data realizations consistent with the statistics of the environment characterizations, to assess the corresponding observability and stability of target signatures/features, and to provide a valuable resource for future discrimination algorithm development and • a comprehensive analysis synthesizing findings from the real and synthetic data results including recommendations for future directions. Our technical approach combines informed use of both real and simulated data inputs to assess sonar-based target detection/discrimination capability. Practical data processing of real data sets invariably requires use of simplifying assumptions, especially for propagation effects in complex environments. The influence of the environment and these assumptions is difficult to untangle for real collections. In terms of target classification, any practical deployment will most likely use predicted target features for a target buried in a simplified background medium. By performing controlled high-fidelity simulation experiments, it will be possible to precisely quantify the influence of simplifications used in the practical data processors on the ultimate classification performance. For example, realistic simulations can be performed for a fully realized complex shallow water environment with a buried target. The processing sequence (such as that described for BOSS below) can then be applied to the synthetic data set using homogeneous background assumptions. In this way, robustness of features used for classification with respect to environmental and target variability can then be precisely characterized. Numerous stochastic realizations of such environments can then be used to provide a meaningful statistical understanding. In summary, our technical execution plan consisted of the following key steps: 1. Process data from the BOSS collections for phenomenology understanding and empirical characterization of performance. Data was processed from known target fields in which clutter and targets could be separately identified on the basis of prior knowledge. 2. Extract features from BOSS image products and derive classifiers using these features. 3. Perform high-fidelity numerical simulations of the acoustic phase histories for standard SAS data collection experiments including BOSS. The spectral element method was used for this purpose. Perform data product generations (e.g., SAS images) using the phase histories from the resulting simulation. 4. The original objective was to perform SEM simulations for many different stochastic realizations of the propagation models, and for each corresponding SAS image, determine the target signatures and features. The objective of this is to explicitly determine the coupled medium-target influence on the SAS image, to determine the stability of target-related features and the degree to which they can be identified and extracted in the presence of the environment complexity and variability. However, the extended development time of the SEM for poroelastic media, and the comptutation time required for each simulation prevented this particular objective from being fully realized. Sensor Phenomenology and Feature Development for Improved Sonar-based Detection & Classification of Underwater UXO 8 BAE SYSTEMS AIT Applied Signal Technology, Inc. Princeton University BAE SYSTEMS AIT SERDP Project MM-1533 2 Background Over the past decade there has been a dedicated research effort for the development of detection and classification methods for underwater objects in shallow and littoral waters. The sonar modality has been a method of choice due to its far range, wide-area coverage capability and diagnostic value. Applications have included real-aperture side-looking sonar (SLS) and the higher resolution synthetic aperture sonar. Example theoretical and observational studies of acoustic responses for surface and buried targets include [8, 30, 6, 7, 5, 25, 16, 13, 17, 2, 22, 21, 19, 23, 34, 33, 14]. Figure 1 depicts some of the major wave processes that are involved. Many of these studies have focused on the sea mine discrimination problem; however, considerations for the UXO problem are