Karl Thomas Hjelmervik

Estimating temperature and salinity profiles using empirical orthogonal functions and clustering on historical measurements

Oceanographic climatology is normally estimated by dividing the world’s oceans into geographical boxes of fixed shape and size, where each box is represented by a climatological salinity and temperature profile. The climatological profile is typically an average of historical measurements from that region. Since an arbitrarily chosen box may contain different types of water masses both in space and time, an averaged profile may be a statistically improbable or even non-physical representation. This paper proposes a new approach that employs empirical orthogonal functions in combination with a clustering technique to divide the world’s oceans into climatological regions. Each region is represented by a cluster that is determined by minimising the variance of the state variables within each cluster. All profiles contained in a cluster are statistically similar to each other and statistically different from profiles in other clusters. Each cluster is then represented by mean temperature and salinity profiles and a mean position. Methods for estimating climatological profiles from the cluster information are examined, and their performances are compared to a conventional method of estimating climatology. The comparisons show that the new methods outperform conventional methods and are particularly effective in areas where oceanographic fronts are present.

Classification of acoustically stable areas using empirical orthogonal functions

Sonar performance modeling is crucial for submarine and anti–submarine operations. The validity of sonar performance models is generally limited by environmental uncertainty, and particularly uncertainty in the vertical sound speed profile (SSP). Rapid environmental assessment (REA) products, such as oceanographic surveys and ocean models may be used to reduce this uncertainty prior to sonar operations. Empirical orthogonal functions (EOF) applied on the SSPs inherently take into account the vertical gradients and therefore the acoustic properties. We present a method that employs EOFs and a grouping algorithm to divide a large group of SSPs from an ocean model simulation into smaller groups with similar SSP characteristics. Such groups are henceforth called acoustically stable groups. Each group represents a subset in space and time within the ocean model domain. Regions with low acoustic variability contain large and geographically contiguous acoustically stable groups. In contrast, small or fragmented acoustically stable groups are found in regions with high acoustic variability. The main output is a map of the group distribution. This is a REA product in itself, but the map may also be used as a planning aid for REA survey missions.

Predicting Sonar False Alarm Rate Inflation Using Acoustic Modeling and a High-Resolution Terrain Model

False alarm rates several orders of magnitude higher than the designed false alarm rate are frequently observed on active, low-frequency, towed array sonars. Increased false alarm rate originates from at least two effects: clutter and false alarm rate inflation (FARI). Clutter, or non-Rayleigh probability distributions of the matched-filter (MF) envelope, is often observed on high-resolution sonars, since too few scatterers are resolved for the central limit theorem to hold. FARI is a signal-processing-induced source of false alarms that occurs when the reverberation is nonstationary in the normalizer window. For example, the reverberation in the analyzed sample originates from a seamount, while most of the normalizer window falls to the side of the seamount, resulting in an underestimated background power estimate, and therefore, increased false alarm rate. By combining a fast and accurate acoustic model with a high-resolution terrain model, occurrence of FARI may be predicted. The described method outputs the modeled probability of false alarm, which is the probability that a false alarm is generated at a given location. The method is tested by comparing spatial concentrations of measured false alarms to modeled probability of false alarm. Comparison shows that a significant amount of false alarms is generated due to FARI, and that occurrence of FARI can be predicted given detailed environmental input.

In ocean evaluation of low frequency active sonar systems

Sonar performance measurements in the sea are always affected by uncontrollable and/or uncertain environmental conditions as sound speed variations, bottom topography or the acoustic properties of the sea floor. This paper presents a method to determine a sonar ‐ target geometry which minimizes the uncertainty in target signal excess due to environmental variability. An acoustic model is used to estimate the signal excess for a large number of sound speed profiles measured in the relevant area. The results are compared seeking a target range and depth where the estimated signal excess is robust with respect to the variation of sound speed profiles to be expected in the actual area. Results from sea trials will be presented as well as simulated examples used to demonstrate the robustness or sensitivity of the signal excess to environmental changes at different test geometries.

Time-calibrated estimates of oceanographic profiles using empirical orthogonal functions and clustering

Oceanographic climatology is widely used in different applications, such as climate studies, ocean model validation and planning of naval operations. Conventional climatological estimates are based on historic measurements, typically by averaging the measurements and thereby smoothing local phenomena. Such phenomena are often local in time and space, but crucial to some applications. Here, we propose a new method to estimate time-calibrated oceanographic profiles based on combined historic and real-time measurements. The real-time measurements may, for instance, be SAR pictures or autonomous underwater vehicles providing temperature values at a limited set of depths. The method employs empirical orthogonal functions and clustering on a training data set in order to divide the ocean into climatological regions. The real-time measurements are first used to determine what climatological region is most representative. Secondly, an improved estimate is determined using an optimisation approach that minimises the difference between the real-time measurements and the final estimate.

Finding Acoustically Stable Areas Through Empirical Orthogonal Function (EOF) Classification

Validity of sonar performance models is generally limited by environmental uncertainty, and particularly uncertainty in the sound-speed profile (SSP). Rapid environmental assessment (REA) missions, e.g., using gliders, and advanced ocean models may be used to reduce this uncertainty before sonar operation in hostile waters. This work shows how data from ocean models may be used for planning of REA missions. The area of operation is divided into oceanographically stable subareas using empirical orthogonal functions (EOFs) and different methods of clustering analyses on SSPs from the ocean model. The acoustic stability of each subarea is assessed using sonar performance modeling. Acoustically unstable areas are divided into smaller subareas. Acoustically stable groups are represented by a single SSP. A map of acoustically stable areas in the area of operation is the main output. Large, geographically contiguous groups indicate acoustically stable areas where frequent SSP measurements are unnecessary, e.g., low concentration of gliders. Small and noncontiguous groups indicate the opposite. Other applications include determination of suitable locations for sonar tests that require stable sonar conditions, and efficient optimization of sonar operation in acoustically stable areas.

A comparison of different machine learning algorithms for automatic classification of sonar targets

A well-known problem with modern anti-submarine warfare sonars with narrow beamwidths and wide frequency bandwidths, is the frequent occurence of false alarms, particularly in littoral environments. This increases the workload of sonar operators and also reduces the usefulness of automatic systems such as autonomous underwater vehicles, since their limited communication abilities hinder them from sharing large amounts of contacts. In this paper, four traditional machine learning algorithms are tested on sonar data with a high amount of false alarms together with synthetic submarine echoes. It is shown that some of the algorithms can outperform simple signal to noise ratio (SNR) thresholding by a significant amount, but that the performance is highly dependent on the parameter values chosen for each algorithm. These parameters are therefore investigated in order to determine their relative significance.

Demonstration of Comparison Methods for Ocean Model Validation

In recent years an increasing amount of oceanographic data has become available. This includes observations as well as data from hydrodynamic ocean models. Validation is required for establishing the necessary confidence in new sources of data. Generally ocean models and other data sources such as satellite imagery are validated by comparing the output to conventional observations or the output of established ocean models. Methods of comparison used in literature range from refined statistical methods to comparisons of snapshots. This work collects descriptions of some of the most widely used comparison methods. The capabilities and limitations of each method are demonstrated using examples from modelled and observed oceanographic data. The work has a particular focus on how to determine discrepancies on vertical gradients in the oceanographic parameters since acoustic propagation is sensitive to errors in the sound speed gradient.

Comparison of measured and modelled reverberation and echo level in littoral waters

Sonar performance models are used both when planning and carrying out anti-submarine warfare operations. The models allow the sonar operator to get an estimate of what ranges the sonar is able to detect an enemy submarine. For active sonar, the transmission loss and reverberation levels are important input to the sonar equations which are used to model the sonar performance. Here we run two different sonar performance models, Lybin and MSTPA, for an environment in the Norwegian Trench on the Western coast of Norway. A present pipeline is used as a target of opportunity. The reverberation and echo levels are modelled. The output of the two models are compared to each other, but also to measurements made in the same area. These measurements were collected during a NAT III experiment in 2002. NAT III was a joint research program that included TNO, Thales Underwater System, FFI, and the Dutch, French and Norwegian navies. The modelled reverberation compares fairly well to measurements, particularly in the flat region in the Trench. The complex terrain close to the coast is an exception, which is expected since only a single set of bottom parameters were used in the modelling. For improved modelling of this complex environment a detailed map of the bottom parameters would be needed. The two models differ strongly on the echo level, and both models fail to estimate the echo level. This may be due to the presence of a strong sound channel at the sonar depth. The oceanography in the area of the experiment is known for its strong variability, and the presence of such sound channel are therefire usually local. By removing the sound speed minimum in the profile, the estimated echo levels of both models improve.

Adjusting modelled sound speed profiles for use in sonar operations

Realistic descriptions of vertical sound speed profiles are essential for modelling underwater acoustic fields. Errors in the vertical sound speed profile will have negative impact on the acoustic propagation modelling. Ocean models provides vertical profiles of temperature and salinity, from which sound speed can be derived, covering large areas with high spatiotemporal resolution. Previous experience with ocean models shows that vertical sound speed profiles are difficult to model with sufficient accuracy to be useful for acoustic modelling. A method for adjusting sound speed profiles from an ocean model to better represent measured profiles is proposed. The method is based on replacing the mean sound speed profile from the ocean model data set with mean profile of an observed data set. The method is illustrated on data from the Norwegian coast where the coastal current causes well-defined fronts and eddies. The proposed method reduces the mean root-mean-square error in the model data, particularly in the upper layers. On the other hand, comparisons of observed and modelled sound speed profiles on a one-to-one basis, is still challenging in both space and time.