Improved Ship Classification in Littorals Through Sensor Fusion

Improved Ship Classification in Littorals Through Sensor Fusion

Lindgren, David

A Combined Hydroacoustic and Electromagnetic Surveillance System Provides Robust Performance in Difficult Environments

The hydrophone is by far the most used sensor in underwater surveillance systems. However, hydrographic and bathymetric conditions in littoral environments often restrict the operational range of hydroacoustic systems. High background noise levels due, among other things, to distant shipping are also common in these environments, which typically degrade the hydroacoustic performance. One way to make these surveillance systems robust to occasionally poor acoustic conditions is to combine them with other sensor types, and use sensor fusion to enable the system to rely more on the other sensor types when the hydroacoustic signal-to-noise ratio is low. Non-acoustic sensors may also provide additional target information.

An interesting supplementary sensor in this respect is the electrode sensor for extremely low-frequency electric fields modulated by propellers and the shafts of passing ships. This work is aimed at fused classification based on real-world hydrophone and electrode measurements. Can the displacement of passing ships be determined by this sensor combination? Two displacement classes are defined, and the objective of the test system is to classify the passing ships as either heavy or light. The task is approached both with and without fusion methods, with the purpose of identifying the potential benefits of sensor fusion. Fused ship classification is conducted in three major steps: feature extraction, classification and classifier fusion. The feature extraction aims to refine measurement signals to a compact and informative representation. The classification maps the extracted features to a class, either heavy or light. The classifier fusion finally joins classifications made separately on hydrophone and electrode signals to a unified result. The classifier fusion also joins classifications made at different time instances.

Of course, other schemes to solve this task exist, such as schemes where fusion of measurement signals precedes feature extraction and classification. However, it is believed that the decision fusion described in this article copes better with situations of varying signal-to-noise ratios.

Feature Extraction

As outlined above, the feature extraction is a preprocessing that serves to give the measurement signals a compact and informative representation. Computationally efficient classification and fusion can then follow up on refined data. Ideally, the features contain exactly what is needed to perform accurate classification. Different methods have different success in solving this task. However, in a given application, it is not always obvious which method is the most appropriate. Two candidates are compared in this work: the autoregressive model and the discrete Fourier transform.

The autoregressive model is adapted to a time window of a sensor signal by the tuning of a set of model parameters. The feature set is then defined as the parameters that best make the model fit the signal. The discrete Fourier transform maps the time window of a sensor signal to the frequency domain, where the response (power) at certain frequencies defines the feature set.


Each new feature set is classified by comparing it to feature sets in a class library. This library is based on feature sets where the ship class is already known. If the new features match the features in the library, it can probably be classified correctly. Here, a support vector machine attends to this task of maintaining a class library and matching features. The support vector machine is chosen due to its ability to adapt to complex feature patterns. Another virtue of the support vector machine is that it offers a compact representation of the class library.

Classifier Fusion

Through data fusion or classifier fusion, different sensors at different time instances are combined to a unified result or classification. Spatial fusion combines classifications from different sensors, and temporal fusion combines classifications from different time instances. The problem of data association will not be addressed in this article, which means that it is assumed that any set of classifications to be fused is based on signals from exactly one ship. Three classifier fusion schemes have been found to be particularly interesting to compare: the Bayes rule, voting and weighted voting. Using the well-known Bayes rule to fuse classifiers implicitly assumes that each classifier outputs the posterior probability for each class, rather than a hard classification. Under certain assumptions on independence, classifier fusion according to the Bayes rule boils down to forming the product of the individual classifier’s posterior probabilities, and then selecting a class according to the maximum product.

Voting is a fusion scheme where the number of separate classifications in favor of one class is compared to the number in favor of the other class. The finally selected class is the one in the majority. With the weighted voting scheme, only the most reliable classifications are allowed to vote. This requires that classifications are assessed with a measure of reliability, indicating how well the feature set matches the library. Posterior probabilities are used to form this reliability measure.


The data was recorded in the Baltic Sea at a sea trial in October 2000 using a multi-sensor platform with a hydrophone and electrode sensors. A total of 23 passages of 15 different surface ships with known displacements were recorded. All the ships are passenger and cargo ships in commercial, scheduled traffic with displacements ranging from 2.9 to 60 kilotons. Based on the displacement distribution in the data set, ships below 20 kilotons are labeled light, and ships above 20 kilotons are labeled heavy.

Can ship weight be estimated merely by the recorded hydroacoustic signal power? In this data set there are no apparent dependencies between weight and power, so the answer is no. Other signal processing methods are motivated.

The autoregressive model parameters are estimated with the Burg method. Model order 15 is used for the electrode data and model order 20 for the hydroacoustic data. Thus, the feature sets contain 15 and 20 elements, respectively.

When the discrete Fourier transform is used to extract features, the first 4,000 components of an 8,000-point transform are used. Principal component analysis is then used to reduce the feature set from 4,000 to 20 components for the hydrophone signal and to two components for the electrode signal.

Regardless of sensor type and feature extraction method, all data is translated and transformed prior to the classification, so that the features all are in the interval (-1,1), and so that the library data set cross correlations all are zero.

The support vector machine uses a nonlinear radial basis kernel.


With the limited data set at hand, a leave-one-out iteration over the 15 ships is conducted. Thus, the ship that is left out at each iteration is classified with a class library consisting of the remaining 14 ships. Three different tests are conducted: classifying each feature set separately without fusion, temporally fusing the classifications from a whole ship recording and spatially fusing hydrophone and electrode data, and temporally fusing over the whole ship passage,


First, the results using the autoregressive model are considered. Without fusion, the hydrophone alone gives a 13 percent error, while the electrode is not much better than fair coin flipping with a 45 percent error. By temporal-weighted voting, the performance improves from a 13 percent error down to 8.7 percent, and from a 45 percent error down to 35 percent, respectively.

Using both spatial and temporal fusion, the error can be as low as 4.3 percent, which is achieved by the weighted voting scheme. This is followed by the Bayes rule, which has an 8.7 percent error, and voting with a 26 percent error.

Now, the results using the discrete Fourier transform are considered. The hydrophone alone, in conjunction with temporal weighted voting, performs at as low an error rate as 4.3 percent. Fusion with the more uncertain electrode data does not improve this result.

In every instance with a 4.3 percent error rate, there is only one misclassified recording, always the lightest ship.


In the study, classification works slightly better with features extracted by the discrete Fourier transform compared to the autoregressive model. Furthermore, the hydrophone alone is more reliable than the electrode. However, when the autoregressive model is used, fusing the electrode with the hydrophone reduces the classification error rate from 13 percent to 4.3 percent.

The authors conclude that electrodes for measuring extremely low-frequency electromagnetic fields are interesting complements to hydrophones in a system for passive surveillance applications. In conjunction with signal processing and fusion methods, the electrodes can improve the reliability in situations of degraded hydroacoustic conditions.


The authors thank E. Dalberg and M. Levonen for valuable contributions to this work. This article is based on a contribution to the Marine Technology Society/Institute of Electrical and Electronics Engineers Oceans Conference in Boston, Massachusetts, in 2006.

By David Lindgren

Senior Scientist

Ron K. Lennartsson



Leif Persson

Deputy Research Director

Swedish Defence Research Agency

Stockholm, Sweden

David Lindgren received his M.S. degree in computer science and engineering and his doctoral degree in automatic control from Linköpings Universitet, Sweden. Since 2005, he has worked at the Swedish Defence Research Agency, where his research mainly concerns estimation and sensor fusion in networks and multi-sensor systems.

Ron K. Lennartsson received his M.S. degree in electrical engineering in 1998 from the Royal Institute of Technology in Stockholm, Sweden. Since 1998, he has worked at the Swedish Defence Research Agency. His specific research areas include statistical signal processing and data fusion for underwater applications.

Leif Persson received his B.S. and doctoral degree from Uppsala University, and his M.S. degree from Chalmers in Gothenburg, Sweden. Since 1984, he has worked at the Swedish Defence Research Agency, currently as deputy research director. His research interests include spectral analysis, statistical signal processing and data fusion.

Copyright Compass Publications, Inc. Jun 2007

Provided by ProQuest Information and Learning Company. All rights Reserved