Detection of outliers in radar signals is a considerable challenge in maritime surveillance applications. High-Frequency Surface-Wave (HFSW) radars have attracted significant interest as potential tools for long-range target identification and outlier detection at over-the-horizon (OTH) distances. However, a number of disadvantages, such as their low spatial resolution and presence of clutter, have a negative impact on their accuracy. In this paper, we explore the applicability of deep learning techniques for detecting deviations from the norm in behavioral patterns of vessels (outliers) as they are tracked from an OTH radar. The proposed methodology exploits the nonlinear mapping capabilities of deep stacked autoencoders in combination with density-based clustering. A comparative experimental evaluation of the approach shows promising results in terms of the proposed methodology’s performance.
Detection of targets and outliers in radar signals is a research issue that has gained significant attention in the academic and industrial research community, mainly because of the important associated impact of relevant applications in surveying of large areas. High-Frequency Surface-Wave (HFSW) radars are a category of radars that operate at the frequency band 3–30 MHz and, in contrast with other radars, use ground wave or sky wave propagation and ionospheric reflections of the electromagnetic waves for target detection, which allows for achieving longer ranges, where microwave radars cannot perform [
The promising capabilities of OTH radars have attracted significant interest from the research community and have already resulted in various approaches (e.g., [ Different targets may present similar dielectric and frequency properties thus making it hard to make a clear distinction among them. Given multipath propagation effects of rough surfaces, scattering from some objects tends to overwhelm the weak backscattering of targets. Due to the changes in atmosphere and ground conditions, noise is added which can confuse the analysis of a radar signal. Ocean and ionospheric clutter generate noise especially for HFSW radars.
On a different note, the surge of deep learning and the great results it has produced in other signal analysis domains, such as computer vision, speech recognition, and natural language processing, create certain expectations regarding its potential efficacy in radar signal analysis applications. Deep learning allows computational models of multiple processing layers to learn and represent data with multiple levels of abstraction mimicking how the brain perceives and processes multimodal information, thereby implicitly capturing intricate structures of large‐scale data. Complex abstractions are learnt at a given level based on relatively simpler abstractions formulated in the preceding layer in the hierarchy.
The goal of this paper is to present a framework for detecting deviations from the norm in behavioral patterns of vessels (henceforth called
The remainder of this paper is structured as follows: Section
In the literature, several signal processing and machine learning methods have been investigated and proposed to acquire more reliable data with lower noise and extract semantic information from radar signals. Kouemou and Opitz [
Denoising techniques for radar signals include low level processing such as the median filter or other nonlinear convolution schemes [
Regarding deep versus “shallow” learning schemes, traditional machine learning techniques exploit shallow architectures; that is, they use a single layer for data/feature transformation, even in a highly nonlinear space. Shallowness refers here to the simplicity of these architectures that use only one (or few) layer(s) of processing, responsible for transforming the raw input signals or features into the problem-specific feature space. Instead, in a deep learning paradigm, the architectures are composed of many (deep) nonlinear processing stages [
The proposed methodology exploits the nonlinear mapping abilities of stacked autoencoders (SAs) [ The history of a naval vessel, in terms of speed, position, course, signal frequency, or other related data, provided by a ground radar, suffices to extract meaningful features. Unexpected deviation from the norm is observed for a few ships, denoted henceforth as outliers.
The approach is relatively straightforward: Given a set of OTH data entries, SAs are used in an unsupervised way to map the track history of any vessel into a compact and informative feature vector. Then, at any moment all tracked ships are projected into a new feature space and clustered using OPTICS [
Proposed approach flowchart.
Heterogeneous data, such as automatic identification system (AIS) data, high-frequency surface wave (HFSW) radar data, and synthetic aperture radar (SAR) data, have been exploited in research for maritime surveillance purposes [
The OTH radar data used for the setting and evaluation of the presented work was acquired by the HFSW STRADIVARIUS radar by Diginext [
On a different note, AIS is an automatic tracking system used for collision avoidance on ships and by vessel traffic services. AIS information supplements marine radar, which continues to be the primary method of collision avoidance for water transport. Vessels equipped with AIS transceivers can be tracked by AIS base stations located along coast lines. The International Maritime Organization’s International Convention for the Safety of Life at Sea requires AIS to be present aboard international voyaging ships with gross tonnage of 300 or more and all passenger ships regardless of size [
Clustering refers to the task of identifying groups or clusters in a dataset. In density-based clustering, a cluster is a set of data objects spread in the data space over a contiguous region of high density of objects. Density-based clusters are separated from each other by contiguous regions of low density of objects. Data objects located in low-density regions are typically considered noise or outliers [
OPTICS computes a Minimum Spanning Tree (MST) of the data, where edge weights represent pairwise distances. These distances are smoothed by a density estimator, called core distance. The core distance of a point
Density-based algorithms, traditionally, use the Euclidian distance metric [
Let
An autoencoder is a neural network that is trained to attempt to copy its input to its output. Internally, it has a hidden layer
Usually, training the autoencoder to perform the input copying task will result in
The learning process is described simply as minimizing a loss function, for example,
A sparse autoencoder is simply an autoencoder whose training criterion involves a sparsity penalty
The core idea of our work lies in using stacked autoencoders to capture a representation of the main patterns present in the data. By doing so, any outlier in data samples will not be explained well using that representation. In other words, outliers will have significant variations from the rest of the data.
The outlier detection is a combinatory threshold-based approach built on the interquartile range rule, as in [
OPTICS outputs (i.e., reachability distances of the ordered ships) are treated as a continuous signal, over which we identify the peaks. Peaks correspond to significant changes between the closest compared vehicles. As such, anything that varies from the norm has a peak, allowing the easy identification of a possible outlier. Then, we calculate a threshold value
In case that an outlier provides AIS data, the detection regarding that ship is ignored. At first, for a specific time instance, ships are ordered in a density-reachable way (Figure
(Best viewed in color) illustration of an instance of the outlier detection mechanism at a specific time moment.
As explained in Section
Figure
FOR each time instance
END
Initialize FOR each OTH tracked ship
END WHILE AIS vessels remain unmatched
END
(Best viewed in color) an illustration of the investigated ship trajectories. Ground radar trajectories are plotted in grayscale. The fading colors correspond to past times.
(Best viewed in color) illustration of matched trajectories between ground radar and AIS data (a) and matched trajectories despite the noise, due to minor course deviations (b).
The matching process is based on a voting mechanism. For each of the radar tracked ships
In the following subsections, we describe the dataset utilized for the experiments, the performance evaluation metrics employed, and the system setup details, before presenting the experimental evaluation of the proposed framework.
Data preprocessing creates a set of
Computational complexity of the different processing steps.
Processing step | Data preprocessing | Data mapping | Data clustering | OTH and AIS matching |
---|---|---|---|---|
Complexity |
|
|
|
|
The utilized dataset pertains to approximately 6 hours of data captured from the Mediterranean coast of France by Diginext in July 2016 in the context of the RANGER EU Horizon 2020 project. AIS data for the same period were also obtained for use as ground truth.
A total of 556 ship entries were in this 6-hour dataset. The following data provided entries are used: Longitude and latitude: position values provided in degrees. The typical range is Course and speed: course is calculated in degrees, typically in the range Doppler frequency: it is calculated in Hz, typically in the range Raw Rx azimuth: azimuth angle from the Rx site in the raw spatial grid (equivalent to the reception beam), typically in the range Local noise: noise level in the surrounding of the plot. It is calculated in dBm, in the range Global noise: background noise level of all range-Doppler map. It is calculated in dBm, in the range
Formally, a cluster analysis can be described as the partitioning a number of
The Calinski–Harabasz index (CHI) [
Calculated for each possible cluster solution, the maximum CHI value indicates the best cluster partitioning of the data.
The Davies–Bouldin index (DBI) [
For each cluster
The silhouette value is a measure of how similar an object is to its own cluster (cohesion) compared to other clusters (separation). The silhouette ranges from −1 to 1, where a high value indicates that the object is well matched to its own cluster and poorly matched to neighboring clusters. If most objects have a high value, then the clustering configuration is appropriate. If many points have a low or negative value, then the clustering configuration may have too many or too few clusters.
For each datum
The first step should be the definition of the feature space on which radar data are mapped. As a starting point, we investigated the dimensional space provided by PCA, maintaining 99.1% of the original variation. The adopted stacked autoencoder approach consists of three layers or four layers, depending on the PCA outcome. The loss function was the well-known mean square error [
Ships track history is composed of 9 consecutive frames, each containing all data as described in Section
OPTICS algorithm outcomes depend on the selection of minimum cluster size. We have investigated the clustering outputs assuming at least 2, 5, 8, 11, 14, 17, 20, 23, and 26 members in each cluster. Clustering over SA mapped data performed better than using raw or PCA mapped data, for most of investigated cases.
According to CHI (Figure
The impact of minimum cluster size (OPTICS input parameter) on Calinski–Harabasz index average score. Stacked autoencoders CHI scores are better in all the investigated cases, compared to PCA and raw data based clusters.
The next step was the investigation of DBI scores for the same minimum cluster size setup (Figure
The impact of minimum cluster size (OPTICS input parameter) on Davies–Bouldin index average score. Stacked autoencoders CHI scores are better in six out of eight investigated cases, compared to PCA based clusters, and five out of eight cases compared to raw data.
The last cluster performance metric was the average silhouette distance (Figure
Impact of minimum cluster size (OPTICS input parameter) on silhouette average values. Stacked autoencoders silhouette scores are better in five out of eight investigated cases, compared to raw based clusters.
Another significant performance metric is the average reachability distance itself. The smaller the reachability distance of a point is, the higher the density is around it. The core idea of the proposed approach is that only outliers should vary significantly from the norm, on the projected feature space. Thus, all the ships, minus the outliers, should have similar feature values, which results in reduced reachability distances.
Providing more training data allows SA to adjust the mapping process to the norm. As illustrated in Figure
An illustration of how the number of training paradigms affects the average reachability distances (OPTICS outputs). Raw data average RD value exceeds 10, in each of the cases.
Illustration of the training period span effect on the variance in reachability distances. Raw data RD variance exceeds 40, in each case.
Average number of generated clusters given various mapping approaches. In all of the investigated cases (i.e., different minimum cluster size), SAs provide more clusters.
Regardless of the adopted feature mapping approach, OPTICS outputs are at least four times less in value, compared to calculated RDs using raw data (see Figures
(Best viewed in color) comparison of OPTICS outputs over the same time instance, setting as minimum cluster size (a) 20 and (b) 26 ships. Stacked autoencoders result in more clusters than PCA or raw data that implies more peaks in the signal, which leads to more outliers’ detection.
The last step of the performance analysis provides empirical findings. In most of the cases, SAs mapped data results in detection of more outliers compared to the other approaches (Figure
(Best viewed in color) illustration of the detected outliers through time. Using SAs’ mapped data results in more outliers compared to the other approaches. Some of the selected outliers correspond to ships equipped with AIS transmitters.
There was the possibility of unwanted outlier identification. In particular, ships providing AIS data were considered, a few times, possible outliers. Figure
(Best viewed in color) illustration of the ships identified as possible outliers, while providing AIS data. Such cases are not considered as outliers.
In our article, a novel approach that identifies unexpected behavior in ship plot and track patterns, as captured by an OTH radar, has been presented. The core idea is the unsupervised development of a mapping process, which can project the raw data in a compact, lower feature space. Outliers projected to the same space should have significantly different values. Stacked autoencoders and PCA were used for the mapping process and compared against the exploitation of raw data, for the identification of unusual ship behavior. Density-based clustering algorithms (OPTICS) were employed for clustering-based outlier detection. Experimental results suggest that the approach based on SAs outperforms the other approaches in both generated cluster quality and outliers’ identification.
The authors declare that there are no conflicts of interest regarding the publication of this paper.
The research leading to these results has received funding from the European Commission’s H2020 Research and Innovation Programme, under Grant Agreement no. 700478 (RANGER project). The authors would like to thank all project partners for their collaboration and especially their partners from Diginext for the provision of the OTH and AIS data and documentation.