The sea area accounts for approximately 71% of the earth. The sea is the most stable ecosystem on the planet. It has vast space and rich resources and plays a significant role in the global ecological environment regulation. With the development of industries worldwide, sea pollution is becoming increasingly serious. At present, sea pollution includes oil pollution, red tide, toxic substances, plastic pollution, and nuclear contamination. Oil pollution, which originates mostly from accidental and deliberate discharges into the ocean of hydrocarbon products from large ships/tankers and oil platforms, constitutes the predominant portion of sea pollution. However, ship and oil-rig accidents, such as the Prestige disaster in 2002 off the west coast of Spain and the more recent Deepwater Horizon well blowout in 2010 in the Gulf of Mexico, exert a considerable effect on the ecology and wildlife of the polluted area. Chronic illegal releases of oil from ships and offshore platform operations are extremely frequent along the main shipping route and over zones of offshore oil extraction structures (Brekke and Solberg, 2005; Zheng et al., 2010). The accurate detection and forecast of oil spill trajectory would be beneficial to fisheries, wildlife, and resource management for the monitoring and conservation of the marine environment. The detection and forecast are some of the most important applications for operational oceanography (Cheng et al., 2011).

The monitoring and detection of oil spills by using synthetic aperture radar (SAR) systems have received considerable attention over the past few years because of their wide area coverage and day and night capabilities. Several space-borne SAR systems are used for oil spill monitoring, including the SAR systems, which were launched by the Russian Space Agency, the European Space Agency, and the Canadian Space Agency (Topouzelis, 2008). A scan SAR narrow mode can monitor the Gulf of Mexico oil disaster (Marghany, 2015). Bern et al. (1993) monitored oil spill using ERS data. Pavlakis et al. (1996) monitored an oil spill accident happened in the Mediterranean Sea using the ERS data. Nunziata et al. (2011) employed fully polarimetric SAR measurements for a sea oil slick observation. However, at wind speeds below 2–3 m/s, a backscatter cannot be observed from the sea surface; hence, the signature of oil slicks is absent. Ideal conditions for the SAR imaging of oil slicks could be wind speeds between 3 and 10 m/s (Solberg, 2012). In addition, oil slicks dampen the Bragg waves on the ocean surface and reduce the radar backscatter coefficient, resulting in dark regions or spots in a SAR image. The complexity of the marine system radar detection of oil spills is highly susceptible to false detections or look-alikes, including natural films/slicks, grease ice, threshold wind speed areas, wind sheltering by land, rain cells, shear zones, and internal waves (Brekke and Solberg, 2005). The differences between oil spills and quasi-oil spills reflected in the SAR image exist. For example, the low wind speed region has several characteristics of larger area, irregular shape, and vague edges generally. Internal waves manifests itself on the radar image as the alternate stripes of bright and dark. Ship wakes tend to be V-shaped lines or strip lines in the SAR images. The efficiency of manual monitoring is low. Thus, this detection poses a fundamental challenge in discriminating between oil spills and look-alikes by an automatic method. In recent years, numerous approaches have been developed and tested to distinguish oil spills from other similar oceanic features in the marine SAR images (Nirchio et al., 2005). In general, the detection of oil spill from the SAR imagery includes two steps: image preprocessing and oil spill identification. The preprocessing can provide the regions of interest (ROIs) for identification by segmenting the SAR images. The second step mainly identifies between oil spills and lookalikes by extracting features and using several classification algorithms.

The main purpose of the preprocessing phase is to segment the SAR image that subdivides an image into its constituent parts and extracts dark regions. Image segmentation approaches are currently based on detecting discontinuities and similarities (Chaudhuri et al., 2012). These approaches include image segmentation algorithms, such as edge detection, thresholding, region growing, clustering, and neural networks. In this study, this step uses a histogram to obtain the ROI.

In the oil spill identification phase, several researchers detect oil spills from look-alikes by using multi features. Most published papers discuss features, such as texture features, gray features, and physical characteristics. Frate et al. (2000) considered the extension and shape of the dark object and used 11 geometrical characteristics. Zhang et al. (2008) extracted five textural features based on a co-occurrence matrix. Salberg et al. (2014) described several physical mechanism features of the backscatter. Guo and Zhang (2014) analyzed the importance of individual features and selected nine shape features. Chaudhuri et al. (2012) presented a statistical-based approach to detect the dark curvilinear features. Stine Skrunes et al. (2014) extracted two multipolarization features that can be used to distinguish between the simulated biogenic slicks and mineral oil types. However, feature extraction and applications by using decomposition technologies, such as Fourier and wavelet decomposition, are rarely discussed because basic functions are fixed in these approaches and do not necessarily match the varying nature of signals. Fourier transform can introduce false harmonic components that can cause the spread of the energy. Wavelet decomposition results largely depend on the selection of a wavelet basis function, so adaptability is weak. An empirical mode decomposition (EMD), as a new technology of a time frequency analysis, has multi scale and multi resolution characteristics. It is a process of constant loop iteration and is suited to solve the problem of nonlinear and non-stationary systems. It is applied to texture segmentation by analyzing the instantaneous frequency of the obtained components called intrinsic mode functions (IMFs) (Liu and Peng, 2005; Chen et al., 2012). Dong et al. (2010) used the EMD to extract several IMF energy features, such as inputted parameter of neural networks, to classify oil slicks and sea water. Nunes et al. (2003) developed an algorithm based on the bidimensional EMD (BEMD) to extract features at multiple scales or spatial frequencies. Nunes et al. (2005) applied the Hilbert-Huang transform and the BEMD for a texture analysis. Chen et al. (2012) combined a BEMD with a local binary pattern for a texture image classification. He et al. (2013) proposed a hyper spectral image classification method based on an improved BEMD. Chen et al. (2014) proposed a novel infrared small target detection method based on the BEMD. The present study proposes an oil spill detection method from the SAR imagery based on the BEMD. The framework for oil spill and look-alike classification based on the BEMD method is show in Fig. 1.

Four research institutions, namely, the University of Hamburg, the Russian Academy of Sciences, the National University of Singapore, and the National Central University, cooperated and competed on a project named “the tropical and subtropical ocean viewed by ERS SAR”. These institutions created a website to attract the attention of a broad readership in countries that lie in tropical and subtropical zones to the ERS SAR imagery acquired over the oceans and to motivate scientists from these countries to use the images for oceanographic and meteorological investigations (http://www.ifm.zmaw.de/fileadmin/files/ers-sar/index.html). These institutions assembled 180 sample ERS SAR images acquired over tropical and subtropical ocean areas that show typical sea surface manifestations of several oceanic and atmospheric phenomena. The oceanic phenomena that can be delineated on the SAR images of the sea surface include ocean surface waves, internal waves, eddies, oceanic fronts, underwater bottom topography, ship wakes, oceanic wakes behind islands, oil slicks, river plumes, and upwelling areas. The atmospheric phenomena include katabatic wind fields, land-sea breeze, atmospheric wakes behind islands and coastal mountains, gap winds, boundary layer rolls, convective cells, atmospheric gravity (or internal) waves, and rain cells. The data set contains several sea states, and all images contain a certain number of dark objects. These images are used to test the quality of our method. The samples of the images are shown in Figs 2 and 3.

The SAR aboard European remote-sensing satellites ERS-1 and ERS-2, which were launched in 1991 and 1995, respectively, provide the images of the earth’s surface with a resolution of 25 m. In the so-called “image mode”, the SAR images are taken along a 100 km wide swath to the right of the satellite track. The SARs aboard the two satellites possess identical designs. They operate at a frequency of 5.3 GHz, transmit and receive the radar signal at vertical polarization (VV polarization), and view the earth’s surface under an incidence angle centered at approximately 23°. In Fig. 2, the bright area in the image is caused by an anticyclonic warm eddy in the Gulf of Tehuantepec, which is located on the Pacific coast of Mexico. In this location, the water temperature is several degrees Celsius higher than that in the surrounding area (black area in the image). These eddies, which typically possess a diameter of approximately 200 km, are generated by strong winds that originate in the Gulf of Mexico and blow through a narrow gap in the mountain range in the Isthmus of Tehuantepec onto the eastern tropical Pacific Ocean. The dark streaks that are visible in the left-hand section in Fig. 3 result from the oil discharge from ships. In one case, the ship (white spot) is visible at the front of the oil trail. The dark streaky features that are visible in the right-hand section of the image could originate from oil seeps.

To evaluate the performance and feasibility of the proposed features in this paper, we generated two sets of samples, namely, oil spill and look-alikes. A kind of classifier is used to predict the category for each sample. The Mahalanobis distance is widely used in a cluster analysis and classification technologies. This paper used the Mahalanobis distance as a classifier and the algorithm as follows:

Step 1. Calculate the feature selected above vector ${\bar {\text{X}}}$ for each sample:

Step 2. Calculate the covariance matrix C to feature vector ${\bar {\text{X}}}$:

where E₁₁=E[(x₁–u₁)(x₁–u₁)], E₁₂=E[(x₁–u₁)(x₂–u₂)], E_1n=E[(x₁–u₁)(x_n–u_n)], E₂₁=E[(x₂–u₂)(x₁–u₁)], E₂₂=E[(x₂–u₂)(x₂–u₂)], E_2n=E[(x₂–u₂)(x_n–u_n)], E_n1=E[(x_n–u_n)(x₁–u₁)], E_n2=E[(x_n–u_n)(x₂–u₂)] and E_nn=E[(x_n–u_n)(x_n–u_n)], u_i is the expectation value of the ith feature, ${u_i}{\rm{ = }}E\left({{x_{{i}}}} \right)$.

Step 3. To use the Mahalanobis distance to classify a test point as belonging to one of two classes, we first compute the mean vector $\overline M _j^i(j{\rm{ = 1, 2)}}$ of samples known to belong to each class. This paper uses a training set as mentioned in Section 3.4. $\overline M _j^i$ is shown as below:

Step 4. Compute the Mahalanobis distance r_j to each class by

and classify the test sample as belonging to the class with minimal Mahalanobis distance. If r₁ is less than r₂, the test sample is the image of oil film; otherwise, the image is a look-alike.

A dataset of 100 samples was selected as a test data set for classification. The top 50 samples are oil spill images, whereas the remaining number belongs to the class of look-alikes. The Mahalanobis distance of each sample is shown in Fig. 10. As shown in the figure, 45 samples possess r₁ that is less than r₂ in the top 50, and 46 samples possess r₂ that is less than r₁ in the remaining 50. The right is 45 oil spills and 46 look-alikes. Therefore, the overall accuracy is 91%. And once the ROIs (80×80) have been selected, the classification step needs about less than 1 second to complete their actions.

The paper compared the proposed method with a typical method on several data sets to evaluate its performance. A typical approach for oil spill detection involves locating dark areas in the SAR image and then computing texture and shape features for dark segments to define whether the target is an oil slick or something else. The first step is to segment out oil spill candidates as selected by a local threshold method. The threshold value corresponds to the local minimum between the two peaks in the histogram. Second, the features are extracted from previous studies, as shown in Table 2. The described set of features extracted from the dark object is inputted to a Mahalanobis distance classifier, which estimates the probability of being an oil slick or a look-alike. Different methods were tested in the same data set. Table 3 shows a comparison of the proposed method with a typical method.

A new algorithm is proposed for detecting oil spill from the SAR image based on Hilbert spectrum of the BEMD. Regions of interest in the SAR image are obtained by a histogram analysis. Each dark path is decomposed by the BEMD, and its Hilbert spectrum and Hilbert marginal spectrum are calculated. Afterward, the features are extracted by a relief algorithm. To demonstrate the validation, we divided different data sets into two categories, which are used to test the accuracy rate of the proposed method by the Mahalanobis distance. The accuracy of the classification increased. On the basis of the above results, the feature analysis of dark formations that are proposed in this paper is feasible. This method can effectively detect oil spills from the SAR image.

Although the BEMD method is based on the data itself, thus presenting good adaptability, several aspects remain to be investigated with regard to the BEMD, including the end effect, the interpolation method, parametric forms, and normalization processing. In addition, more intelligent algorithms (the neural network, fuzzy clustering, etc.) and data sets (Radarsat-2, ship-borne radar, shore-base radar, etc.) will be used to verify the validation of the proposed method in the future.