Oil spill identification in X-band marine radar image using K-means and texture feature

Study area and experimental data

Experimental data

In our study, the “Sperry Marine” radar system was used to monitor and record the wave clutter signals. The signals output from the radar transceiver was directly connected to the computer processing system, and the image was displayed by the monitor after processing. With the rotation of the radar antenna, the radar system can digitize and store radar images of the sea surface. Table 1 displays the main parameters of radar. In this work, the experimental data is the X-band horizontal polarization image collected at 23:19 on July 21, 2010, by the teaching-practice ship Yukun of Dalian Maritime University (Fig. 1). The image size was 1,024 × 1,024 pixels with a detection range of 0.75 nautical miles (NM), and the actual area represented by each pixel is 7.36 m².

Table 1:

The parameters of radar.

Parameter	Value
Band	X-band
Detection distance	0.5/0.75/1.5/3/6/12/24 NMs
Range resolution	3.75m
Antenna type	Waveguide split antenna
Polarization mode	Horizontal
Horizontal detection angle	360°
Rotation speed	28–45 revolutions/min
Length of antenna	8ft
Pulse recurrence frequency	3,000 Hz/800 Hz/785 Hz
Pulse width	50n/ns/ns
Automatic image acquisition speed	28 images/min
Spatial resolution	7.36 m2

DOI: 10.7717/peerjcs.1133/table-1

Data preprocess

The collected experimental image adopts the polar coordinate system, which needs to be converted into the Cartesian coordinate system to facilitate subsequent processing. The image after the coordinate transformation takes the azimuth angle as the horizontal axis and distance as the vertical axis. The image size is 512 × 2,048 pixels (Fig. 2). The bright line in Fig. 2 is the co-frequency vertical interference noise, and the lighter-colored areas at the bottom of the image are wave echoes and oil film targets. The original image needs to be preprocessed to extract the oil film region successfully. The pretreatment process referred to the methods adopted by Xu et al. (2021a); Xu et al. (2021b), and the specific process is shown in Fig. 3. First, the vertical noise detection operator is used to convolve with the image in the Cartesian coordinate system. Second, the Otsu algorithm is used for detecting vertical noise processing. Third, the distance weighted linear interpolation method is used to suppress the vertical noise. Finally, the de-noised image is obtained (Fig. 4). In this article, the method of oil film extraction was studied using the de-noised image and based on the Matlab R2020b platform.

Experimental method

In order to realize an intelligent, and rapid marine oil spill monitoring method, effectively improve the efficiency of oil spill monitoring, we compared and tried to select a feature in GLCM with typical machine learning classification methods, and then it combined with the local adaptive threshold to effectively extract oil film information from the marine radar image. The experiment process is shown in Fig. 5. First, the denoised image was cut into slices the size of the local window. Second, the texture features of each slice were calculated based on the GLCM. Third, according to the texture features, the K-means clustering algorithm was used to extract the effective oil spill area from the cut images. Finally, using the Sauvola algorithm, the oil film was segmented, and the final extracted oil film region was overlapped on the image.

Texture feature extraction based on the GLCM

The texture is caused by the different physical attributes of the object surface and mainly reflects the diversity of grayscale or color information. Image texture is one of the attributes of images, which usually be represented by the gray distribution of a pixel and its surrounding spatial neighborhood (Mohanaiah, Sathyanarayana & Gurukumar, 2013). GLCM is a common method to describe texture by studying the spatial correlation characteristics of gray. It is defined as the probability that two pixels with step distance d and direction θ appear in the image (Barburiceanu, Terebes & Meza, 2021), which is expressed as formula Eq. (1), the mechanism of GLCM is shown in Fig. 6. Through the GLCM, the image texture features can be extracted, and Table 2 lists the calculation formula of each texture feature characterization quantity. (1)${\displaystyle p_{ij}=\frac{p\left(i,j,d,\theta \right)}{\sum _{i=0}^M\sum _{j=0}^{N}p\left(i,j,d,\theta \right)}}$

K-means clustering algorithm

The K-means clustering algorithm is an iterative algorithm based on unsupervised learning. It is widely used in data classification because of its simple realization, strong explanatory power, and effective clustering effect (Nitta et al., 2020). The algorithm process is as follows (Hartigan & Wong, 1979): (1) K initial cluster center c_i ( i = 1, 2, 3,…,K) is selected for the sample set S, where S = {x₁, x₂ ,…, x_n}. (2) The Euclidean distance between each object and K cluster centers is calculated, and the data object to the cluster that is closest to the cluster center is classified. (3) The mean value of data objects in each cluster is calculated, and the mean value is taken as the new cluster center. (2)${\displaystyle c_i=\frac{1}{\left|S_i\right|}\sum _{x_j\in x_i}{X}_j}$ where $\left|S_i\right|$ isthe total number of instances that are in cluster i.

(4) The distance of each data object to the new K initialization cluster centers is calculated and redivided.

(5) The next iteration proceeds until the object category stops changing and the clustering ends.

Local adaptive threshold segmentation algorithm

The local adaptive threshold segmentation algorithm is based on the distribution of pixel gray values in the window. The gray mean s(i, j) and gray standard deviation $m\left(i,j\right)$ were used to calculate the threshold $T\left(i,j\right)$. Compared with global thresholds, local adaptive thresholds can avoid segmenting the noise in images as well as deal with the problem of low-resolution images.

Niblack (1986) proposed a local threshold method for the image pixel-level process. This process involves adjacent pixel values within a region window. The threshold T(i,j) can be estimated as: (3)${\displaystyle T\left(i,j\right)=m\left(i,j\right)+ks\left(i,j\right)}$ where k is a constant.

Sauvola & Pietikinen (2000) improved on the Niblack algorithm. Its expression is as follows: (4)${\displaystyle T\left(i,j\right)=m\left(i,j\right)\times \left[1+k\left(\frac{s\left(i,j\right)}{R}-1\right)\right]}$ where R is the dynamic range of standard deviations, and k is the influence factor of standard deviation, which reflects the intensity of the influence of standard deviation on the threshold T(i, j). The range is between 0 and 1.

Phansalkar et al. (2011) modified the Sauvola local adaptive threshold segmentation algorithm, which is used for processing low-contrast images. $T\left(i,j\right)$ can be expressed as follows: (5)${\displaystyle T\left(i,j\right)=m\left(i,j\right)\left[1+p{e}^{-q\cdot m\left(i,j\right)}+k\left(\frac{s\left(i,j\right)}{R}-1\right)\right]}$ where p and q are constants.

After comparing the experiments, the Sauvola algorithm was chosen for oil film identification in our work.

Image slices

The size of the each slice is generally the common factor of the length and width of the image, and the commonly used sizes are 8 × 8 pixels, 16 × 16 pixels, 32 × 32 pixels, 64 × 64 pixels, 128 × 128 pixels, and 256 × 256 pixels (Xu et al., 2021a; Xu et al., 2021b). Considering the accuracy and efficiency, In this article, the de-noised image is cut into 64 × 64 pixels and 256 sub-images are generated.

Texture feature extraction and selection

According to the definition of the GLCM model and the requirements of eigenvalue calculation, parameters of the texture feature extraction algorithm based on the GLCM are selected as follows:

(1) Selection of gray level.

The commonly used gray levels are 16, 64, 128, and 256. To reduce the amount of computation, in our work 16 is chosen as the gray level.

(2) Selection of step length.

The step size adopted in this article was d = 1, that is, the central pixel and its adjacent pixels were calculated.

(3) Selection of direction.

Generally, θ is 0°, 45°, 90°, and 135°. However, considering that the differences between the four directions are not obvious, this article takes the average values of these four directions.

According to the formula in Table 2, the characteristic values of each texture were calculated for the de-noised image. GLCM was used to extract texture features of each slice. In this article, texture feature entropy was selected as the classification feature. Figure 7 shows the output result of the entropy value.

Image classification

The K-means clustering algorithm was adopted to classify the sliced images. The final clustering result of this algorithm depends on the arbitrary selection of the cluster center and the size of the K value. Different texture features have different abilities to identify targets. In our work, texture feature entropy was taken as the input feature of the classifier. In general, marine radar images mainly record the regions with valid waves, weak wave echo signals, and wave disturbance (Liu et al., 2021). To classify these three regions and lock the oil films position, the initial number of clustering K = 3 and the number of iterations 50 were set to classify images. The result of classification is shown in Fig. 8A. Black is the disturbance zone, white is the valid wave area and gray is the region with weak wave echo signals. Considering that the oil films appear on the valid waves, and the wave is at the bottom of the image, the white part of the classification result is retained and superimposed with the de-noised image. The oil film target region is finally generated (Fig. 8B).

Threshold segmentation

The Sauvola algorithm as used to segment oil spill images, the window size was set to 32 × 32 pixels with the dynamic range of standard deviations R = 128 and the influence factor of standard deviation k = 0.5. The oil spill image was segmented to obtain the initial resulting image, as shown in Fig. 9A. There are lots of small spots in the preliminary result. Because the oil film is usually continuous, the small areas of black spots were removed. Then the image is inversed, repeating the segmentation operations to remove small areas of white spots. The final oil film extraction result is obtained by superposition with the classification diagram (Fig. 9B). And the image is converted into coordinates. Figure 10 shows the oil spill identification results in the polar coordinate system.

Validation

In our work, the visible light image (Fig. 11A) and thermal infrared image (Fig. 11B) were used to validate the result. Because the collection time of the radar image was at night, it is impossible to capture the same real-phase offshore oil film with the visible sensor. Figure 11A shows the visible image near the acquisition location taken during the daytime, and it is obvious that an oil film exists. Figure 11B shows the oil film captured with the thermal infrared sensor at the same location as in the radar image, and the grayscale value of the area where it is located is slightly lower than the grayscale value of the neighboring area. Meanwhile, some scholars have conducted studies on this oil spill, and eight scenes of remote sensing images of the oil pollution impact area were collected by the HJ1 A/B satellite during the critical period of oil spill response (July 16, 2010–August 2, 2010) (Lan, Ma & Chen, 2012). These images from different sensors proved that oil spill films were present in the sea at that time.

Comparison of texture features

Multiple texture features were extracted using the GLCM method. Ma, Li & Niu (2010) selected texture feature correlation and mean as the texture feature index for oil spill extraction from SAR image. In our work, these two texture features were used as the input to the K-means algorithm to classify the images. Figures 12A and 12C are the visualizations of the texture feature correlation and the mean value. The extracted effective oil spill areas are shown in Figs. 12B and 12D. The results obtained by selecting these two features have the problem of missing oil film, while using the texture feature entropy as the input classifier, the oil film regions are effectively extracted, and the main two strip-shaped oil films remain intact, as shown in Fig. 8B.

Comparison of local adaptive thresholds

Xu et al. (2021a); Xu et al. (2021b) used the Phansalkar algorithm for the segmentation of marine radar images to extract oil films. Yu et al. (2017) adopted the improved Otsu algorithm for oil spill detection on SAR images. In our work, the improved Otsu and Phansalkar methods were used to segment the de-noised image to compare with the Sauvola method, the parameters of the Phansalkar method recommend k = 0.25, R = 0.5, p = 2, and q = 10, the window size of the Otsu method is set to 128 × 128 pixels., The Otsu algorithm for segmentation could not identify the oil film well, as shown in Fig. 13A. There were many noises mistakenly segmented out. As shown in Fig. 13B, oil film can be extracted by the Phansalkar algorithm. However, some false positive targets were produced. The final oil spill results extracted by the three methods are displayed under the polar coordinate system (Fig. 14), and the oil pixel number and the area of the oil are counted as shown in Table 3. The area calculation shows that the area of the oil film obtained by the Otsu algorithm is much larger than the extraction results of the other two methods, which is caused by a large number of false positive targets being misclassified into oil films. The extracted oil film area by the Phanlakar method is also larger than that of the Sauvola method. Therefore, the proposed Sauvola algorithm is superior to the other two algorithms in segmentation accuracy, and it is suitable for the extraction of marine oil films.

Comparison with other methods in oil spill identification

Xu et al. (2020) adopted the support vector machine and local adaptive threshold method to identify and extract oil spills from shipboard radar (hereafter referred as Method 1). In our work, we use the same method for oil film extraction experiments. First, the support vector machine method is used to distinguish waves from the background. Then the image is processed by image restoration techniques to generate a gray distribution matrix (Fig. 15A). Second, the gray distribution matrix threshold is set to “100” to obtain the effective wave monitoring range (Fig. 15B). Finally, the oil spill identification result is obtained after adaptive threshold segmentation of the effective wave area and removal of small spots (Fig. 15C). Also, in this method, the effective wave region is segmented by manually selecting threshold, which is not an intelligent approach, in other words, the option of threshold value affects the extraction range of the oil film directly.

Xu et al. (2021a); Xu et al. (2021b) conducted an oil spill extraction experiment using a combination of LBP and K-means (hereafter referred as Method 2). Here we use the same method to complete our experiment, the sliding window size is set to 128 pixels. The classification result, the effective oil spill range, and the final oil film identification are shown in Fig. 16, although this method can reject the interference of ship wake flow, there is a problem that many oil films are missing.

The final oil spill extraction results of Method 1 and Method 2 in the polar coordinate system are shown in Fig. 17. The area of oil films and the compute time are obtained in Table 4. Compared with the result in this article (Fig. 10) can get, the oil film area identified by Method 1 is slightly smaller than that obtained by the method used in this article. However, from the extraction effect, although the two strip-shaped oil films were completely recognized, there were also many small noises. The oil film area identified by Method 2 is larger than the result obtained by the method adopted in this article. The reason is that some false-positive targets were misclassified, and a large number of oil films were missing, the integrity of two-strip oil films is poor. In terms of computing time, Method 1 takes a long time, because when oil-water separation is performed by the support vector machine, foreground samples and background samples will be selected, which will consume a certain amount of time. Although Method 2 takes the shortest time, it is inferior to the method used in this article in extraction accuracy. In a word, the method used in this article avoids the problems of the above two methods. Two obvious strip-shaped oil films were extracted as well as fewer non-oil films.

Comparison of texture feature slice window sizes

The selection of texture feature local window size is 64 × 64 pixels, as Fig. 8A. The texture feature window size is reduced to 32 × 32 pixels, and the same method is used to extract oil film regions in Fig. 18A. The effective wave area is shown in Fig. 18B after classification. From the extracted results, the smaller texture feature window is unable to distinguish the oil film targets from the ship wake interferences, some invalid wave monitor regions may influence the segmentation as well. Table 5 shows the time consumption in different texture feature local window sizes, when the window becomes smaller, the required slice time and feature map generation time will increase. Therefore, in the selection of texture feature window size, we prefer 64 ×64 pixels to have the experiment.

Comparison with other machine learning classifiers

According to the same experimental process, the SVM and KNN classifier were used to process the radar image, to test the effect of the K-means adopted in this article. From the experimental results in Figs. 19 and 20, some invalid wave areas were classified, resulting in false positive targets and ship of wake in the final oil spill identification. In general, the classification results of these two methods depend on the selection of samples, which is subjective and random. So in terms of extracting effective oil spill areas, the K-means is recommended for classification in our study.