Oil Slick Detection in X-Band Marine Radar Imagery: Leveraging a Boundary-Aware SBR Feature and an Improved Whale Optimization Algorithm

Abstract

Marine oil spills pose a persistent threat to marine ecosystems and coastal economies, and their rapid and unpredictable spread requires timely and reliable monitoring. In X-band marine radar images, oil slicks usually appear as low-contrast dark targets embedded in heterogeneous sea clutter, making accurate segmentation particularly challenging. To address this problem, this study proposes a training-free two-stage oil slick detection framework that combines an improved Slick Boundary Ratio (SBR) feature with an improved Whale Optimization Algorithm (WOA). First, the improved SBR feature is used to extract the oil slick region of interest (ROI). Then, the improved WOA is employed to determine the global threshold for oil slick segmentation. Experimental results show that the proposed method achieves accurate and spatially coherent oil slick segmentation in complex radar backgrounds, with an Accuracy of 99.36%, a Precision of 85.73%, a Recall of 84.42%, an F1-score of 85.07%, and an Intersection over Union (IoU) of 74.01%. These results indicate that the proposed framework can effectively suppress false positives while maintaining strong detection sensitivity, thereby improving segmentation robustness in low-contrast marine radar scenes. Owing to its training-free design, the proposed method shows potential for shipborne and coastal oil spill monitoring applications.

1. Introduction

Oil spill is a persistent threat to marine ecosystems and coastal economies [1,2,3], especially as offshore oil and gas development and global shipping continue to intensify [4,5,6]. Even medium-scale spills can spread rapidly under winds and currents [7,8,9,10], causing long-lasting ecological damage and substantial economic loss if not detected and contained in time [11,12]. Reliable, near-real-time monitoring of the sea surface is therefore a key component of modern marine environmental management and maritime safety [13,14,15].

Research on marine oil spill detection has evolved from manual interpretation toward automatic radar-based analysis. Early monitoring relied mainly on visual interpretation of aerial or satellite imagery, which provided useful observational information but was labor-intensive and susceptible to subjective errors [16,17]. With the development of synthetic aperture radar (SAR), automated dark-spot detection became a major research direction because oil slicks usually dampen short surface waves and appear as low-backscatter regions in radar images. Representative studies established automatic SAR oil spill detection frameworks by combining dark-spot segmentation, statistical feature extraction, and contextual information such as shipping lanes and offshore platforms [18]. Subsequent studies further emphasized that the key difficulty in SAR-based oil spill monitoring lies not only in detecting dark regions, but also in distinguishing true oil slicks from look-alikes such as low-wind areas, natural films, and other ocean surface phenomena with similar radar signatures [19,20,21]. These classical SAR-related studies provide the physical and methodological basis for understanding radar dark-spot detection and the long-standing look-alike problem, which also affects oil slick interpretation in marine radar imagery.

In recent years, artificial intelligence and learning-based techniques have further advanced SAR oil spill detection. Machine learning methods, such as Support Vector Machines (SVMs) and Random Forests, and deep learning models, such as Convolutional Neural Networks, U-Net, DeepLabv3+, and other encoder–decoder networks, have improved the ability to learn texture, spatial, and contextual features for oil spill segmentation [22,23,24]. More recent studies have further explored self-evolving deep learning algorithms, near-real-time detection and early-warning systems, DeepLabv3+-based segmentation frameworks, and lightweight attention networks for Sentinel-1 SAR oil spill detection [25,26,27,28]. These studies demonstrate that AI-assisted SAR monitoring remains an active research direction and has achieved substantial progress in segmentation accuracy, look-alike discrimination, and operational monitoring. However, their performance often depends on representative training samples and accurate annotations, and their generalization ability may be affected when imaging conditions, sea states, or sensor configurations differ from those included in the training data.

Despite these advances, SAR-based monitoring still has limitations for local and time-critical oil spill response. SAR is highly suitable for wide-area and all-weather surveillance, but its capability for continuous observation is constrained by satellite revisit intervals and data availability [29,30,31]. In practical emergency scenarios, especially for shipborne or coastal monitoring, high-temporal-resolution sensors are needed to continuously observe the evolution of oil slicks over a local sea area. X-band marine radar provides such a complementary sensing platform because it can acquire sea-surface backscatter at short time intervals and is widely used on ships and coastal stations. Therefore, X-band marine radar has practical potential for real-scene oil spill monitoring, particularly when rapid local response is required.

However, oil slick segmentation in X-band marine radar imagery remains challenging. Oil-covered areas usually appear as low-contrast dark targets embedded in heterogeneous sea clutter, and their radar signatures can be easily confused with wave shadows, speckle noise, and other clutter-induced dark regions [32,33,34,35,36,37]. Hand-crafted structural features, such as the Slick Boundary Ratio (SBR), provide physically interpretable descriptions of oil-induced wave damping and can help distinguish continuous slick regions from fragmented clutter. Nevertheless, classical SBR-based methods mainly rely on local dark-pixel ratios and fixed threshold settings, which may lead to false positives and unstable ROI extraction in complex sea clutter environments [38]. Metaheuristic algorithms, such as the Whale Optimization Algorithm (WOA), offer an adaptive way to search for segmentation thresholds [39]. However, the standard WOA may still be influenced by high-contrast wave clutter and global grayscale variance when explicit dark-target constraints are absent.

To address these problems, this study proposes a training-free two-stage detection framework that integrates a boundary-aware improved SBR feature with an improved WOA. The improved SBR feature jointly considers internal darkness and boundary continuity to extract a cleaner oil slick ROI, while the improved WOA introduces an exponential penalty-based threshold optimization strategy to guide the threshold search toward solutions consistent with the dark-target characteristics of oil slicks. By combining physically interpretable feature representation with constrained global threshold optimization, the proposed framework aims to suppress false positives and improve segmentation robustness in low-contrast X-band marine radar scenes. The present study is intended as a real-scene validation of methodological feasibility based on shipborne radar observations from an actual spill response task. The main contributions of this study are summarized as follows. First, a boundary-aware improved Slick Boundary Ratio (SBR) feature is proposed to jointly characterize the internal dark-pixel response and the boundary structure of candidate oil slick regions. By introducing the boundary-pixel term into the local SBR feature computation procedure, the proposed feature improves the discrimination between spatially coherent oil slick regions and fragmented dark clutter. Second, an improved Whale Optimization Algorithm (WOA)-based threshold optimization strategy is developed by incorporating a dark-target-oriented exponential penalty term into the fitness function. This modification guides the threshold search toward solutions that are more consistent with the low-backscatter characteristics of oil slicks and reduces over-segmentation caused by heterogeneous sea clutter. Third, the proposed training-free framework is validated using real shipborne X-band marine radar observations acquired during an actual oil spill response task. Comprehensive comparisons with thresholding-based, clustering-based, machine-learning-based, ensemble-learning-based, neural-network-based, SBR-Sauvola-based, and metaheuristic baseline methods are conducted under the same experimental setting to evaluate its segmentation robustness. The remainder of this paper is organized as follows. Section 2 introduces the shipborne radar dataset and details the proposed methodology, including data preprocessing, the boundary-aware improved SBR feature, and the improved Whale Optimization Algorithm. Section 3 presents the experimental results and visualizes the step-by-step segmentation process. Section 4 discusses the spatial error characteristics, validates the improved SBR feature and ROI-gating parameters, analyzes the contribution of the improved WOA strategy, compares the proposed method with competitive baseline methods, and discusses environmental limitations.

2. Materials and Methods

2.1. Materials

This study was based on shipborne X-band radar observations acquired during a real oil spill response task in the Bohai Sea near Dalian Bay. The monitored accident occurred near a coastal terminal and was associated with a crude oil release caused by pipeline failure during unloading operations. To support continuous observation of the affected sea surface, radar data were collected aboard the training vessel Yukun of Dalian Maritime University (Dalian, China), which served as the experimental platform in this study. As shown in Figure 1, the vessel was equipped with an X-band radar and the corresponding data acquisition platform. This figure provides contextual information on the real shipborne observation platform and indicates that the radar imagery used in this study was obtained from field observations rather than from a simulated or laboratory dataset.

The onboard radar system operated in horizontal polarization with a monitoring range of 0.75 nautical miles. The observations were recorded at approximately 2 s intervals, generating radar scans of 1024 × 1024 pixels. The original radar echoes were first stored in polar coordinates and then transformed into Cartesian images for subsequent full-image analysis. In the present study, one representative 1024 × 1024 radar scene containing visually identifiable oil slick signatures was selected from the continuous shipborne radar observations for methodological validation and quantitative evaluation. This scene was used together with a manually interpreted reference mask to assess the segmentation performance of the proposed method. A representative raw radar image in polar coordinates is shown in Figure 2, where the oil slick locations are marked for visual reference.

Owing to the emergency nature of the spill response, auxiliary environmental parameters such as wave height, wave period, and wind speed were not synchronously measured at the site during acquisition. Therefore, the dataset provides valid radar imagery and manually interpreted reference masks for pixel-level segmentation evaluation in the selected real spill response scene, but it does not support a stratified analysis of algorithm performance under different sea-state conditions. For quantitative assessment, the manually interpreted oil slick regions were used as the reference masks. Although visual interpretation may still involve uncertainty in diffuse boundary zones, it remains a practical reference for evaluating segmentation performance in real low-contrast radar scenes.

2.2. Overall Workflow of the Proposed Method

The proposed framework for robust oil slick segmentation in X-band marine radar images is illustrated in Figure 3. The overall procedure consists of four consecutive stages. First, the raw radar observations were transformed from polar coordinates into Cartesian images and then preprocessed to suppress co-frequency interference and speckle noise. Grayscale correction and local contrast enhancement were further applied to improve the visibility of low-contrast slick structures. Second, a sea-surface region of interest (ROI) mask was constructed to exclude non-sea areas and constrain subsequent analysis to the valid monitoring region. Third, within the ROI, a boundary-aware improved SBR feature was computed in sliding windows to characterize local dark-target structures, and an initial decision map was generated by Median Absolute Deviation (MAD)-based adaptive thresholding. Finally, an improved WOA was employed to determine the optimal global segmentation threshold under the ROI constraint, and the segmentation result was further refined by area filtering and morphological post-processing.

2.3. Data Preprocessing

The preprocessing significantly improved the image quality, reducing background interference and enhancing oil spill visibility, thus providing a solid foundation for accurate segmentation. The original radar image was converted from polar to Cartesian coordinates to facilitate subsequent processing. A row vector convolution with the kernel [−1, −1, 4, −1, −1] was applied to enhance co-frequency interference, which appeared as bright horizontal structures in the Cartesian image. This kernel was selected according to the observed local intensity characteristics of the co-frequency interference. Specifically, these interference responses usually appear as locally bright noise points or stripe-like structures, while their left and right neighboring pixels are relatively darker. The positive central coefficient enhances the local bright response, whereas the negative neighboring coefficients suppress the adjacent background intensity. Since the sum of the kernel coefficients is zero, this row kernel acts as a directional high-pass filter along the azimuth direction and emphasizes abrupt horizontal intensity variations. Therefore, it facilitates the subsequent threshold-based extraction and masking of co-frequency interference before further oil slick segmentation. Grayscale thresholding was then used to isolate and mask out these interference signals, followed by mean filtering to reduce residual noise. A dual-threshold method was applied to binarize the image again to identify and eliminate speckle noise. A median filter with a 20 × 20 window size was then applied to further reduce speckles and smooth the image. The denoised image was subjected to grayscale correction for normalizing intensity levels. Finally, local contrast enhancement was applied to highlight the oil slick boundaries and improve visual distinction from the background. As shown in Figure 4, the preprocessing significantly reduced background clutter and enhanced the oil slick boundaries, thereby improving detection clarity.

2.4. Boundary-Aware Improved SBR Model

In X-band marine radar imagery, classical SBR measures the proportion of dark pixels inside a sliding window and treats this as evidence of a slick. However, this approach is highly susceptible to speckle depressions and elongated wave streaks, which may cover large areas but lack the typical morphological characteristics of an oil slick. To solve these issues, a boundary-aware improved SBR model was introduced, as shown in Figure 5.

Compared with the classical SBR feature, the improvement in this study lies in the feature computation procedure rather than only in the post-processing stage. The classical SBR mainly evaluates the proportion of dark pixels within a sliding window, and therefore a window containing many scattered dark clutter pixels may still produce a high SBR response. In contrast, the proposed boundary-aware improved SBR first extracts dark candidate pixels using MAD-based adaptive binarization and then computes both the internal dark-pixel response and the boundary structure of the candidate dark region within each local window.

Specifically, for each sliding window, N denotes the number of dark candidate pixels, P denotes the number of boundary pixels of the candidate dark region, and L denotes the side length of the local window. The dark-pixel term describes the internal low-backscatter response, while the boundary term reflects the spatial continuity and boundary complexity of the candidate region. By incorporating the boundary-related term into the SBR formulation, the improved feature suppresses fragmented clutter responses with high boundary complexity and enhances spatially coherent oil slick regions. Therefore, the proposed SBR feature is improved from a dark-pixel-ratio descriptor into a boundary-aware local structural descriptor.

The computation method was as follows:

(1)

Image binarization

Calculate the adaptive threshold T for image I by using the median absolute deviation (MAD):

$$T=m-k\times {\displaystyle \frac{\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\left(\right|I-m\left|\right)}{h}}$$

(1)

where m denotes the median of the pixel intensities in I, h is a scaling factor, which was set to 0.6475 here, and k is a tuning coefficient, which was set to 3 here. Then, the image was binarized with the threshold T.

(2)

The improved SBR feature extraction

Dark candidate pixels are extracted using a sliding window from the binarized image. Within each local window, N denotes the number of dark candidate pixels, P denotes the number of boundary pixels of the candidate dark region, and L denotes the side length of the sliding window.

The improved SBR feature D is computed as:

$$D=\mathrm{l}\mathrm{o}\mathrm{g}({\displaystyle \frac{N}{L^2}}+ϵ)-\mathrm{l}\mathrm{o}\mathrm{g}({\displaystyle \frac{P}{4(L-1)}}+ϵ)$$

(2)

where N is the number of dark candidate pixels, P is the number of boundary pixels of the candidate dark region within the sliding window, L is the side length of the sliding window, and $ϵ$ is a small constant used to avoid division by zero or logarithmic singularity. In this formulation, N characterizes the internal dark-pixel response, whereas P introduces boundary-related structural information.

Therefore, the improved SBR feature is not simply a threshold adjustment, but a modification of the local feature computation procedure by introducing the boundary-pixel term P into the original dark-pixel-ratio-based descriptor. By jointly evaluating the internal darkness and boundary complexity of candidate regions, the improved SBR feature can suppress fragmented clutter responses and better preserve spatially coherent oil slick structures.

Finally, D_norm is linearly normalized:

$$D_{\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}}={\displaystyle \frac{D-min\left(D\right)}{max\left(D\right)-min\left(D\right)}}$$

(3)

Based on the normalized improved SBR feature map D_norm, a two-step ROI extraction strategy was further applied to isolate potential oil slick regions from the surrounding background. First, a coarse ROI was defined using a gating mechanism:

$$ROI(x,y)=\left\{\begin{array}{ll}1,& {\tau }_l \le { D}_{\mathrm{norm}}(x,y) \le {\tau }_u\\ 0,& \mathit{otherwise}\end{array}\right.$$

(4)

where D_norm(x,y) denotes the normalized boundary-aware improved SBR feature value at pixel location (x,y), and ${\tau }_l$ and ${\tau }_u$ denote the lower and upper range-gating thresholds, respectively. In this study, ${\tau }_l$ = 0.01 and ${\tau }_u$ = 0.90 were used as the default ROI-gating thresholds. These two parameters are not final segmentation thresholds; instead, they are loose range-gating thresholds used to remove non-informative responses before the subsequent MAD-based refinement and WOA-based threshold optimization.

Specifically, ${\tau }_l$ is used as a lower gate to suppress near-zero responses after normalization, which mainly correspond to obvious background areas, whereas ${\tau }_u$ controls the upper range of the candidate ROI in the normalized improved SBR feature map (Figure 6). Sensitivity analyses in Section 4.3 show that the final segmentation performance is insensitive to ${\tau }_l$ within the tested lower-threshold range and becomes stable when τu is sufficiently large. Therefore, ${\tau }_l$ = 0.01 and ${\tau }_u$ = 0.90 were adopted as conservative default ROI-gating parameters rather than finely tuned segmentation thresholds. To avoid confusion with the WOA parameters a and b, the lower and upper ROI-gating thresholds are denoted as ${\tau }_l$ and ${\tau }_u$, respectively.

2.5. The Improved Whale Optimization Algorithm for Oil Spill Segmentation

In the present study, the image segmentation problem is formulated as a one-dimensional global threshold optimization problem within the sea-surface ROI. Each whale represents one candidate solution, expressed as ${\overrightarrow{X}}_i$ = [T$i$], where T$i$ denotes the normalized grayscale threshold represented by the i-th whale. Since the input radar image is normalized to the interval [0, 1], the search space of T$\mathrm{i}$ is also constrained to [0, 1]. For each candidate threshold, pixels within the ROI are divided into oil-slick and background classes, and the corresponding fitness value is evaluated using the modified Otsu criterion with a dark-target-oriented penalty term. Therefore, WOA is used here as a population-based global optimizer for threshold selection, rather than as an image classifier.

WOA was selected because it has a compact population-update mechanism, requires relatively few control parameters, and provides a balance between global exploration and local exploitation through encircling behavior, spiral updating, and random search. These characteristics make it suitable for the low-dimensional continuous threshold optimization task considered in this study. Compared with other bio-inspired algorithms such as Particle Swarm Optimization (PSO), the Firefly Algorithm, and Artificial Bee Colony (ABC), WOA does not require velocity updating, attractiveness modeling, or multi-role bee operations, making it convenient to integrate into a training-free segmentation framework. It should be noted that this selection does not imply that WOA is universally superior to other metaheuristic algorithms; rather, WOA was chosen because its simple structure is suitable for the threshold optimization problem addressed in this work.

Nevertheless, the standard WOA also has limitations. Its convergence behavior may depend on the initial population distribution, and premature convergence may occur when the fitness landscape is noisy. In X-band marine radar imagery, strong sea clutter, wave-shadow responses, and heterogeneous low-backscatter regions may distort the standard between-class variance criterion and cause the standard WOA to select thresholds that lead to over-segmentation. Therefore, this study introduces an exponential penalty term into the fitness function to guide the search toward thresholds that are more consistent with the low-backscatter dark-target characteristics of oil slicks. This modification aims to retain the simplicity and global search capability of WOA while reducing its sensitivity to clutter-dominated variance.

The workflow of the WOA used for oil spill segmentation is summarized in Figure 7.

(1)

Initialization

The positions of N whales are randomly initialized in the normalized threshold search space, and the initial parameters were set as follows: a is a linear decreasing coefficient, which decreases linearly from 2 to 0 in order to balance the global and local search. The coefficient vector $\overrightarrow{A}$ is defined as $\overrightarrow{A}$ = 2a·r₁ − a where r₁ ∈ [0, 1] is a random number. The coefficient vector $\overrightarrow{C}$ is given by $\overrightarrow{C}$ = 2·r₂, where r₂ ∈ [0, 1] is another random number. The spiral shape parameter l lies within the range [−1, 1] and is also a random number.

An additional parameter a₂ was introduced to control the spiral shape.

$$a_2=-1+t\cdot (-{\displaystyle \frac{1}{t_{max}}})$$

(5)

where t is the iteration index and t_max is the maximum number of iterations. The spiral constant b is typically set to 1 and controls the tightness of the logarithmic spiral, while the probability threshold p is typically set to 0.5 and is used to switch between different search strategies.

(2)

Perform fitness evaluation

First, the fitness value of each whale is computed, with the current best solution identified as the one yielding the highest fitness. Then, the position of each whale is updated:

For each whale ${\overrightarrow{X}}_i\left(t\right)$, choose the search strategy based on ∣$\overrightarrow{A}$∣ and p; there are three cases:

$$\begin{array}{r}Behavior=\left\{\begin{array}{ll}Encirclingbehavior& if \left|\overrightarrow{A}\right|<1 \mathrm{a}\mathrm{n}\mathrm{d} r<p\\ Spiralupdate& if \left|\overrightarrow{A}\right|<1 \mathrm{a}\mathrm{n}\mathrm{d} r\ge p\\ Randomsearch& if \left|\overrightarrow{A}\right|\ge 1\end{array}\right.\end{array}$$

(6)

where r₁∼U(0, 1).

In the encircling phase, each whale updates its position toward the current global best, moving along its direction and distance to progressively approach the optimum:

$${\overrightarrow{X}}_i(t+1)={\overrightarrow{X}}^{\ast }\left(t\right)-\overrightarrow{A}\cdot \overrightarrow{D}$$

(7)

where $\overrightarrow{D}$ represents the current distance between the whale (agent) and the global best.

$$\overrightarrow{D}=|\overrightarrow{C}\cdot {\overrightarrow{X}}^{\ast }(t)-{\overrightarrow{X}}_i(t\left)\right|$$

(8)

where $\overrightarrow{C}$ denotes the coefficient vector. When simulating the humpback whale’s spiral update, the agent adjusts its position along a logarithmic-spiral around the current global best, enabling gradual approach and fine-grained exploration; the position is updated as:

$${\overrightarrow{X}}_i(t+1)={\overrightarrow{D}}^{\prime }\cdot e^{b\cdot l}\cdot \mathrm{c}\mathrm{o}\mathrm{s}\left(2\pi l\right)+{\overrightarrow{X}}^{\ast }\left(t\right)$$

(9)

Current distance between the whale (agent) and the global best:

$${\overrightarrow{D}}^{\prime }=\left|{\overrightarrow{X}}^{\ast }\right(t)-{\overrightarrow{X}}_i(t\left)\right|$$

(10)

(3)

The procedure terminates when either the maximum number of iterations t_max is reached or a convergence condition is satisfied. In the initial implementation of the improved WOA, t_max = 200 was used as a conservative upper bound to ensure sufficient iterations for convergence during algorithm design. However, the maximum iteration number is not a fixed theoretical parameter of WOA and is commonly adjusted according to the specific optimization task and convergence behavior. In this study, the convergence condition was defined as the fitness value remaining unchanged for 10 consecutive iterations, and the best solution X* was then returned.

(4)

Fitness Function

This study extended the Otsu criterion by introducing an exponential penalty that gently favored lower thresholds in dark-target scenes. The penalty strength was fixed, ensuring that separability remained the primary driver while the search was nudged toward darker cutoffs within the sea-surface ROI:

$$fitness={\displaystyle \frac{{\sigma }^2}{\mathrm{e}\mathrm{x}\mathrm{p}(s\cdot threshold)}}$$

(11)

where threshold denotes a threshold condition or the degree of constraint satisfaction; exp is the exponential function, with exp(x) = ex, s is a tuning coefficient (set to 3 in this work); and σ² denotes the maximum between-class variance:

$${\sigma }^2={\displaystyle {\sum }_{i=1}^K} {\omega }_i\cdot ({\mu }_i - {\mu }_T{)}^2$$

(12)

where K is the total number of classes, μ_T is the global mean, and ω_i denotes the proportion of samples in the i-th class within the dataset.

3. Results

3.1. ROI Extraction and Refinement

For the representative 1024 × 1024 radar scene used in this study, the boundary-aware improved SBR feature was first computed and normalized, as shown in Figure 6. Based on this normalized feature map, the binary ROI mask shown in Figure 8 was obtained using the range-gating strategy described in Section 2.4. Therefore, Figure 8 represents the ROI selection result derived from the improved SBR feature, rather than the raw SBR feature map itself.

By comparing Figure 8 with the original radar scene and the ROI visualization results in Figure 9, it can be observed that a large portion of irrelevant background responses outside the candidate slick region is suppressed, while the main candidate oil slick region and its boundary structure are retained. This provides a cleaner search region for the subsequent WOA-based threshold optimization.

To further visualize the ROI extraction effect, the sea-surface ROI mask was applied to the preprocessed image through element-wise multiplication, namely I_ROI = I_final⊙ROI. In this way, the image intensity within the candidate sea-surface region was preserved, whereas the excluded regions were suppressed, as shown in Figure 9a. In addition, the inverted ROI mask was superimposed on the original image to verify the spatial coverage of the extracted candidate region, as shown in Figure 9b. These results indicate that the improved SBR-based ROI extraction stage can reduce irrelevant background interference while preserving the main candidate slick structures for subsequent segmentation.

3.2. WOA-Based Global Threshold Optimization and Post-Processing

The improved WOA was employed to determine the optimal global threshold for precise segmentation. The population size was set to 50, and the maximum number of iterations was set to 50. In the initial implementation, t_max = 200 was used as a conservative upper bound for the improved WOA. However, in the present radar scene, the fitness value became stable within 50 iterations, and no further improvement in the segmentation threshold was observed in subsequent tests. Therefore, the maximum number of iterations was set to 50 in this study to reduce redundant computation while maintaining stable threshold optimization. First, the initial binary segmentation map generated by the WOA-optimized threshold is shown in Figure 10a. Subsequently, morphological post-processing was applied to remove small isolated noise spots, resulting in the cleaner mask shown in Figure 10b. The refined segmentation result was then transformed from the radar polar domain into the Cartesian coordinate system for spatial representation, as shown in Figure 10c. The final detection result, superimposed on the original radar image to verify alignment, is presented in Figure 10d.

3.3. Quantitative Evaluation Metrics

To quantitatively assess the segmentation performance of the proposed method, a series of standard statistical metrics were adopted based on the pixel-wise comparison between the predicted results and the manually interpreted Ground Truth.

Let True Positives (TP) denote correctly detected oil spill pixels, False Positives (FP) denote background pixels incorrectly classified as oil spill, False Negatives (FN) denote oil spill pixels missed by the algorithm, and True Negatives (TN) denote correctly rejected background pixels. Based on these quantities, Accuracy, Precision, Recall, F1-score, and Intersection over Union (IoU) were calculated.

Accuracy measures the proportion of correctly classified pixels over the entire image:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(13)

Precision measures the proportion of predicted oil spill pixels that are truly oil spill pixels, reflecting the algorithm’s ability to suppress false positives:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(14)

Recall reflects the detection sensitivity and quantifies the proportion of actual oil spill pixels that are successfully identified:

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(15)

F1-score is the harmonic mean of Precision and Recall, providing a balanced assessment of segmentation performance:

$$F1={\displaystyle \frac{2TP}{2TP+FP+FN}}$$

(16)

IoU evaluates the spatial overlap consistency between the predicted oil spill region and the Ground Truth:

$$IoU={\displaystyle \frac{TP}{TP+FP+FN}}$$

(17)

In practical marine oil spill monitoring scenarios, both missed detections and excessive false alarms should be carefully controlled. Therefore, Precision-, Recall-, and overlap-based metrics (F1-score and IoU) were emphasized to reflect the algorithm’s capability to balance detection sensitivity, false-positive suppression, and spatial consistency in complex sea clutter environments.

These metrics collectively provide a comprehensive evaluation framework for assessing segmentation robustness, spatial consistency, and operational applicability under low-contrast X-band radar conditions.

4. Discussion

4.1. Performance Validation and Spatial Error Analysis

The results of visual interpretation are shown in Figure 11a, which serve as the reference standard (Ground Truth) for performance validation. Although visual interpretation relies on expert knowledge to distinguish oil spills from complex sea clutter, this manual process remains constrained by operational conditions and the intrinsic ambiguity of radar backscatter. The diffuse and low-contrast boundaries of oil slicks require sustained attention for accurate delineation, and human fatigue as well as inter-observer variability may lead to conservative or inconsistent annotations.

Based on the quantitative metrics defined in Section 3.3, a spatial error analysis was further conducted to investigate the distribution characteristics of segmentation discrepancies. Figure 11b,c present the spatial comparison between the predicted results and the reference annotation, highlighting True Positives (green), False Positives (red), and False Negatives (blue).

The spatial distribution of the segmentation discrepancies was further examined based on the overlay results in Figure 11. It should be noted that the following interpretation is qualitative and is based on the observed spatial relationship among the predicted mask, the manually interpreted reference mask, and the original radar backscatter pattern.

As shown in Figure 11b,c, a substantial proportion of the FPs appears near the peripheral transition zones of the main oil slick region. This suggests that part of the discrepancy may be related to the diffuse and low-contrast nature of oil slick boundaries. In manual interpretation, experts tend to annotate high-confidence core regions, whereas the proposed boundary-aware feature may retain some weak dark responses around the slick margins. Therefore, these peripheral detections are counted as FPs in the pixel-wise evaluation, although they are spatially adjacent to the main slick region.

Additional scattered FPs are observed in heterogeneous dark-clutter regions outside the main slick body. These regions may be associated with wave shadows, speckle depressions, or other low-backscatter clutter structures that exhibit visual similarity to oil-induced dark signatures in X-band radar imagery. Since synchronous wave and wind measurements were not available in this dataset, these interpretations should be regarded as plausible explanations rather than direct environmental attribution.

Conversely, FNs are mainly located in discontinuous or weak-response parts of the manually interpreted slick region. This indicates that fragmented slick structures and low local contrast remain challenging for both manual annotation and automated segmentation. Therefore, the spatial error analysis mainly provides a qualitative assessment of where segmentation discrepancies occur, while the quantitative evaluation in Section 3.3 provides the numerical basis for performance validation.

Overall, the spatial analysis suggests that the observed discrepancies are mainly concentrated around diffuse slick boundaries, heterogeneous dark-clutter regions, and weak-response fragmented areas. These observations indicate that the remaining errors are closely related to the intrinsic ambiguity of low-contrast X-band radar imagery and the uncertainty of manual annotation. Therefore, the spatial error analysis should be interpreted as a qualitative complement to the quantitative metrics, rather than as independent proof of environmental causality.

4.2. Validation of the Improved SBR Feature

To clarify the individual contributions of the two key components, the proposed framework was analyzed separately in terms of feature representation and threshold optimization. This subsection focuses on the improved SBR feature to examine its physical rationale and evaluate its effectiveness compared with the classical SBR model, as shown in Figure 12. While the traditional SBR model is derived from the hydrodynamic damping characteristics of oil slicks, it relies mainly on the ratio of dark pixels within local windows for discrimination. As a result, wave shadows and speckle interference with weak spatial continuity may be incorrectly identified as oil slick targets. As shown in the red boxed regions of Figure 12, the classical SBR model still retains scattered noise responses and fragmented false positives in complex sea clutter scenes. In contrast, the improved SBR feature in Equation (2) introduces joint evaluation of internal darkness and boundary continuity during feature extraction. This formulation helps suppress the response of fragmented clutter regions with high perimeter-to-area ratios, thereby improving discrimination between cohesive oil slick regions and irregular interference. As illustrated in Figure 9b, many previously retained artifacts are effectively reduced, and the extracted ROI becomes cleaner while preserving the main morphology and boundary structure of the oil slick. These results indicate that the improved feature formulation reduces the shape sensitivity of the traditional SBR model and improves ROI extraction quality without requiring additional morphological post-processing.

4.3. Parameter Sensitivity of the ROI-Gating Thresholds

To examine whether the ROI-gating thresholds strongly affected the final segmentation performance, sensitivity analyses were conducted for both the lower threshold ${\tau }_l$ and the upper threshold τu. During the τu test, ${\tau }_l$ was fixed at 0.01 and τu was varied from 0.10 to 0.90 with an interval of 0.10. The reviewer-suggested value τu = 0.99 was also tested. During the ${\tau }_l$ test, ${\tau }_u$ was fixed at 0.90 and ${\tau }_l$ was varied among 0, 0.005, 0.01, 0.02, and 0.05. For all tests, the subsequent WOA-based threshold optimization, area filtering, and morphological post-processing steps were kept unchanged.

As shown in

Table 1. Sensitivity analysis of the ROI-gating thresholds.

${\mathit{\tau }}_{\mathit{l}}$	${\mathit{\tau }}_{\mathit{u}}$	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	IoU (%)
0.01	0.1	97.83	0	0	0	0
0.01	0.2	97.85	74.80	1.61	3.14	1.60
0.01	0.3	98.01	74.69	12.70	21.71	12.18
0.01	0.4	98.63	80.07	48.77	60.62	43.49
0.01	0.5	99.14	83.96	74.84	79.14	65.48
0.01	0.6	99.34	85.20	83.95	84.57	73.27
0.01	0.7	99.34	85.20	83.95	84.57	73.27
0.01	0.8	99.34	85.20	83.95	84.57	73.27
0.01	0.9	99.34	85.20	83.95	84.57	73.27
0.01	0.99	99.34	85.20	83.95	84.57	73.27
0	0.9	99.34	85.20	83.95	84.57	73.27
0.005	0.9	99.34	85.20	83.95	84.57	73.27
0.02	0.9	99.34	85.20	83.95	84.57	73.27
0.05	0.9	99.34	85.32	83.91	84.61	73.32

, when ${\tau }_u$ was too small, the candidate ROI became overly restrictive, and many true oil slick pixels were excluded before WOA-based segmentation. This resulted in low Recall, F1-score, and IoU values. As ${\tau }_u$ increased, the performance improved progressively. When ${\tau }_u$ reached 0.60, the final segmentation performance became stable, and the same results were obtained for τ_u = 0.60, 0.70, 0.80, 0.90, and 0.99.

The lower threshold ${\tau }_l$ showed even lower sensitivity. The results were identical for ${\tau }_l$ = 0, 0.005, 0.01, and 0.02. When ${\tau }_l$ was increased to 0.05, only a very small change was observed: Precision and IoU slightly increased, while Recall slightly decreased. This indicates that ${\tau }_l$ mainly suppresses near-zero background responses and does not critically determine the final segmentation result. Therefore, ${\tau }_l$ = 0.01 was selected as a conservative lower gate to preserve weak slick-related responses, while ${\tau }_u$ = 0.90 was selected as a conservative upper gate within the stable performance range. These results confirm that the reported segmentation performance is not artificially determined by exact choices of ${\tau }_l$ and ${\tau }_u$.

4.4. Validation of the Improved WOA Strategy

The purpose of this comparison is not to claim that WOA is universally superior to other metaheuristic algorithms, but to evaluate whether the dark-target-oriented modification introduced in this study can overcome the over-segmentation tendency of the standard WOA in the evaluated X-band marine radar scene. This subsection evaluates the contribution of the improved WOA strategy through quantitative and qualitative comparison with the traditional WOA method, as summarized in

Table 2. Quantitative comparison between the improved WOA and traditional WOA.

	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	IoU (%)
Improved WOA	99.36	85.73	84.42	85.07	74.01
Traditional WOA	82.19	10.85	99.98	19.58	10.85

. In marine radar images, oil slicks usually appear as low-contrast dark targets. Without explicit dark-target constraints, the traditional WOA-based threshold optimization tends to be strongly influenced by high-contrast wave clutter and the global variance criterion, which may lead to severe over-segmentation and numerous false positives in visually complex scenes, as shown in Figure 13. Quantitatively, although the traditional method achieves an extremely high Recall (99.98%), its very low Precision (10.85%), F1-score (19.58%), and IoU (10.85%) indicate limited ability to distinguish weak oil slick backscatter from heterogeneous background interference, while its Accuracy remains relatively low at 82.19%. In contrast, the improved WOA strategy introduces an exponential penalty term to guide the threshold search toward solutions more consistent with the dark-target characteristics of oil slicks, thereby suppressing the influence of high-reflectivity wave clutter. As reported in Table 2, the improved strategy increases Accuracy to 99.36%, Precision to 85.73%, F1-score to 85.07%, and IoU to 74.01%, while maintaining a Recall of 84.42%. These results indicate that the improved WOA strategy provides substantially more balanced and reliable threshold selection for oil slick segmentation in low-contrast marine radar scenes.

4.5. Overall Comparison with Competitive Baseline Methods

To further evaluate the competitiveness of the proposed framework, additional baseline methods were implemented and applied to the same X-band marine radar image. The compared methods include classical global thresholding (Otsu), clustering-based segmentation (K-Means), machine-learning-based classification (SVM), ensemble learning (Random Forest), neural-network-based classification using a Back-Propagation Neural Network (BP Neural Network), a closely related SBR-based adaptive thresholding method (SBR + Sauvola), and metaheuristic threshold optimization (GWO). The selection of these baseline categories was guided by the literature reviewed in the Introduction, including machine-learning- and deep-learning-based oil spill detection studies [22,23,24,25,26,27,28], SBR-related radar dark-target analysis [38], and metaheuristic optimization-based threshold searching [39]. All methods were evaluated using the same manually interpreted reference mask and the same pixel-level metrics, including Accuracy, Precision, Recall, F1-score, and IoU. This setting ensures that the comparison reflects algorithmic differences under the same radar scene and annotation standard.

It should be noted that deep semantic segmentation models such as U-Net, DeepLabv3+, and attention-based encoder–decoder networks usually require a large number of annotated training samples and independent testing scenes. Since the present study is based on one representative real spill-response radar scene, training such models on the same image would lead to an unfair and potentially overfitted comparison. Therefore, deep-learning methods are discussed as important state-of-the-art references, while the quantitative comparison focuses on methods that can be reasonably applied or reproduced under the same single-scene X-band marine radar setting.

In Figure 14, colored boxes are used to mark representative local regions for visual comparison. These boxes are not quantitative error-category labels and do not correspond to fixed meanings such as TP, FP, or FN. Instead, they highlight typical local segmentation phenomena, including false-positive clutter responses, over-segmentation, fragmented detections, and missed weak slick structures. Different colors are used only to distinguish the annotated regions in different subfigures and to facilitate visual comparison among the baseline methods and the proposed method.

As summarized in

Table 3. Quantitative comparison of segmentation performance across different methods.

	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	IoU (%)
Our Method	99.36	85.73	84.42	85.07	74.01
SVM	96.94	41.36	98.43	58.24	41.08
Otsu Method	81.76	10.06	100.00	18.27	10.06
K-Means (K = 2)	82.41	10.39	99.96	18.82	10.38
GWO	82.19	10.86	99.98	19.59	10.86
SBR + Sauvola	99.30	80.99	67.50	73.63	58.27
BP Neural Network	96.63	38.81	96.24	55.32	38.23
Random Forest	97.62	47.64	98.14	64.14	47.21

, the traditional unsupervised methods, including Otsu and K-Means, achieved extremely high Recall values of 100.00% and 99.96%, respectively. However, their Precision, F1-score, and IoU values were very low. Otsu achieved only 10.06% Precision, 18.27% F1-score, and 10.06% IoU, while K-Means achieved 10.39% Precision, 18.82% F1-score, and 10.38% IoU. This indicates severe over-segmentation and poor discrimination between oil slick targets and heterogeneous sea clutter. The visual results in Figure 14b,c further support this observation. In the boxed regions, a large amount of background clutter is incorrectly segmented as oil slicks, suggesting that thresholding or clustering based mainly on grayscale distribution is insufficient for complex low-contrast marine radar scenes.

A similar tendency can be observed for the GWO-based thresholding method. Although GWO achieved a Recall of 99.98%, its Precision, F1-score, and IoU were only 10.86%, 19.59%, and 10.86%, respectively. As shown in Figure 14d, the boxed regions still contain extensive false-positive responses in clutter-dominated areas. This suggests that metaheuristic optimization alone cannot ensure accurate oil slick segmentation if the objective function lacks explicit dark-target or spatial constraints. In other words, using an optimization algorithm without physically meaningful constraints may still lead to thresholds dominated by heterogeneous sea clutter.

The learning-based baselines, including SVM, BP Neural Network, and Random Forest, generally achieved high Recall values, indicating that they could detect a large proportion of oil slick pixels. Specifically, SVM, BP Neural Network, and Random Forest achieved Recall values of 98.43%, 96.24%, and 98.14%, respectively. However, their Precision, F1-score, and IoU values remained substantially lower than those of the proposed method. SVM achieved 41.36% Precision, 58.24% F1-score, and 41.08% IoU, while BP Neural Network achieved 38.81% Precision, 55.32% F1-score, and 38.23% IoU. Random Forest performed better than SVM and BP Neural Network, with 47.64% Precision, 64.14% F1-score, and 47.21% IoU, but it still remained clearly below the proposed method. The boxed regions in Figure 14a,f,g show that these learning-based methods still produce fragmented false-positive responses in low-contrast clutter regions. This indicates that, under the limited single-scene training condition, learning-based classifiers are still affected by the representativeness of sampled training pixels and handcrafted features.

The SBR + Sauvola method represents a closely related training-free SBR-based adaptive thresholding strategy. Compared with Otsu, K-Means, and GWO, it achieved much higher Accuracy, Precision, F1-score, and IoU values, reaching 99.30%, 80.99%, 73.63%, and 58.27%, respectively. This demonstrates the usefulness of SBR-based feature representation and local adaptive thresholding for marine radar oil slick detection. However, its Recall was only 67.50%, which was lower than that of the proposed method. As shown in Figure 14e, SBR + Sauvola suppresses many false-positive responses, but the detected oil slick structures are more sparse and discontinuous. This suggests that local adaptive thresholding alone may miss weak or fragmented slick responses, especially in diffuse boundary regions and low-contrast areas.

In contrast, the proposed method achieved the most balanced overall performance among all compared methods, with an Accuracy of 99.36%, a Precision of 85.73%, a Recall of 84.42%, an F1-score of 85.07%, and an IoU of 74.01%. Compared with the traditional unsupervised and metaheuristic baselines, the proposed method substantially improved Precision and IoU, indicating effective suppression of false positives. Compared with the learning-based baselines, it achieved better F1-score and spatial overlap consistency, showing stronger robustness under the single-scene X-band marine radar setting. Compared with SBR + Sauvola, the proposed method improved Recall, F1-score, and IoU, indicating better preservation of weak oil slick structures while maintaining low false-positive responses.

Overall, the quantitative results in Table 3 and the visually annotated regions in Figure 14 consistently demonstrate the advantage of the proposed framework. The boundary-aware improved SBR feature provides a physically interpretable candidate-region representation, while the dark-target-constrained WOA strategy further improves threshold selection under heterogeneous sea clutter. By combining these two components, the proposed method achieves a better balance between detection sensitivity, false-positive suppression, and spatial consistency. Therefore, the proposed training-free framework provides a more reliable solution for oil slick detection in the evaluated low-contrast X-band marine radar scene.

4.6. Environmental Limitations and System Deployment Constraints

The practical application of X-band marine radar for oil slick monitoring is affected by both environmental conditions and system-level constraints. Because the radar response depends strongly on the damping effect of oil slicks on capillary waves, the performance of the proposed method may vary under extreme sea states. Under very calm sea conditions, the lack of sufficiently developed capillary waves can reduce the contrast between oil-covered and oil-free areas, thereby weakening detection sensitivity. In contrast, under rough sea conditions and severe weather, the oil slick may become fragmented, while intensified sea clutter and atmospheric attenuation can further obscure the target structure and interfere with reliable segmentation.

In addition to these environmental factors, practical deployment on coastal or shipborne platforms also requires appropriate engineering support. For example, the radar system should maintain sufficient azimuth resolution to preserve target geometry, while the processing framework should support efficient handling of continuous observation data. The absence of synchronous environmental measurements further limits the interpretation of model generalizability. Without wave height, wave period, and wind speed records, the relationship between segmentation performance and specific sea-state conditions cannot be quantitatively established. Therefore, the reported results should be interpreted as evidence of practical feasibility in the evaluated real spill response scene, while broader applicability under different wind-wave conditions requires further validation with multi-scene datasets and matched environmental observations. Future research will focus on improving robustness under diverse sea states and exploring multi-sensor fusion strategies to enhance monitoring reliability.

5. Conclusions

This paper proposes a training-free oil slick detection framework based on a boundary-aware improved Slick Boundary Ratio (SBR) feature and an improved Whale Optimization Algorithm (WOA) to address the challenges of low contrast, heterogeneous sea clutter, and background interference in X-band marine radar imagery. The improved SBR feature incorporates both internal dark-pixel responses and boundary structural information, thereby improving the extraction of candidate oil slick regions and reducing fragmented clutter responses. The improved WOA further introduces an exponential penalty-based threshold optimization mechanism to guide the global threshold search toward the low-backscatter characteristics of oil slicks. By integrating these two stages, the proposed framework achieves accurate and spatially coherent segmentation results in the evaluated real radar scene, with an Accuracy of 99.36%, a Precision of 85.73%, a Recall of 84.42%, an F1-score of 85.07%, and an IoU of 74.01%.

Despite these promising results, several limitations should be acknowledged. First, the present validation is based on one representative shipborne X-band marine radar scene acquired during an actual oil spill response task. Although this provides a realistic field-observation scenario, the generalizability of the method across different sea states, oil types, wind-wave conditions, and radar acquisition settings still requires further verification. Second, synchronous environmental measurements, such as wind speed, wave height, and wave period, were not available during the emergency data acquisition. Therefore, the relationship between segmentation performance and specific environmental conditions could not be quantitatively analyzed. Third, the manually interpreted reference mask may contain uncertainty in diffuse slick-boundary regions, which can affect pixel-level evaluation, especially for weak or fragmented oil slick structures. Finally, although the proposed framework is training-free and the ROI-gating thresholds were shown to be relatively stable in the sensitivity analysis, some parameters and post-processing settings are still empirically configured and may need further adaptation for more complex operational scenarios.

Future work will focus on several possible improvements. More shipborne and coastal X-band marine radar datasets should be collected under diverse sea states and environmental conditions to evaluate the robustness and generalizability of the proposed method. Matched wind-wave observations should also be incorporated to analyze the influence of environmental factors on radar-based oil slick segmentation. In addition, more self-adaptive parameter selection and post-processing strategies should be developed to reduce empirical dependence and improve robustness for weak-target detection. Finally, multi-sensor fusion with complementary observations, such as SAR, optical imagery, or infrared data, may further improve oil slick discrimination and reduce false alarms in complex marine environments. These improvements will help extend the proposed framework from single-scene methodological validation toward more reliable operational oil spill monitoring.