Oil Film Detection for Marine Radar Image Using SBR Feature and Adaptive Threshold

Abstract

Marine oil spills pose a serious and persistent threat to marine ecosystems, coastal resources, and global environmental health. These incidents not only disrupt ecological balance by damaging marine flora and fauna but also lead to long-term economic consequences for fisheries, tourism, and maritime industries. Owing to their rapid spread and often unpredictable occurrence, timely and accurate detection is essential for effective containment and mitigation. An efficient detection system can significantly enhance the responsiveness of emergency teams, enabling targeted interventions that minimize ecological damage and economic loss. This paper proposes a marine radar-based oil spill detection method that combines the Significance-to-Boundary Ratio (SBR) feature with an improved Sauvola adaptive thresholding algorithm. The raw radar data was firstly preprocessed through mean and median filtering, grayscale correction, and contrast enhancement. SBR features were then employed to extract coarse oil spill regions, which were further refined using an improved Sauvola thresholding algorithm followed by a denoising step to obtain fine-grained segmentation. Comparative experiments using different threshold values demonstrate that the proposed method achieves superior segmentation performance by better preserving oil spill boundaries and reducing background noise. Overall, the approach provides a robust and efficient solution for marine oil spill detection and monitoring.

1. Introduction

With the continuous expansion of offshore oil exploration and maritime transportation, the frequency and scale of marine oil spill incidents have increased significantly [1,2,3,4,5]. During offshore oil exploration, production, and transportation, unexpected spills may occur due to overpressure, mechanical failures, pipeline corrosion, or ship collisions [6,7,8]. Such incidents pose serious threats to marine ecosystems, coastal economies, and public health. Given the sudden onset, rapid spread, and long-lasting impact of oil spills, timely and accurate detection is essential for effective emergency response and environmental damage mitigation [9,10,11].

Currently, mainstream marine oil spill monitoring technologies include SAR satellites, optical satellites, and UAVs. SAR enables all-weather, day-and-night monitoring with wide coverage, but suffers from low revisit frequency and delayed data processing [12,13]. Optical satellites offer high-resolution images, yet are limited by cloud cover and lighting conditions [14,15,16]. UAVs provide flexible, high-precision observation, but their short endurance and poor performance in bad weather limit their application [17,18,19].

In recent years, marine radar, particularly X-band shipborne radar, has gained attention as a promising tool for oil spill detection, thanks to its all-weather operability, low cost, platform adaptability, and continuous coverage capabilities [20,21,22]. Widely deployed on shore-based, shipborne, and offshore platforms, X-band radar offers a maximum detection range of up to 72 nautical miles and supports early warning and continuous observation of target sea areas [23,24,25,26,27,28]. However, research on X-band radar-based oil spill detection started relatively late. In 1991, Atanassov et al. first demonstrated that the suppression effect of oil slicks on ocean wave echoes could be leveraged for marine oil spill detection using X-band radar [29]. Subsequently, Zhu et al. proposed a power attenuation correction method to address uneven echo distribution and validated its effectiveness in oil film extraction [30].

To improve the accuracy and robustness of radar image-based oil spill detection, researchers have recently explored the use of image processing and deep learning techniques. Convolutional Neural Networks (CNNs), U-Net, and Transformer-based architectures have achieved promising results in remote sensing image segmentation tasks, particularly for high-contrast data [31,32,33,34,35]. However, these models typically rely on large-scale annotated datasets and exhibit limited generalization capabilities in low-contrast radar images under complex sea conditions, hindering their real-time applicability and engineering deployment [36].

To address these challenges, this paper proposes a marine radar oil spill detection method that combines the Significance-to-Boundary Ratio (SBR) feature with an improved Sauvola adaptive thresholding algorithm. The SBR feature captures local structural complexity to highlight potential oil spill regions, while the improved thresholding algorithm adapts to low-contrast imagery to enhance segmentation accuracy. Comparative experiments using different threshold settings demonstrate that the proposed method achieves superior performance in boundary preservation and background noise suppression, offering a robust and efficient solution for practical marine oil spill monitoring.

2. Materials and Methods

2.1. Materials

The experimental data was collected using a platform equipped with an integrated X-band marine radar transceiver and an onboard image processing system. The radar system includes a real-time display and supports continuous acquisition and storage of sea clutter echo data. The X-band radar features a short wavelength that enables high-resolution imaging with fine-grained echo responses. It is also capable of long-range target detection and operates reliably under adverse weather conditions such as rain and snow, making it well-suited for oil spill monitoring. During data acquisition, a horizontally polarized rotating antenna was used to scan the sea surface, generating 28 to 45 images per minute. The detection range was adjusted by changing the transmitted pulse width, and for this experiment, it was set to 0.75 nautical miles to improve image detail. At this range, the radar produces images with a resolution of 1024 × 1024 pixels, clearly capturing fine-scale oil slick features. The data used in this study were collected from a real incident near a coastal terminal in Dalian Bay caused by a pipeline failure during tanker unloading. All data were obtained by the training vessel Yukun of Dalian Maritime University during routine patrol missions, specifically at the site of the major oil spill accident that occurred in Dalian on 16 July. The radar platform and a sample image are shown in Figure 1 and Figure 2, respectively.

Due to the emergency nature of data collection following a sudden oil spill incident, specific measurements such as wave height, wave period, and wind speed at the spill site were not recorded.

2.2. The Methodology Flow

The experimental workflow consists of four main steps: data preprocessing, SBR feature extraction and classification, improved adaptive threshold segmentation, and noise reduction. As shown in Figure 3, the raw marine radar images were first preprocessed through coordinate transformation, filtering, grayscale correction, and contrast enhancement to improve image quality and target visibility. Then, the Significance-to-Boundary Ratio (SBR) feature was extracted and used to classify candidate oil spill regions. Afterwards, an improved Sauvola adaptive thresholding algorithm was subsequently applied to segment potential oil slick regions. Finally, post-processing was conducted to remove speckle noise and refine the segmentation, producing clean and accurate detection results.

2.3. Data Preprocessing

The complete data preprocessing workflow is shown in Figure 4. The detailed steps are described as follows:

(1) The original radar image in polar coordinates was converted to Cartesian coordinates to facilitate subsequent processing, as shown in Figure 4.

(2) A row vector convolution was applied using the kernel [−1, −1, 4, −1, −1] [−1, −1, 4, −1, −1], which enhances co-frequency interference lines appearing as bright horizontal structures in the Cartesian image.

(3) A grayscale thresholding operation was used to extract the co-frequency interference signals. This was followed by mean filtering to smooth out residual noise.

(4) The result was binarized again using a second grayscale threshold to improve contrast and suppress weak background clutter.

(5) Speckle noise was removed using an area-based filtering method that eliminates small connected regions below a pixel quantity threshold.

(6) A median filter with a 20 × 20 window size was then applied to further eliminate isolated noise artifacts and improve smoothness, as shown in Figure 4.

(7) The denoised image was then subjected to grayscale correction to normalize intensity levels.

(8) Finally, local contrast enhancement was applied to highlight the oil slick boundaries and improve visual distinction from the background, as illustrated in Figure 4.

The preprocessing steps performed before the detection procedure are shown in Figure 5. The oil slicks observed in this experiment were primarily crude oil spills from tankers or storage facilities near coastal terminals. These targets exhibit typical radar backscatter characteristics and were well-suited for validating oil spill detection algorithms.

Figure 4. Data preprocessing workflow for X-band marine radar images.

Figure 4. Data preprocessing workflow for X-band marine radar images.

2.4. SBR Feature Extraction

Image features are widely used in computer vision analysis, especially in tasks such as image segmentation and object detection. The SBR is a local statistical feature designed to highlight regions with high internal intensity and well-defined boundaries.

In marine radar images, oil slicks often produce weak but coherent echo responses. These characteristics make them difficult to detect using traditional edge detection or texture-based feature methods. SBR addresses this issue by calculating the ratio between the number of foreground pixels and the perimeter length of a local window. This ratio reflects the structural complexity of the region.

Compared with the surrounding background, oil slicks usually have higher SBR values. They show more concentrated structures and clearer boundaries. Therefore, the SBR feature provides a reliable basis for region extraction and fine segmentation.

The calculation of the SBR feature follows these steps: First, the radar image is converted into an 8-bit grayscale image. Then, a sliding window of size L × L is applied across the entire image. In each window, local binarization is performed using the following formula:

(1) The original image is converted to 8-bit grayscale image.

(2) Each local window (L × L) was binarized using the threshold T:

$$T=m-fs$$

(1)

where m denotes the local mean, $s$ represents the local standard deviation, and f is a constant used to adjust the threshold flexibility, typically set to 2.5 in this experiment.

(3) The SBR for each local window is calculated by determining the number of foreground pixels and the local window boundary length:

$$\mathrm{SBR}={\displaystyle \frac{log\left(N\right)}{log\left(P\right)}}$$

(2)

where $N$ represents the number of foreground pixels. $P$ denotes the local window boundary length as follows:

$$P=4\left(L-1\right)$$

(3)

where L is the side length of the local window.

(4) The SBR feature values are normalized.

(5) A threshold is set to binarize the SBR feature values for extracting the oil spill regions.

The process of SBR extraction, including sliding window scanning, local thresholding, and logarithmic feature computation, is visually summarized in Figure 6.

2.5. The Improved Sauvola Adaptive Thresholding Algorithm

Threshold-based segmentation is a widely used technique in image processing. Among them, local adaptive thresholding methods like Niblack and Sauvola have shown promising results in document and texture segmentation. However, their effectiveness on marine radar images is limited due to low contrast, weak edges, and strong background noise.

Niblack [37] introduced a simple local thresholding method, where the threshold T is computed as follows:

$$T=m+k·s$$

(4)

where m and s are the local mean and standard deviation, respectively. k is a user-defined parameter, typically negative. This method tends to over-segment in noisy regions.

To improve robustness, Sauvola and Pietikäinen [38] proposed a modified version:

$$T=m\times \left[1+k\left({\displaystyle \frac{s}{R}}-1\right)\right]$$

(5)

where R represents the dynamic range of standard deviation (usually 128), and k is a positive constant (typically 0.5). This formulation reduces the effect of noise in homogeneous regions. However, it is still sensitive to global brightness and less effective for radar images with weak gradient transitions.

To address these issues, Xu et al. [39] introduced a variance-based variant for oil slick segmentation:

$$T=m\times \left[1+k\left({\displaystyle \frac{v}{R^2}}-1\right)\right]$$

(6)

where v is the local variance, replacing the standard deviation. This approach strengthens threshold response in high-variance regions.

This study enhance this method by introducing global variance normalization:

$$T=m\times \left[1+k\left({\displaystyle \frac{v}{v_{max}}}-2\right)\right]$$

(7)

where $v_{max}$ denotes the maximum variance across all local windows in the image. The use of $v_{max}$ allows the threshold to adapt proportionally to local fluctuation intensity relative to the global extremum, enhancing contrast in low-intensity regions such as radar oil slicks. The constant k was set to 0.5, following empirical tuning based on segmentation performance.

Compared to previous approaches, this improved Sauvola method offered better boundary preservation and noise suppression for weak and irregular oil spill targets. Its adaptivity to local variance ensures more stable segmentation results under varying sea states and environmental conditions.

2.6. Postprocessing

After the initial segmentation using the improved Sauvola adaptive thresholding algorithm, the result may still contain isolated speckle noise and small false-positive regions. To further improve the accuracy and overall quality of the segmentation, several postprocessing steps were applied.

First, small connected components were removed using area-based filtering. These small regions were usually caused by radar clutter or background disturbances and do not exhibit typical oil spill characteristics. To adapt to different image resolutions and noise distributions, an adaptive regions threshold was set based on empirical observations rather than using a fixed pixel value. This improved the flexibility and robustness of the postprocessing.

Next, morphological closing was applied to the initial segmentation mask. This operation helped to smooth the boundaries of oil spill regions and fill small holes within them, improving the continuity and structural integrity of the segmented areas.

Finally, the processed binary mask was overlaid on the original radar image. Through noise removal and boundary refinement, the visibility of the oil spill targets was enhanced, and background interference was significantly reduced. These postprocessing steps validated the effectiveness and practicality of the proposed strategy.

3. Experiment

3.1. Pre-Treatment Process

Before segmentation, the radar images were preprocessed to remove noise and improve image quality. The full preprocessing steps are described in Section 2.3.

The experimental images were collected using the X-band radar system on the training vessel Yukun. The oil slicks in this experiment mainly came from crude oil leaks near tankers and storage terminals. These slicks showed typical radar features such as dark patches with smooth texture and clear boundaries.

As shown in Figure 5e, the preprocessing steps removed co-frequency interference and speckle noise. Grayscale correction and contrast enhancement helped make the oil slicks more visible. These results provided clean and reliable input for the next stage of segmentation.

3.2. Extraction of the Rough Oil Spill Regions

To obtain an initial estimate of oil spill locations, the Significance-to-Boundary Ratio (SBR) values were calculated in each local window of the radar image and normalized to the range [0, 1]. Based on the normalized values, a threshold of 0.55 was applied to extract rough oil spill regions.

This threshold was selected empirically after comparing different settings. A lower threshold such as 0.35 increases recall but introduces a large number of false positives. In contrast, a higher threshold such as 0.75 reduces noise but may lead to missed detections. Experimental results demonstrated that the intermediate value of 0.55 achieves a good trade-off between precision and recall.

The rough extraction phase was designed to locate potential spill regions without requiring pixel-level precision. It identified continuous dark patches with strong local structure, which were typical of oil slicks in marine radar images. Although some background noise or non-oil regions may be included, this stage ensures that true spills were retained with high probability.

As shown in Figure 7, the extracted regions provided a rough but representative outline of potential oil spill regions. The resulting mask maintained boundary continuity and spatial coverage, which supported the accuracy of the next step. These results served as the input for the subsequent adaptive threshold segmentation, laying the foundation for more refined and accurate oil spill detection.

3.3. Preliminary Segmentation Result

To further refine the rough oil spill regions, an improved Sauvola adaptive thresholding algorithm was applied. This method dynamically adjusted the threshold based on local intensity variance, allowing better segmentation of low-contrast regions. As shown in Figure 8a, the initial segmentation outlines the oil slick shapes more precisely than simple global thresholding.

Next, isolated black and white speckles were removed, which are common noise artifacts in radar images. These speckles were eliminated using a combination of area filtering and median smoothing. This step helped clean the mask and enhanced boundary consistency, as shown in Figure 8b.

Finally, the refined segmentation mask was fused with the denoised radar image for visualization. This fusion highlights the spatial alignment of the detected oil slicks with the background sea surface, making it easier to interpret the results, as shown in Figure 8c. The fusion result is also useful for downstream processing such as visual inspection or decision support.

4. Discussion

4.1. The Impact of Threshold Selection in SBR Feature Extraction

The selection of threshold values in the Significance-to-Boundary Ratio (SBR) feature extraction stage plays a decisive role in the quality of oil spill detection. Since the SBR value determined whether a local window was classified as part of an oil spill candidate region, the threshold directly affected the segmentation outcome. To analyze this effect, comparative experiments using three different thresholds were conducted: 0.35, 0.55, and 0.75. The corresponding results are illustrated in Figure 9.

When the threshold was set to 0.35, the system exhibited a tendency to over-detect. A large number of regions were classified as potential oil spills, including many areas that did not actually contain oil. This threshold was sensitive to weak or noisy SBR values, causing background fluctuations and surface textures to be misclassified as oil slicks. However, this setting preserved the overall morphology of the spills more completely, with fewer internal discontinuities. As shown in Figure 9a, while the region continuity was high, the false positive rate increased significantly.

In contrast, the threshold of 0.75 led to the opposite effect. The model became highly conservative, retaining only the strongest SBR regions while discarding weaker but still valid oil spill regions. This setting effectively reduced false detections but introduced a high risk of under-segmentation and fragmented regions. In Figure 9c, many genuine oil spill regions were missed entirely, leading to poor completeness and broken spill contours.

The intermediate threshold value of 0.55 offered the most balanced performance. It preserved the main oil spill contours while effectively suppressing irrelevant background noise. The boundaries of the detected regions were clearer and better defined, which contributed to improved overall segmentation accuracy. As depicted in Figure 9b, this setting provided a good compromise between recall and precision.

The results of this experiment demonstrated that threshold selection in SBR extraction must consider both the target integrity and background suppression. While lower thresholds increased sensitivity, they were more prone to noise. Higher thresholds improved specificity but risked losing valuable information. Based on this analysis, a threshold value of 0.55 was recommended for typical marine radar scenarios, offering a robust trade-off suitable for general applications.

4.2. Comparison of Different Thresholds in Oil Film Segmentation

To evaluate the performance of different thresholding strategies for oil spill segmentation, three algorithms were compared: the original Sauvola method (Equation (4)), a previously improved variant (Equation (6)), and our proposed method (Equation (7)). The comparison focused on boundary clarity, shape completeness, and robustness under noisy conditions. Results are shown in Figure 10.

The original Sauvola algorithm worked well when the contrast between oil spills and background was high. It can segment slicks accurately in ideal conditions. However, in low-contrast scenes with complex ocean backgrounds, the algorithm was easily affected by noise. As shown in Figure 10a, many false positives occurred, and the detected boundaries were blurry and fragmented.

The previously improved Sauvola algorithm (Equation (6)) introduced modifications to better adapt to local variance. This method improved oil spill recognition compared to the original version. However, segmentation errors such as over-detection and missed regions still remained. As shown in Figure 10b, the boundaries were more complete but not fully reliable in certain regions.

Based on these observations, an enhanced version of the Sauvola algorithm (Equation (7)) was proposed, which adjusts the threshold using normalized local variance. Our method achieved clearer segmentation by effectively suppressing background noise while preserving the overall shape of the oil slicks. As shown in Figure 10c, the boundaries were sharper and better aligned with the true extent of the oil spill.

It is worth noting that some regions with weak wave echoes were not classified as oil spills. These regions may represent suspected slicks or background variation. The experiment focused on identifying actual oil spills with higher confidence, and did not attempt to segment ambiguous or low-confidence regions.

Overall, the improved algorithm demonstrated better robustness and segmentation performance compared to the previous approaches, particularly in noisy and low-contrast marine radar environments.

4.3. Validation of Experimental Results

Under nighttime conditions, infrared or laser-induced fluorescence imagery is often used to verify the accuracy of marine radar-based oil spill detection methods. In this study, marine radar data were collected during nighttime operations, along with corresponding thermal infrared images. To visually validate the detected oil spill regions, Figure 11 presents a visible-spectrum image captured at the accident site. The image was taken near a coastal terminal in Dalian Bay during the Dalian 7.16 major oil spill incident. In the thermal infrared image, the grayscale intensity of the oil slick appears slightly lower than that of the surrounding seawater, helping to distinguish its features. The visible image clearly illustrates the spatial distribution of surface oil, further confirming the feasibility and reliability of the proposed detection method.

4.4. Segmentation Accuracy Evaluation

As shown in Figure 12, in the process of expert visual interpretation, oil spill targets are manually delineated based on subjective judgment. This manual annotation inherently introduces boundary inaccuracies due to human error. In contrast, the proposed method extracts oil spill targets using an adaptive thresholding approach, which provides relatively more consistent and objective segmentation boundaries. This discrepancy is one of the key reasons behind the observed differences in precision and recall scores.

In the process of expert visual interpretation, oil spill targets need to be manually delineated, which inevitably introduces boundary inaccuracies due to human subjectivity. In contrast, the proposed method employs an adaptive thresholding approach for automatic extraction, yielding more consistent and objective segmentation boundaries. This fundamental difference contributes to the deviation observed in precision and recall scores.

As shown in Figure 13, several typical error sources further explain the discrepancies. In the blue box region, the presence of vessels or moderate-intensity reflective elements affects the performance of adaptive thresholding, making it difficult to accurately extract oil slicks in that area, whereas expert interpretation can intuitively exclude such interference. Meanwhile, in the yellow box, the oil spill exhibits high similarity to the background, which hinders the thresholding algorithm from achieving the same level of discrimination as manual interpretation. Additionally, the green box highlights the impact of ship wake interference, which can lead to misidentification of oil spill boundaries in the automated result.

Despite these sources of error, the proposed method still achieved favorable evaluation results, with a precision of 81.0%, a recall of 67.5%, and an F1 score of 73.6%, indicating that the algorithm performs reliably even when compared to expert-delineated ground truth under complex conditions.

5. Conclusions and Future Work

5.1. Conclusions

This study proposed an oil spill detection method based on the SBR feature combined with an improved Sauvola adaptive thresholding algorithm. The approach effectively addressed the challenges of weak contrast and background interference in marine radar imagery. Experimental results demonstrated that the method achieves superior segmentation performance compared to traditional and previously improved thresholding techniques, particularly in complex sea environments.

The proposed technique provides a reliable and efficient tool for the timely identification of oil spills, which is critical for emergency response and marine ecological protection. By preserving the morphology of oil slicks while suppressing noise, the method enhances both detection accuracy and visual interpretability.

However, the current approach still relies on parameter tuning for optimal threshold settings, which may affect its generalizability in varied scenarios. In future work, we plan to introduce automatic parameter selection mechanisms and explore the detection of suspected oil slicks with lower confidence levels, thereby improving the adaptability of the model to diverse marine conditions.

5.2. Limitations and Future Work

While the proposed method demonstrates robust performance in detecting clearly visible oil slicks using X-band marine radar, several limitations remain.

First, the threshold value (e.g., 0.55 for the SBR map) is empirically selected based on experimental observations. This fixed threshold may not generalize well across varying marine environments or image conditions. Future work will explore adaptive thresholding techniques, such as Bayesian optimization or local-statistics-based dynamic adjustment, to improve the adaptability of the method.

Second, although the method is lightweight and does not rely on large-scale training data, it lacks the learning capability offered by modern machine learning approaches. Due to the limited availability of labeled real-world radar data and the need for low computational complexity in field deployment, deep learning models were not included in this preliminary study. Nevertheless, integrating simple learning-based modules, such as Random Forest or SVM classifiers, may enhance robustness when detecting ambiguous regions or dealing with complex backgrounds.

Third, the current validation dataset mainly consists of well-defined oil slicks collected from a real accident. The method’s performance under more challenging conditions, such as diffuse slicks, shadow zones, and non-oil anomalies, has not been fully assessed. Additional datasets and scenario-specific evaluations will be considered in future studies.

Finally, to further improve segmentation precision, especially around oil spill boundaries, the integration of lightweight deep learning models, such as simplified U-Net architectures, will be explored in subsequent work.