Explainable deep learning with novel marine domain metrics for oil spill detection

1 Introduction

Oil spills threaten marine environments, requiring rapid detection for effective response. Traditional methods for monitoring oil spills have proven to be limited in terms of speed, coverage and weather conditions (Li et al., 2008; Wan and Cheng, 2013; Yekeen and Balogun, 2020). Over the years, Synthetic Aperture Radar (SAR) imagery has emerged as an effective alternative, allowing large area coverage, regardless of weather conditions (Longman et al., 2016; Prastyani and Basith, 2018; Zeng and Wang, 2020). However, differentiating true oil spills and look-alikes remains a challenge (Chen and Wang, 2022; Goes et al., 2023).

Deep learning models have achieved high accuracy in oil spill detection and monitoring but due to their “black-box” nature, it remains hard to totally trust them in high stakes environmental monitoring (Zhan et al., 2024). Stakeholders hesitate to act on unexplained alerts, especially given the consequences of false positives. This necessitates the incorporation of Explainable AI (XAI) into detection systems to aid human-interpretable explanations for model predictions.

Oil spill detection systems are crucial to achieving several United Nations (UN) Sustainable Development Goals, including Goal 6 (Clean Water and Sanitation), Goal 7 (Affordable and Clean Energy), and Goal 14 (Life Below Water) (United Nations, 2025). This work demonstrates how explainable AI connects technological innovation with environmental protection through automated and reliable monitoring systems.

We propose an explainable deep learning framework for oil spill detection based on SpillNet, a CNN tailored for SAR imagery. Five post hoc XAI techniques—Gradient-weighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), Integrated Gradients and Layer-wise Relevance Propagation (LRP) are integrated to explain SpillNet’s decisions and the prominent features it uses to differentiate oil spills from look-alikes. Furthermore, we introduce a comprehensive 7-metric evaluation framework combining traditional XAI metrics (Explanation Fidelity, Localization Accuracy, Context Concordance, Explanation Confidence) with novel marine-specific metrics (Marine Domain Relevance, False Positive Analysis and a domain-tailored composite score, Domain Alignment Score) to quantitatively evaluate explanation quality in oil spill monitoring contexts. This study directly supports several UN Sustainable Development Goals, including SDG 6 (Clean Water), SDG 7 (Clean Energy), and SDG 14 (Life Underwater), by improving environmental protection through reliable AI-based monitoring systems.

Recent developments in the use of AI have contributed to timely and effective detection and mitigation of oil spills with improvement in the precision of environmental monitoring (Umaha et al., 2025). This is because AI models have the capability to detect and classify oil spills from different and complex datasets such as the dataset from sensors and satellite imagery or combination of both. Nevertheless, the successful deployment of AI models is limited by interpretability, sometimes requiring experts to derive useful insights for effective decision making. XAI solves the problem of interpretability by providing precise and explanations on how the results are obtained. Thus, the use of XAI for oil spill detection demonstrated in this study can promote trust and reliability in the deployment of automated systems for oil spill detection by providing explanations on why certain areas are flagged including information on the features that enabled the detection. This study bridges the gap between technological solution deployment and practical environmental scenarios, thereby enhancing a reliable and effective data driven decision geared towards oil spill detection and mitigation.

This study is novel in that the deployment of XAI for oil spill detection geared towards SDGs 6,7 and 14 have not been sufficiently discussed in the literature.

The proposed framework bridges the gap between deep learning models that are considered “black boxes” and operational decision-making. Unlike the conventional object detection model such as the SAR that is prone to “look-alikes” (Zakzouk et al., 2025), the XAI-integrated learning model isolates the physical scattering features in SAR imagery, precisely suppressing the speckle noise, thus, ruling out the possibilities of marine look-alikes. This will allow domain experts conduct the model’s spatial analysis effectively to ensure that the AI detection is based on physical damping characteristics instead of superficial artifacts. Although the SAR is excellent for weather monitoring, the proposed model in this study can translate complex deep learning features into transparent, parameters that the marine experts utilize (Murugan et al., 2026).

This study contributes to the conservation of freshwater bodies from oil pollution and ensuring safer water resources in line with SDG 6. It also contributes to SDG 7 by proposing the deployment of a more sustainable technique to detect and prevent oil leaks to ensure energy production as well as safe crude oil exploration. In addition, it provides a real time technique for conserving marine biodiversity, to prevent ecological destruction. Through effective crude oil spill detection oil producing communities as well as those depending on oceans can be preserved. Hence, this study leverages on XAI for oil spill detection, thereby providing a sustainable and responsible roadmap for this study highlights a pathway toward more resilient, sustainable, and ethically responsible for environmental monitoring.

2 Related work

Over the years, deep learning technologies have come to be effective for oil spill detection in marine environments. Akhmedov et al. (2024) showed that combining synthetic aperture radar (SAR) data with machine learning algorithms achieves high detection accuracy. Existing studies have demonstrated the application of deep learning models such as the Deep Belief Network, Residual Network, Convolutional Neural Network (CNN), for oil spill detection, segmentation and classification analysis in SAR images (Huang and Wang, 2014; Guo et al., 2017; Gallego et al., 2018). Also, Shaban et al. (2021) as well as Zeng and Wang compared the performance of the traditional machine learning models with the deep learning models and highlighted that deep learning significantly outperforms traditional methods in segmentation tasks [[11]](https://www.zotero.org/google-docs/?Wo1uOV). For instance, Chen and Wang (2022) employed the integrated attention gate U-Net model for oil speed detection on the marine surface and found out that the performance of the proposed model exceeded the traditional machine learning. Umaha et al. (2025) stated that the different layers and networks allow it to learn important and complex features in the dataset automatically unlike the machine learning models that rely on explicit pre-processing and feature engineering. Zhan et al. (2024) proposed the use of the integrated CNN-DBSCAN model for oil spill detection and found out that the proposed model excelled in detection speed compared to other models such as the CNN-visual transformer.

By using polarimetric and contextual data, studies on model architecture, such as the Attention U-Net by Chen and Wang (2022), have shown improved detection. Akhmedov et al. (2024) also introduced a framework using large datasets, improving detection in marine environment. Vasconcelos et al. (2023) showed the importance of domain-specific metrics to address classification challenges in oil spill detection.

Explainable artificial intelligence (XAI) methods have undergone major transformation to improve transparency and trust among stakeholders in various fields, including environmental monitoring. Singh et al. (2024) highlighted the benefits of combining descriptive models with deep learning models to reduce potential biases due to environmental factors such as “look-alikes”. Similarly, Linardatos et al. (2021) emphasized the importance of increasing trust in AI systems by using XAI techniques in conjunction with deep learning models to improve oil spill detection. Arnob et al. (2025) integrated the CNN with the XAI for oil spill segmentation task and found that the proposed method outperformed the traditional machine or deep learning models by 32.09%. Bui et al. (2024) achieved a mean Intersection over Union by 72.49% via deep learning and data augmentation.

Despite the advancements made so far, there remains challenges which include high noise levels in SAR images and the processes involved in distinguishing oil spills from look-alikes, as noted by Fan et al. (2021). The introduction of explainability in deep learning models proved pivotal for enhancing decision-making processes in environmental monitoring, ultimately reshaping oil spill detection in SAR images.

Thus, existing studies clearly demonstrate that use of the XAI model for oil spill detection boasts of high detection accuracy which implies a significant progress in the deployment of technology for environmental monitoring marine environments, and mitigation of the adverse impacts of oil spills. However, there still exist the challenge of look-alikes which leads to false alarm or false detection thus affecting the accuracy of detection. Hence, the need for the introduction of novel metrics that can mitigate the effect of look-alikes on the accuracy of oil spill detection.

Table 1 captures the comparative advantages of the proposed evaluation metrics vis-à-vis the conventional metrics.

TABLE 1. Comparative advantage of the proposed evaluation metrics vis-à-vis the conventional metrics.

Thus, existing studies clearly demonstrate that use of the XAI model for oil spill detection boasts of high detection accuracy which implies a significant progress in the deployment of technology for environmental monitoring marine environments, and mitigation of the adverse impacts of oil spills. However, the introduction of the proposed evaluation metrics will enable proper capturing of the edge boundaries to prevent false alarm due to natural look-a-likes (Brekke and Solberg, 2005) and effectively mitigate the effect of look-alikes and speckle noise to prevent false alarms and its associated cost implications.

3 Methodology

3.1 SpillNet architecture evolution and XAI implementation

The developed XAI model integrates the SpillNet, a customised Convolutional Neural Network (CNN) architecture with five XAI techniques. The proposed model leverages the potentials of SpillNet and XAI evaluation framework for oil spill detection in SAR images. For this study, we implemented an enhanced version of SpillNet (Umaha et al., 2025), optimized for better generalization and XAI compatibility.

3.1.1 Enhanced SpillNet for XAI analysis

First, we implemented a simplified yet robust version of SpillNet designed for improved generalization and explanation quality. The modified and improved architecture replaces the custom encoder with a pre-trained MobileNetV2 backbone, selected over alternatives (ResNet, EfficientNet) for its optimal parameter efficiency (6.4M vs. 25M + parameters) while maintaining feature extraction quality suitable for XAI analysis, and this provides a balance between better feature extraction and computational efficiency. Secondly, we implemented transfer learning by partially freezing the backbone layers during training. We used a simplified attention module that applies global average pooling in place of the Squeeze and Excite blocks in the original Spillnet model, followed by channel-wise attention using fully connected layers as expressed in Equation 1.where x ¯ is the GAP x, W ₁, b₁ are the weights and bias of first fully connected layer and W ₂, b₂ are weights and bias of second fully connected layer.

A t t e n t i o n x = x ⊙ σ W 2 R e L U W 1 x ¯ + b 1 + b 2 (1)

The decoder uses three up-sampling blocks with progressive channel reduction (256→128→64→32), with each convolutional layer incorporating L2 regularization, batch normalization for training stability, dropout for regularization (rates: 0.4, 0.4, 0.2, 0.0). This was empirically determined through validation experiments to get the best efficiency compared to uniform dropout strategies and bilinear up-sampling for smooth recovery of spatial resolution. The final model achieved a total of 6,417,541 parameters with an enhanced architecture optimized for XAI analysis.

The developed SpillNet model was benchmarked against foundational and contemporary state-of-the-art (SOTA) segmentation model. This was done to confirm its operational effectiveness and performance in relation to other SOTA models. A comprehensive baseline evaluation was introduced across five SOTA architectures: namely U-Net, Attention U-Net, DeepLabV3+, SegFormer (Transformer-based), and Swin Transformer (Hierarchical Transformer). This stage compared the binary segmentations obtained for each model determine the model architecture that is operationally superior, competitive in mapping the oil spills using evaluation metrics such as MDR, FPA, and DAS directly from the raw final outputs of each of the models.

3.2 XAI techniques and iterative feedback loop

To improve the explainability of SpillNet’s predictions, we integrated five state-of-the-art post hoc XAI techniques into the model development pipeline. Without modifying the model, these techniques provide feature attribution explanations.

Grad-CAM creates heat maps of the most influential regions by back-propagating class gradients to the final convolutional layer. LIME identifies superpixels and approximates the CNN locally by fitting a simple surrogate model. SHAP attributes importance values to each feature using a game-theoretic approach, clearly distinguishing pixels that increase oil spill probability (red) from those that decrease it (blue). Integrated Gradients works by comparing the model’s response to a null matrix and the input image, tracking how the prediction changes along this path, and adding up the pixel-wise gradients. Layer-wise Relevance Propagation (LRP) traces the model’s prediction backwards through the layers, distributing the prediction score across all neurons while keeping the total unchanged, producing clear, pixel-specific heatmaps.

We applied all five XAI techniques to the trained enhanced SpillNet model and evaluate their explanation quality using existing metrics and the three-novel metrics. These include the Marine Domain Relevance (MDR) for the quantification of oil spill, False Positive Analysis (FPA) for look-alike discrimination and Domain Alignment Score (DAS); an expert-based checklist with composite metric. This allows for a comparative analysis among the explanation techniques. The process of spill monitoring using the SAR relies on pixel-level accuracy which may not directly translate directly into accuracy or operational validity. This is due to the fact that oil-induced capillary wave damping produces distinct edge gradients that separate mineral slicks from look-alikes like biogenic films (As established (Brekke and Solberg, 2005). Thus, the introduction of MDR solves this by examining the physical edge and spatial coherence of the detection instead of relying on a simple pixel count. Look-alikes and speckle noise could also induce significant false alarms in standard segmentation models. Hence, with the introduction FPA false alarms can be reduced by suppressing large contextual look-alikes (Solberg et al., 2007). Furthermore, standard metrics may fail to capture the explainability of the model’s spatial priors whereas the introduction of DAS can enable the integration of structural, radiometric, and contextual metrics into a combined score. This will ensure that the model makes classifications based on true physical oceanographic signatures rather than background noise or clutter artifacts (Topouzelis, 2008).

The robustness of the proposed model was enhanced with the use of strategies such as bilinear up-sampling, and conservative regularization to enable explanation of image quality during analysis.

3.3 Comprehensive seven-metric evaluation framework

Our framework evaluates marine XAI explanations using four existing metrics and three novel domain-specific ones.

3.3.1 Traditional XAI metrics

Explanation Fidelity (EF): Measures how faithfully the explanation represents the model’s actual decision-making process by comparing predictions with and without important features identified by the explanation.

Localization Accuracy (LA): Evaluates the spatial precision of explanations by computing the Intersection over Union (IoU) between highlighted regions and ground truth oil spill locations.

Contextual relevance (CR): Assesses whether explanations focus on contextually relevant features within the broader context of the image, rather than on spurious correlations.

Explanation Confidence (EC): Reflects attribution concentration by computing normalized entropy of the attribution map. This is expressed by Equation 2.Where H is expressed as Equation 3.and p i which represents the normalized attribution values, and maximum possible entropy is expressed as Equation 4, while H max is expressed as Equation 5. Where a _{*i* *​*}represents the attribution value for pixel i, and N is the total number of pixels. The higher the EC the more focused the explanations and vice versa. This implies that the model can identify precise decisive or critical regions.

E C = 1 − H H max (2)

H = − ∑ i p i ⁡ log ⁡ p i (3)

p i = a i ∑ j a j (4)

H max = log ⁡ N (5)

3.3.2 Novel marine domain-specific metrics

Equations 6a,b, presents the Marine Domain Relevancy (MDR) expression. This is a metric suitable for the assessment and detection of marine oil spill detection. It is an advancement over the conventional XAI assessment metrics which is vulnerable to false positive and missed detection. However, the inclusion of domain-specific knowledge of the features of the spill can reduce false positives and missed detections. MDR integrates three different characteristics for the marine environment. These include: edge sharpness, spatial consistency and background reduction. The edge sharpness indicates how the features accurately reflect the oil slick’s edge characteristics, which is necessary for expert identification standards while the spatial consistency assesses the uniformity across a uniform oil slick. This reflects the expected physical distortions in slick oil formation. Lastly, the background reduction feature which evaluate the reduction in contrast to the water surface. This is necessary for the separation of signal from noise.where: E P means edge prominence while S C depicts spatial coherence and B S represents the background suppression. w 1 , w 2 and w 3. These three parameters represent the min-max normalized scores confined within the range [0, 1].

M D R = w 1 E P + w 2 S C + w 3 B S (6a)

Where w₁ = 0.4, w₂ = 0.3, w₃ = 0.3, ensuring that ∑ i = 1 3 w i = 1, MDR ϵ 0 , 1. Hence, 1 is an ideal and interpretable oil spill detection.

Thus,

M D R = 0.4 E P + 0.3 S C + 0.3 B S (6b)

The EP was assigned the highest weight of 0.4 based on the relative importance its feature. The sharp backscatter gradient boundary between the oil slick and the surrounding clean sea surface is the most important feature. XAI saliency maps relies on the edge gradients to distinguish between actual oil slicks from the natural look-alikes such as the low-wind zones which may have blurry and poorly-defined boundaries (Topouzelis, 2008).

Secondly, the SC was assigned a weigh of 0.3 to ensure the XAI-highlighted pixels form a continuous, cohesive geometry that depicts a realistic oil slick structure, eliminating fragmented speckle noise or random isolated pixel activations that could affect the image resolution and accuracy of detection. Furthermore, BS was also assigned a weight of 0.3 because it measures the ability of the model ability to actively suppress ocean clutter and radar noise in the background. By this, the model will be able to maximize the signal-to-clutter ratio so that the model’s confidence is entirely driven by the spill region for improved accuracy of detection (Solberg et al., 2007).

The initial weight boundaries were established after rigorous literature review and consultation with the operational manual of the European Maritime Safety Agency (EMSA) CleanSeaNet framework. Experts’ opinions were also sought from 10 independent maritime and remote sensing experts. There was a consensus among the experts that EP take precedence and allocated the highest weight because the sharp backscatter gradient boundary which serves as the main physical feature that separates actual mineral oil spills from diffuse natural look-alikes.

Sensitivity analysis was also conducted under varying weight distributions as shown in Table 2. The outcome produced optimal weight configurations (w₁ = 0.40, w₂ = 0.30, w₃ = 0.30).

TABLE 2. Sensitivity analysis for the various weight configurations.

Equation 7 expresses the False Positive Analysis (FPA). This is a metric that addresses the major issue of environmental monitoring by the stakeholders. It evaluates the ability to correctly identify and suppress attention on known oil spill look-alikes. It provides weighted regional analysis in areas where the look-alike discrimination receives the highest weight (0.6) as a result of operational criticality in differentiating between oil spills from natural and non-natural phenomenaWhere: w l o o k a l i k e = 0.6, w o i l = 0.3 , w b a c k g r o u n d = 0.1 and A r is the accuracy score.

F P A = ∑ r ϵ R e g i o n s w r × A r (7)

The Domain Alignment Score (DAS) (Equation 8) is a novel composite metric that analytically quantifies the alignment of deep learning feature attributions with standard marine oceanography principles. The structural criteria rooted in DAS are directly derived from the operational guidelines of the marine domain expertise from the European Maritime Safety Agency (EMSA) CleanSeaNet operational guidelines (European Maritime Safety Agency, 2025), and the National Aeronautics and Space Administration (NASA). It uses remote sensing training protocols, as well as peer-reviewed literature on marine remote sensing. This metric is an advancement over the general weighted combinations, in that it employs an equal weighting framework of 8 criteria that encompasses spatial features (such as even homogeneity, boundary clarity, spatial continuity), look-alike discrimination (such as algal bloom, low wind and ship wake suppressions, as well as contextual factors (such as thickness indication and background suppression).

The criteria were derived through a systematic literature review of over fifty papers on marine remote sensing and validated against operational protocols from EMSA’s 15 years of experience in CleanSiNet and NASA’s SAR training curriculum. This literature-based approach eliminates subjective bias while ensuring a comprehensive assessment consistent with the operational requirements of marine monitoring.

D A S = 1 N ∑ i = 1 N C i (8)

Where: N is the total number of expert-defined criteria and C _*i* is the individual criterion score defined according to expert operational guidelines detailed in Table 3 (normalized to [0,1]).

TABLE 3. Expert checklist criteria for domain alignment score.

Table 3 presents the expert checklist criteria for domain alignment score used as the validation mechanism used to prove Equation 8.

A multi-disciplinary panel of 10 independent maritime and remote sensing experts comprising of Radar Oceanographers (4), Marine Environmentalists (3), and Deep Learning/XAI Engineers (3) to balance physical radar science with AI interpretability with experience ranging from 10 to 15 years employed the checklist presented in Table 3 to independently grade the significance and alignment of the metric’s parameters across a representative validation subset.

To ensure the scientific validity of the checklist, Inter-Rater Reliability (IRR) testing was conducted which yielded a Fleiss’ Kappa coefficient (k = 0.821, p < 0.001). According to Fleiss (1971), a Kappa score between 0.81 and 1.00 indicates “Almost Perfect Agreement” among the raters. This shows that there is a significant consensus among the rater which further suggests that the checklist evaluation is robust, reproducible, and bias-free.

The validation of the checklist items was achieved using a Content Validity Index (CVI). This produced a high Scale-level index (S-CVI/Ave = 0.92), thus, validating its alignment with operational standards (Lynn, 1986; Polit and Beck, 2006).

To separate the precise performance the proposed metric components, a systematic ablation study was conducted to assess the performance of the proposed SpillNet model under three stripped configurations: (i) excluding Edge Prominence, EP (ii) excluding Spatial Coherence, SC and (iii) excluding Background Suppression, BS.

3.4 Implementation and computational setup

Enhanced SpillNet Training: Our enhanced SpillNet implementation employed a conservative training approach, designed to improve model generalization and explainability. The model was trained with MobileNetV2 as the backbone, using transfer learning, where early layers were initially frozen and progressively unfrozen during training. The following configuration was used during training:

• Input resolution: 224 × 224 pixels.
• Batch size: 8.
• Learning rate: 1e-4 with ReduceLROnPlateau scheduling.
• Regularization: L2 weight decay (1e-4) and dropout (0.4).
• Class weights: Conservative weighting scheme (Oil spill: 15.0, Sea surface: 1.0, Look-alike: 3.0, Ship: 30.0, Land: 3.0).
• Optimizer: AdamW with Weight Decay = 1 × 10⁻⁴ to ensure training stability and to prevent prevents learning rate decay errors.
• Initialization Strategy: He (Kaiming) Normal/Pre-trained ImageNet. This optimizes the initial gradient flow across the MobileNetV2 backbone.
• Training Time: Approx. 3.2 h (30 Epochs) to achieve full convergence GPU Specifications: Single NVIDIA Tesla T4 (16 GB VRAM) or an NVIDIA GTX 1080 Ti (11 GB VRAM). The algorithm latency limited the samples to 20 because of computational time, cost and intensiveness,
• Random speed: 42. It guarantees deterministic reproducibility for weights and data splits.
• Inference Time: 45.5 Milliseconds (ms) per image patch.
• XAI Generation Time: (a) Grad-CAM: 80.2 m per image, (b) LRP: 100.4 m per image (c) Integrated Gradients: 1,720.0 m (1.72 s) per image (d) LIME: 8,100.0 m (8.1 s) per image € SHAP (KernelSHAP): 13,850.0 m (13.85 s) per image.

A comprehensive multi-stage ablation analysis across six variables namely: MobileNetV2 backbone, transfer learning, attention module, dropout, regularization and preprocessing pipeline was conducted to the contribution of introduced variables modification.

Dataset and Preprocessing: This study utilizes Sentinel-1C-band (with approximately 5.405 GHz) satellite SAR images from the European Space Agency’s Copernicus program. This Satellite sensor ensured stable temporal tracking and uniform radar backscatter sensitivity. Dual polarization node (VV + VH): Vertical-Transmit, Vertical-Receive and VH Vertical-Transmit, Horizontal-Receive) was utilized. While VV optimizes the oil-to-water contrast; VH on the other hand, minimizes the uniform look-alike background clutter. The dataset comprises 1,112 images with verified oil spill masks from the European Maritime Safety Agency (2025). We identified 1,002 valid image-mask pairs for our analysis acquired via the interferometric Wide (IW) Swath at a wind speed threshold of 3.0 m/s to 10.5 m/s.

The architecture of the proposed explainable SpillNet model that demonstrates the data workflow from raw dual-polarized C-band SAR input patches through the pre-processing stages to the regularized deep learning framework, and to the prediction stage evaluated by domain-specific metrics is displayed in Figure 1.

Table 4 captures the other information on the dataset employed.

TABLE 4. Dataset profile (N = 1,002).

To avoid bias due to random selection and to ensure that the 20 samples are representative of the 1,002-image dataset, the systematic stratified sampling protocol was employed as captured in Table 5 (Cochran, 1977; Kuhn and Johnson, 2013). This method eliminates sampling bias to ensure that the XAI analysis is generalizable over the operational scenarios concerns. The first step is the portioning of the selection matrix into five different strata that reflect the particular physical challenges of operational maritime surveillance. The 20-image samples comprise of 12 oil spill scenarios and 8 complex look-alike conditions. The geographical and environmental representation was ensured by drawing the samples from the four major oceanographic regions namely: the Gulf of Mexico (6), the North Sea, the Mediterranean Sea, and the South China Sea. Sampling bias was mitigated by implementing two criteria. The first criteria ensures that the wind speed distributions within the 20 patches correspond to the mean distribution of the broader 1,002-image dataset (Mean: 6.2 m/s). Secondly, it was ensured that the selected visual boundaries did not contain any no processing artifacts or land-clutter bleeding.

TABLE 5. Systematic stratified sampling protocol.

The preprocessing phase comprises of five major steps. These are (1) localization and cropping of established oil spills, (2) rescaling of images to 224 × 224 pixels to promote computational efficiency, (3) execution of radiometric calibration to correct the variations in sensor (4) application of median filtering to lessen speckle noise, and (5) normalization of the pixel values to [0,1] range for a consistent and reliable model input.

For the XAI implementation phase, the analysis was carried out on 20 representative samples extracted from the dataset. The implementation of each of the explanation techniques was done using standard libraries such as the Gradient Class Activation Mapping (Grad-CAM) through TF-Keras-vis, Local Interpretable Model-Agnostic Explanations (LIME), Shapley Additive Explanations (SHAP) with superpixel segmentation, Integrated Gradients via Tensor Flow, and LRP.

All the experiments were carried out on Google Colab Pro with GPU acceleration, employing the TensorFlow 2.x model.

The enhanced SpillNet model achieved training convergence within 30 epochs using early stopping with validation accuracy monitoring, demonstrating improved generalization compared to the original architecture.

3.5 Statistical analysis

Statistical analysis such as mean, standard deviation, -statistic/Z-value and statistical significance were conducted to validate the included metrics (MDR, FPA and DAS). The mean represents the average score of performance of the metrics across the test images. This is necessary to determine if the XAI model performs consistently on a collective basis, rather than performing optimally on one or two isolated images. The standard deviation (SD) measures the spread, variance, or volatility of the scores around the average mean which provides an insight into the model’s stability. A model with high mean and high SD is volatile and unstable. It may perform well on one and fail on others (Field, 2013). Conversely, a model with a low SD indicates model’s explainability and detection abilities are robust and stable across different sea states and wind conditions. The t-statistic/Z-value is the magnitude of difference that measures the deviation of the framework’s performance from a baseline model, relative to the variation in the data. A high t-statistic or Z-value indicates a significantly large difference in the performance gap between the proposed marine metrics and the traditional computer vision metrics and vice versa. The statistical significance (p-value) is the probability of chance that indicates the probability that the developed model is superior to random chance. Values less than 0.05 indicates that there is less than a 5% probability the output of the model is by coincidence or random chance thus indicating the robustness of the model (Field, 2013).

In addition, to compare the various XAI models, the Wilcoxon Signed-Rank test and paired t-test as well as the confidence interval and the effect size measure tests were conducted to compare and validate the superiority of the various XAI models employed. The Friedman test is a non-parametric omnibus test that examines all the XAI techniques simultaneously to show mathematically if a significant difference exists somewhere across the entire group of XAI models. A significant p-value (< 0.05 indicates the existence of a significant difference somewhere across the entire group of XAI models and vice versa.

The Wilcoxon Signed-Rank test and paired t-test are post hoc pairwise comparisons tests conducted to match the superior model directly against other alternatives once the Friedman test proves a global difference. The paired t-test examines if the average (mean) metric improvement of the superior model is statistically valid with a normally distributed data while the Wilcoxon Signed-Rank Test acts as the non-parametric counterpart that compares the medians and ranks of the paired differences to ensure a valid conclusion (Demšar, 2006). This helps in determining if the dataset contains skewed distributions or outliers triggered by complex SAR look-alikes. While the p-value identifies if statistical difference exists between the XAI models, the confidence intervals determine the range of the difference. At 95% Confidence Interval, if determines if the performance of the superior model will fall within the same range of values if the experiment is replicated severally. This indicates the high precision and predictability of the superior model (Demšar, 2006).

Lastly, the Effect Size Measures (Cohen’s d) measures how large the operational gap is between superior model and others. A large effect size (Cohen’s d > 0.8) indicates that the superior model is large enough to considerably influence oil spill detection accuracy and vice versa (Cohen, 1988).

The statistical analysis was conducted in the Python 3.11.8 environment.

4 Results and discussion

Table 6 presents the outcome of the multi-stage ablation analysis across six variables. The results show that, for the MobileNetV2 backbone and transfer learning), transitioning to MobileNetV2 integrated with pre-trained weights produced a robust baseline which significantly improved the overall segmentation convergence. The inclusion of the attention mechanism produced the most substantial single improvement in the domain-specific metrics, resulting in an increase in the MDR. This will enable the model to focus on subtle oil boundaries.

TABLE 6. Multi-stage ablation analysis evaluating Key SpillNet components (N = 1,002).

These inclusion of the final components (Dropout, Regularization, and Preprocessing) serves as critical stabilization instruments. They sufficiently prevented the model from overfitting against heavy radar speckle noise. This resulted in reduction in the FPA.

4.1 Enhanced SpillNet performance and architecture validation

Our enhanced implementation of SpillNet demonstrates superior generalization characteristics in comparison with the original model. With conservative training strategies, the model converged in 30 epochs. The performance metrics demonstrate robust learning without overfitting, and the low accuracy variation (σ = 0.024) indicates stable and well-calibrated predictions, with validation accuracy of 90.5%, ideal for the evaluation of Explainable AI (XAI).

The improved SpillNet model with MobileNetV2 as backbone had a total of 6,417,541 parameters. This provides an optimal balance between model performance and computing efficiency. The attention mechanism and the design of the progressive decoder maintained spatial coherence, thereby reducing the architectural complexity, which may have made the generation of explanations more difficult.

4.2 Outcome of the comparative analysis with SOTA models

Comparative analysis with SOTA models was conducted and findings presented in Table 7 revealed that the SpillNet significantly outperforms the Convolutional and Attention Baselines (U-Net, Attention U-Net, DeepLabV3+) tin terms of standard semantic segmentation metrics (Intersection-over-Union, IoU) and the proposed domain-specific metrics (MDR, FPA, DAS). Although, the Transformer Baselines (SegFormer, Swin Transformer) demonstrated superior pixel-level standard IoU, particularly the Swin Transformer with IoU of 0.844, the proposed SpillNet however achieved superior capabilities in terms of look-alike discrimination and background noise suppression. This results in a higher MDR and FPA. This implies that the other models may struggle to separate biogenic films and wind-shadow look-alikes from true mineral oil spills leading to false detections (false positives). The performance of the.

TABLE 7. Benchmarking SpillNet against SOTA segmentation models.

SpillNet may be attributable to the specialized spatial priors and boundary-gradient constraints (Umaha et al., 2025; Wang et al., 2026) which yields MDR of 0.885 and significantly low FPA value of 0.071. This shows that the SpillNet is operationally superior and highly competitive for marine applications such as oceanographic remote sensing.

This benchmarking validates the fact that the proposed SpillNet model is significant, highly efficient, superior and highly competitive for domain-specific SAR oil spill tracking in relation to other SOTA models.

The result of the ablation study carried out indicate that by EP, there was significant increase in the look-alike resulting in significant false alarms. This increases the FPA to 0.308. This confirms that edge prominence, EP is an important factor for effective discrimination between actual oil slicks from the natural look-alikes. Mathematically, this validates the decision to assign it the highest priority weight of 0.4 in Equation 6b.

4.3 Comprehensive seven-metric quantitative results

The analysis of 20 representative samples out of 1002 image-mask pairs revealed significant differences in performance between the XAI methods across all seven-evaluation metrics as shown in Table 8. Grad-CAM achieved the highest domain alignment score (0.608), 14.7% better than LIME (0.530), the second-best method. Although SHAP achieved a near-perfect false positive rate (0.999), its localization accuracy was very low (0.003).

TABLE 8. Results with standard deviations from 20 samples.

Figure 2 shows important trade-offs between the XAI methods. Grad-CAM shows the most balanced performance across all spatial measures (LA, CC, DAS, MDR), making it the most suitable for operational use. In contrast, SHAP, Integrated Gradients, and LRP achieve near-perfect False Positive Analysis (FPA), but lag behind in spatial accuracy. This contrast illustrates the trade-off between discrimination and localization strengths.

The LIME technique provides a middle ground by integrating high explanatory confidence with practical spatial performance.

Figure 3 shows the heat map for the evaluation of the XAI performance in terms of clustering. The Figure shows two distinct clusters. The Grad-CAM technique indicates a regular spatial strength as illustrated by the areas mapped with warm colours with moderate FPA. There was significant improvement in the FPA metrics for the SHAP, Integrated Gradients, and LRP perform as indicated in the areas mapped with green colour with evidence of slight spatial lag indicated by the dark red colour. The LIME technique gave a contrasting outcome with the EC showing strong spatial performance but with no specific uniformity. Hence, Figure 3 indicates that there is a significant trade-off between spatial precision and discrimination accuracy. Thus, the Grad-CAM technique demonstrates a unique suitability for environmental monitoring. The heat map allocates performance score to the various evaluation metrics, vis-à-vis the using different XAI techniques, to aid the understanding of the spill features. This is necessary for effective decision-making.

4.4 Sample analysis

Figure 4 presents the XAI performance on a SAR image of an oil spill with 5.47% coverage. The Grad-CAM technique gave a heat map that is well aligned with the spill boundary with a DAS of 0.608 and MDR of 0.344. This indicates the suitability of this technique for spatial analysis. The block-based LIME method gave a fairly robust result, highlighting parts of the spill while other parts are missing. The SHAP method provides a fine-grained, colour-coded attributions with a near-perfect false positive control with a FPA score of 0.999, although this method lack spatial accuracy. The integrated Gradients and LRP methods could detect the general area of the spill but could not define the boundary, thus, restraining their practical applications. The domain alignment scores of the Grad-CAM range from 0.458 to 0.732, thus, indication a consistent performance across the 20 samples analysed in this study. These findings are consistent with existing studies such as Arnob et al. (2025) who indicated that the Grad-CAM which improves transparency in the detection process, hence making it ideal for a real-world application.

4.5 Performance analysis across metric categories

In terms of Explanation Fidelity, the SHAP and LRP achieved the highest score of 0.019. This outcome demonstrates their superior reliability to decision-making process. The SHAP also excelled in FPA with a score of 0.999, indicating its ability to distinguish between showing the real scenario and its look-alikes. This is a critical performance capability of the XAI technique which makes it suitable for operational deployment.

In terms of localization accuracy and context concordance, the Grad-CAM demonstrated a dominant superiority with localization accuracy and context concordance scores of 0.184 and 0.660 respectively, thus, signifying an outstanding spatial precision and contextual importance. Furthermore, Grad-CAM also has an MDR score of 0.344 signifying its proper alignment with marine oil spill features such as edge prominence and background suppression. The LIME showed a superior performance in terms of Explanation Confidence (0.429), giving the most focused attributions that evidently differentiate the important regions from the less important. This will guide the decision makers and experts on the areas that requires significant attention and mitigation efforts.

On the overall, the DAS of 0.608 clearly established the Grad-CAM as the optimal performer, followed by the LIME technique with a DAS of 0.530. This result indicates that there is a significant performance gap between the Grad-CAM and the LIME techniques which justifies the outstanding performance of the Grad-CAM and its superior suitability for deployment in operational marine oil spill detection systems.

4.6 Statistical analysis results

Table 9 shows the outcome of the statistical analysis used to validate the performance of the MDR, FPA, and DAS metrics vis-à-vis the baseline CNN model.

TABLE 9. The outcome of the statistical analysis.

Table 10 the outcome of the comparative analysis of the various XAI model employed. The global Friedman test across all five post hoc XAI techniques simultaneously strongly rejected the null hypothesis with a p-value (< 0.05. The outcome shows a highly significant overall variance χ 2 (4) = 0.001. This implies that a significant difference somewhere exists across the entire group of XAI models and vice versa. Pairwise significance comparison was conducted which directly measure Grad-CAM against other alternatives using the Paired t-test and Wilcoxon Signed-Rank. Both tests indicate that the improvement of the Grad-CAM model is statistically robust and significant (p< 0.01 across all the various combinations. The 95% Confidence Intervals was calculated for the mean differences alongside Cohen’s d effect size measures. This produced a significantly large effect sizes (d > 1.1) across all comparisons. This confirms that the Grad-CAM’s has superior operational ability to capture sharp backscatter edge gradients compared to other alternative pixel-attribution methods.

TABLE 10. Comparative analysis of the various XAI models.

5 Discussion of results

This study achieved a validation accuracy of 90.5%. This implies that the SpillNet correctly classified 90.5% of image patches as either containing a true oil spill or being a look-alike. This indicates the ability of the model’s ability to make the correct overall decision per image frame. In terms of segmentation accuracy, it achieved IoU of 0.830 (83%). This is a pixel-level localization metric that evaluates how precisely the predicted binary masks overlap with the ground-truth pixels. In terms of the novel domain-specific detection accuracy (MDR, FPA, DAS) which evaluates whether the pixels identified by the SpillNet model using the Grad-CAM method match the physical boundaries and capillary wave damping properties required by EMSA and NASA guidelines, the proposed model achieved MDR of 0.608, FPA of 0.344 and DAS of 0.973.

The findings obtained in this study is comparable with contemporary studies in SAR oil spill segmentation. For instance, Wang et al. (2026) employed the Frequency-Spatial Coordinate and Strip-Augmented Network (FCS-Net) achieving a mean IoU ranging between 87.78% and 89.62% while Murugan et al. (2026) utilised the XAI-ViR-SC ensemble for SAR oil spill detection achieving an improved performance of 98.45%. Zakzouk et al. (2025) employed the DeepLabv3+ on SAR data achieving an accuracy of 98.14% accuracy as well as mIoU of 0.7872 while Zeng and Wang (2020) achieved high patch accuracy with OSCNet model. Furthermore, Zhan et al. (2024) also achieved high precision with a CNN-DBSCAN approach on hyperspectral data, while Zhang et al. (2022) employed D-S evidence theory achieving over 70% detection success with less than 19% false alarms. These prior studies reported high performance in SAR oil spill detection using different techniques and their outcomes are comparatively close to the IoU of 0.830 (83%) obtained in this study using the SpillNet model.

5.1 Operational deployment implications of the XAI techniques

The Grad-CAM technique has an exceptional DAS with the highest performance score of 0.608. This implies that it is the most suitable technique for real-time oil spill monitoring. The Grad-CAM framework can be integrated into an existing marine monitoring system, to provide a real-time monitoring and explanations on satellite image analysis with an automated alert validation. From the radar chart analysis displayed in Figure 2, the Grad-CAM has a balanced performance profile across all metrics coupled with a clear spatial agreement with expert expectations. This implies that the technique is trustworthy and interpretable thus suitable for deployment as a decision support. This also addresses the limitation of AI model deployment as a black box especially in deployment for environmental monitoring.

5.2 Novel Marine Metrics impact

Traditional AI systems lack the Marine Domain Relevance and False Positive Analysis that capture marine-specific explanation quality. The XAI techniques employed in this study successfully captured these metrics which makes it a novel approach. For instance, heatmap visualisation displayed in Figure 3 shows a critical trade-off between FPA control and localization accuracy. The SHAP technique achieves near-perfect false positive control of with a score of 0.999 (close to 1), but with a poor localisation accuracy of 0.003. These results can be carefully considered before the selection of the suitable XAI technique based on applications and set objectives. This outcome further demonstrates the usefulness of deploying a multidimensional evaluation approach to achieve an informed decision during the selection of suitable XAI technique. This finding agrees substantially with the finding of Chen et al. (2024) that the SHAP technique can effectively explain the discrimination between the actual regions and its look-a-likes. Chen et al. (2024) achieved significant accuracy in water detection from SAR imagery and reduction in missed detection by deploying this technique.

6 Conclusion

This study demonstrates the application of XAI for oil spill detection. The model integrates the SpillNet, with five XAI techniques. The comprehensive evaluation of 20 representative samples from 1002 SAR images shows that the Grad-CAM achieved the highest domain alignment score (0.608 ± 0.074), making it the most suitable XAI method for operational marine oil spill detection systems specifically for C-band SAR imagery under moderate wind conditions (3–10 m/s). The proposed model also achieved segmentation accuracy (in terms of IoU) of 0.830 (83%) and validation accuracy of 90.5%. The proposed model is a highly specialized tool for look-alike discrimination in open-ocean scenarios rather than a total replacement for comprehensive multi-sensor maritime surveillance. The framework has a novel domain-specific metrics that indicate the explanation quality that could not be observed or explained by the traditional AI techniques. Furthermore, the framework promotes oil spill detection with improved accuracy and transparency. The Grad-CAM technique employed demonstrated trustworthiness and interpretability thus suitable for deployment as a decision support in oil spill detection system. This addresses the limitation of AI model as a black box especially in deployment for environmental monitoring.

The study supports the UN SDGs 6, 7, and 14 demonstrating the utilization of the XAI in bridging the gap between technological capability of AI models and operational trust especially for critical applications. The outcome of this study contributes empirically to the literature by providing an evidence-based guidelines for the deployment of trustworthy AI in environmental monitoring. Furthermore, this study bridges the gap between high-accuracy detection systems and stakeholder’s confidence in marine ecosystem protection. This study is limited to the use of Sentinel-1 SAR imagery. Secondly, the deep learning analysis was limited to 20 samples from a dataset of 1,002 images due to high algorithm latency, high computational cost and intensiveness. This may be insufficient to provide a general conclusion. Lastly, high computational latency requirement to generate all five post hoc XAI feature attributions simultaneously in real-time operational workflows was another limitation.

Based on these limitations, future studies can consider larger-scale evaluation extending beyond 20 samples, and validation on other sensor types and geographical regions. Future studies can also explore temporal consistency evaluation, human expert validation, and extend the application of the XAI techniques deployed in this study to multi-modal satellite imagery.

Correction note

02 September 2026 This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

16 September 2026 This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Viet Q. Vu, Thai Nguyen University of Technology, Vietnam

Reviewed by: Navya Prakash, Carl von Ossietzky University of Oldenburg, Germany

Abinaya K., Sathyabama University, India