Transfer learning and single-polarized SAR imagepreprocessingfor oilspill detection

ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 Available online 24 December 2024 2667-3932/© 2024 The Authors. Published by Elsevier B.V. on behalf of International Society of Photogrammetry and Remote Sensing. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). Contents lists available at ScienceDirect ISPRS Open Journal of Photogrammetry and Remote Sensing journal homepage: www.journals.elsevier.com/isprs-open-journal-of-photogrammetry-and-remote-sensing Transfer learning and single-polarized SAR image preprocessing for oil spill detection Nataliia Kussul a,b,c, Yevhenii Salii a,b,∗, Volodymyr Kuzin a,b, Bohdan Yailymov b, Andrii Shelestov a,b aNational Technical University of Ukraine ‘‘Igor Sikorsky Kyiv Polytechnic Institute’’, Department of Mathematical Modelling and Data Analysis, Kyiv, Ukraine bSpace Research Institute NAS Ukraine and SSA Ukraine, Department of Space Information Technologies and Systems, Kyiv, Ukraine cUniversity of Maryland, Department of Geographical Sciences, College Park, United States ARTICLE INFO Keywords: Oil spill detection Synthetic aperture radar (SAR) Deep learning Image preprocessing Transfer learning ABSTRACT This study addresses the challenge of oil spill detection using Synthetic Aperture Radar (SAR) satellite imagery, employing deep learning techniques to improve accuracy and efficiency. We investigated the effectiveness of various neural network architectures and encoders for this task, focusing on scenarios with limited training data. The research problem centered on enhancing feature extraction from single-channel SAR data to improve oil spill detection performance. Our methodology involved developing a novel preprocessing pipeline that converts single-channel SAR data into a three-channel RGB representation. The preprocessing technique normalizes SAR intensity values and encodes extracted features into RGB channels. Through an experiment, we have shown that a combination of the LinkNet with an EfficientNet-B4 is superior to pairs of other well-known architectures and encoders. Quantitative evaluation revealed a significant improvement in F1-score of 0.064 compared to traditional dB-scale preprocessing methods. Qualitative assessment on independent SAR scenes from the Mediterranean Sea demonstrated better detection capabilities, albeit with increased sensitivity to look-alike. We conclude that our proposed preprocessing technique shows promise for enhancing automatic oil spill segmentation from SAR imagery. The study contributes to advancing oil spill detection methods, with potential implications for environmental monitoring and marine ecosystem protection. 1. Introduction The Mediterranean Sea, a biodiversity hotspot hosting 11% of the world’s marine species in less than 1% of the global ocean area, faces increasing threats from pollution. Satellite imagery has revealed 757 oil slicks spanning 1.9 million hectares in the Mediterranean between 2020 and 2024 (SkyTruth,2024), primarily from illegal discharges by vessels. These incidents are particularly alarming as they impact key marine protected areas, including the Mediterranean Cetacean Migration Corridor and the Pelagos Sanctuary. Furthermore, several areas, such as those near Cyprus, remain under-assessed despite their vulnerability to pollution. Given the scale and number of these incidents, traditional oil spill detection methods, such as visual inspections and on-site sampling, are limited in scalability and weather dependency, Synthetic Aperture Radar (SAR) offers a promising alternative with its large area coverage, ∗Corresponding author at: Space Research Institute NAS Ukraine and SSA Ukraine, Department of Space Information Technologies and Systems, Kyiv, Ukraine. E-mail address: [email protected] (Y. Salii). all-weather, day-and-night imaging capabilities (Dong et al.,2023; Hernández-Hamón et al.,2023;Vrînceanu et al.,2023). However, accurate and timely detection of oil spills from SAR imagery remains challenging, particularly due to look-alike phenomena that can be mistaken for oil slicks and the general difficulty in distinguishing oil from similar water dark spots (Krestenitis et al.,2019;Chen et al., 2024;Wu et al.,2024b). Furthermore, a significant challenge in the deployment of data-driven methods for SAR analysis is the lack of sufficient labeled training data for many geographic regions, including our pilot territory near Cyprus. Advances in deep learning, particularly convolutional neural networks (CNNs), have become integral to improving oil spill detection from SAR imagery. These techniques have revolutionized the field by addressing the complexity of SAR data and the subtleties of oil spill signatures that traditional image processing methods often struggle with. https://doi.org/10.1016/j.ophoto.2024.100081 Received 1 July 2024; Received in revised form 15 December 2024; Accepted 16 December 2024 ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 2 N. Kussul et al. A variety of CNN architectures, including U-Net (de Moura et al.,2022; Mahmoud et al.,2022), LinkNet (Ding et al.,2020;Yan et al.,2023), DeepLabV3+(Kong et al.,2021;Li and Kong,2023), U-Net++ (Chen et al.,2023;Yu et al.,2023), MAnet (Wu et al.,2024a;Huang et al., 2023), FPN (Zhang et al.,2021;Sun et al.,2022), and PSPNet (Li et al.,2020;Erten et al.,2023), offer diverse approaches to image segmentation. The selection of an encoder backbone significantly impacts model performance, complementing the choice of architecture. Options like EfficientNet, MobileNetV3, and MobileOne offer various trade-offs between accuracy and computational efficiency (Li et al.,2021;Kolosov et al.,2022;Vasu et al.,2023). This balance is particularly crucial in oil spill detection, where rapid processing of large volumes of satellite data is essential for timely responses to environmental emergencies. In our study, we conduct an experiment through a comprehensive evaluation of state-of-the-art architectures and encoders to determine their effectiveness in the context of the application task. While deep learning models have shown impressive results, their performance heavily depends on the quality and representation of input data. This dependency highlights the critical role of preprocessing techniques for satellite images, which can significantly enhance model performance by accentuating relevant features and suppressing noise (Cristea et al.,2020;Hong et al.,2020;Lin et al.,2021). However, the optimal pipeline remains an open question because of the diversity in SAR image characteristics across different environmental conditions and geographic regions. Our study addresses these challenges within the framework of the HORIZON Europe iMERMAID project, which aims to develop innovative strategies for preventing, monitoring, and mitigating pollution in the Mediterranean Sea. Specifically, we propose an approach that combines advanced preprocessing techniques with the utilization of the best architecture-encoder pair, determined through a comprehensive experiment. By leveraging transfer learning, we mitigate the data scarcity issue by fine-tuning models pre-trained on large datasets to adapt them to the characteristics of our study region. Our contributions and their impacts are as follows: •A comparative evaluation of various CNN architectures and encoder backbones for oil spill detection, identifying effective combinations for this task. •A novel preprocessing technique that normalizes SAR intensity values and encodes extracted features into RGB channels, leading to enhancement of relevant features of oil spills •An experimental evaluation using a dataset spanning multiple geographic regions, demonstrating the model’s ability to generalize to unseen data from the Mediterranean Sea These contributions not only advance the technical field of oil spill detection but also directly support the broader goals of the iMERMAID project in protecting the Mediterranean Sea from chemical pollution. 2. Data and methods In this section, we detail the comprehensive process involved in detecting oil spills from satellite imagery using deep learning models. The entire workflow, from data acquisition to the training of segmentation models, is illustrated in Fig. 1. 2.1. CleanSeaNet CleanSeaNet (Carpenter,2015), a European satellite-based oil spill detection service, represents a cornerstone in maritime environmental monitoring. Operated by the European Maritime Safety Agency (EMSA), this system leverages advanced SAR technology to provide near-real-time detection of potential oil spills across European waters. The service’s strength lies in its ability to capture high-resolution imagery regardless of cloud cover or daylight conditions, making it an invaluable tool for continuous maritime surveillance. CleanSeaNet’s detection capabilities extend beyond mere identification, offering crucial details such as spill location, areal extent, and confidence assessments of detected anomalies. Despite its sophisticated capabilities, the publicly accessible data from CleanSeaNet for the Mediterranean region presents limitations for research purposes. The available information is restricted to annual point data, which lacks the temporal linkage needed to fit machine or deep learning models. This constraint poses a significant challenge for researchers aiming to develop and train high-precision oil spill detection models specific to the Mediterranean Sea. 2.2. Marine Pollution Surveillance Program The Marine Pollution Surveillance Program (Shazif,2022), operated by the NOAA/NESDIS Satellite Analysis Branch (SAB), stands as a pivotal resource in the monitoring and detection of oil spills in U.S. waters. Operational since late 2010, this program has evolved into a robust, round-the-clock surveillance system, utilizing a combination of SAR and multi-spectral satellite imagery. The program’s primary mission extends beyond mere detection, aiming to provide comprehensive analyses of oil spill incidents. These analyses include crucial information such as the location of oil slicks, their spatial extent, and, when possible, assessments of relative oil thickness. This detailed approach supports rapid response efforts coordinated by the National Ocean Service Emergency Response Division. A key strength of the Marine Pollution Surveillance Program lies in its extensive historical dataset. The program offers openly accessible vector data (NOAA,2022) of oil spills from 2011 to 2024, covering a vast array of water bodies including the Gulf of Mexico, the Pacific Ocean, the Atlantic Ocean, and the Great Lakes. This rich temporal and spatial coverage provides an invaluable resource for developing and refining oil spill detection algorithms. The dataset’s comprehensive nature, encompassing various marine environments and conditions, makes it particularly suitable for training machine learning models. It offers researchers the opportunity to develop robust algorithms capable of identifying oil spills across diverse scenarios and environmental conditions. However, it is important to note that while this dataset is extensive for U.S. waters, it does not include data from the Mediterranean Sea. This geographical limitation presents both a challenge and an opportunity for researchers. The challenge lies in the direct applicability of models trained on this data to Mediterranean contexts. Conversely, it opens avenues for exploring transfer learning techniques, where models trained on the rich U.S. dataset can be adapted and fine-tuned for application in Mediterranean waters, potentially bridging the data gap in this region. 2.3. Sentinel-1 SAR data The backbone of our study is Sentinel-1 SAR data, chosen for its consistent quality and global coverage. For our analysis, we specifically utilized the Ground Range Detected (GRD) products from Sentinel-1, which offer a spatial resolution of 10 m. These products are pre-processed to remove thermal noise and provide terrain-corrected data in ground range geometry. We focused on the VV (Vertical–Vertical) polarization channel, as it has been shown to be particularly effective in distinguishing oil spills from the surrounding water surface. The Sentinel-1 data acquisition process was twofold, tailored to our specific study areas. For the images corresponding to the Marine Pollution Surveillance Reports (MPSR), we utilized the NASA Earth Data platform with the Alaska Satellite Facility (ASF). This approach allowed us to precisely target and download the relevant SAR images for the reported oil spill events in U.S. waters. ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 3 N. Kussul et al. Fig. 1. Pipeline illustrating the steps from data acquisition to segmentation model training and evaluation. In contrast, we used the Google Earth Engine platform for the Mediterranean study area. This cloud-based solution allowed us to efficiently access and process Sentinel-1 imagery covering our area of interest between Cyprus and Anatolia. Using the cloud platform was particularly advantageous for handling the large amount of data required for comprehensive coverage of the Mediterranean region. We applied additional pre-processing steps to the Sentinel-1 imagery to ensure optimal data quality. These included speckle filtering to reduce inherent SAR noise and radiometric calibration to ensure consistency between acquisition dates and sensor modes. 2.4. Dataset Our dataset was meticulously compiled from Marine Pollution Surveillance Program reports, focusing specifically on Sentinel-1 data from 2018 to 2023. After careful curation, which involved excluding reports without binary segmentation, we assembled a collection of 300 distinct areas of interest. Thus, the resulting dataset comprises 300 VV-polarized SAR images, each with dimensions of 640x640 pixels and a spatial resolution of 10 m. To ensure robust model evaluation, we reserved all 58 areas from the year 2021 for testing purposes. The remaining 242 areas were randomly divided into training and validation subsets, following an 80:20 ratio respectively. To complement this primary dataset and facilitate qualitative assessment within the iMermaid project framework, we constructed an additional one. This supplementary set consists of 114 unlabeled SAR images, each measuring 6275 ×3489 pixels. These images cover an expansive area of 176,000 hectares in the Mediterranean Sea, specifically the region between Cyprus and Anatolia. According to CleanSeaNet records, this area experienced 32 oil spill events during 2022, making it a valuable resource for evaluating model performance in a real-world, Mediterranean context 2. This comprehensive dataset structure allows for thorough model training, validation, and testing, while also providing a means to assess the model’s transferability to the Mediterranean region, which is crucial for the broader objectives of the iMermaid project. 2.5. Pretrained models Given the limited size of our oil spill dataset, we adopted a transfer learning approach to leverage knowledge from models pretrained on large-scale datasets. This strategy allows us to benefit from features learned on diverse image data, potentially improving our model’s ability to detect oil spills in SAR imagery. We utilized models pretrained on the ImageNet dataset, a database of over 14 million images across 20,000 categories. The ImageNet pretraining provides our models with a robust set of low-level features and high-level semantic understanding, which can be advantageous even when transferring to a domain as specialized as SAR imagery. For our experiments, we explored a range of modern architectures known for their effectiveness in image segmentation tasks. These Fig. 2. A pilot area in the Mediterranean for testing an oil spills detection model. include U-Net, LinkNet, DeepLabV3+, U-Net++, DeepLabV3, MAnet, FPN, and PSPNet. Each of these architectures offers unique strengths in handling multi-scale features and preserving spatial information, which are crucial for accurate oil spill delineation. To complement these architectures, we investigated several encoder backbones, specifically efficientnet-b4, timm-mobilenetv3_small_100 and mobileone_s4. These encoders were chosen for their balance of accuracy, computational efficiency and size, which is particularly important for potential real-time applications of oil spill detection. Through this transfer learning approach, we aim to leverage the power of large-scale pretraining while fine-tuning the models to the specific challenges of oil spill detection in SAR imagery. This strategy allows us to potentially achieve higher accuracy and faster convergence than training from scratch, even with our relatively small domain-specific dataset. 2.6. Computing infrastructure The training of our deep learning models was performed on a dedicated instance within the CreoDIAS platform (Kuzin et al.,2022). CreoDIAS is a cloud-based infrastructure that provides access to Copernicus satellite data and computing resources, enabling efficient processing and analysis of large-scale Earth observation datasets. For our experiments, we utilized a custom virtual machine instance specifically configured for deep learning tasks. The instance, designated as ‘‘vm.a6000.2’’, was equipped with the following specifications: ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 4 N. Kussul et al. •Virtual GPU: NVIDIA RTX A6000 with 12 GB VRAM •RAM: 28 GB •vCPUs: 4 •Storage: 80 GB SSD This high-performance configuration, particularly the NVIDIA RTX A6000 GPU, allowed us to efficiently train and evaluate our deep learning models on the mentioned Sentinel-1 dataset. The ample GPU memory was crucial for handling the complex neural network architectures and relatively large batch sizes used in our experiments. 2.7. Metrics In evaluating the performance of our oil spill detection model, we focused on several key metrics that are particularly well-suited for assessing segmentation tasks in imbalanced datasets, which is characteristic of oil spill detection in vast ocean environments. The primary metrics we employed were the Intersection over Union (IoU), also known as the Jaccard index, and the F1 score, which is equivalent to the Dice coefficient. These metrics were chosen for their ability to provide a comprehensive assessment of our model’s accuracy in identifying oil spills against the predominant background of clean water. The IoU metric measures the overlap between the predicted oil spill segmentation and the ground truth, calculated as the ratio of the intersection to the union of the predicted and actual oil spill areas. It ranges from 0 to 1, with 1 indicating perfect overlap. The IoU is particularly valuable in our context as it penalizes both false positives and false negatives, providing a strict measure of segmentation accuracy. The F1 score, calculated as the harmonic mean of precision and recall, offers a balanced measure of the model’s performance. In the context of oil spill detection, where the positive class (oil spills) is significantly smaller than the negative class (clean water), the F1 score helps in assessing how well the model identifies oil spills without being overly influenced by the large number of true negatives. Additionally, we utilized the Dice loss function during model training. The Dice loss, being the complement of the Dice coefficient, served as an effective optimization target. It helped in addressing the class imbalance issue inherent in our dataset by focusing the model’s attention on correctly identifying the relatively small oil spill areas. 2.8. Data preparation The use of pretrained models, however, comes with a notable challenge: ImageNet-trained models are designed to process color images composed of three channels (red, green, and blue) whereas our application involves using only single-channel VV-polarized images. To address this discrepancy, we adapted our approach by utilizing conversation from VV to RGB, which can be done in various ways, but in the scope of our research, we focus on 2 of them. 2.8.1. Original approach The established technique involves transforming calibrated satellite imagery (𝑉 𝑉) to a logarithmic decibel (𝑑 𝐵) scale: 𝑑 𝐵= 20 ⋅log10 𝑉 𝑉 This conversion enhances the visibility of relative differences in backscatter intensity (Blondeau-Patissier et al.,2023). It is particularly effective for highlighting oil spill pixels in satellite images while ensuring they remain darker than surrounding ocean and land areas. Oil spills in SAR imagery typically have a limited range of intensity values, making them challenging to distinguish from other surfaces. The dB scale addresses this by expanding the lower intensity range associated with oil slicks while compressing higher intensity values from sea and terrain. This transformation amplifies subtle contrasts between oil slicks and their surroundings, maintaining the characteristic darker appearance of oil spills in SAR imagery due to their smoother surface texture. To represent the 𝑑 𝐵values as color components, we apply min– max normalization to scale them to the range [0,255], suitable for RGB channels. 2.8.2. Proposed approach Our novel preprocessing approach aims to enhance the differentiation of oil spills by assigning distinct information to each color channel, thereby emphasizing various aspects of oil slicks while minimizing the influence of confounding factors. Initially, we convert the VV-polarized SAR data to the decibel scale, similar to the original approach and apply min–max scaling: 𝑟= log10 𝑉 𝑉 However, in cases where no oil slicks are present in the image, min–max scaling can be expected to create a significant number of look-alikes, as the spread of water brightness values will be relatively small. To minimize look-alike recognition caused by red channel normalization, it was decided to convert the VV channel to a normal distribution. This conversion works on the hypothesis that the vast majority of pixels in the image belong to the water class, which is true for both training data and real-world images. Since after conversion the values are not limited, but most are close to 0, we used normalization based on the arctangent function. 𝑔=1 𝜋ar ct an (𝑉 𝑉−𝑚𝑒𝑎𝑛(𝑉 𝑉) 𝑠𝑡𝑑(𝑉 𝑉))+ 0.5 One of the features of the training data is that some images have significant intra-image variability in water surface brightness. In these cases, the spatial brightness of the water forms a nonlinear surface with a low frequency of variation from the mean. In many cases, such images have large dark regions that could be mistaken for oil spills. To mitigate this issue, we propose a local brightness equalization technique that tries to approximate the spatial nonlinearity of water brightness: 𝐸𝑃=𝑝𝑜𝑜𝑙 𝑖𝑛𝑔𝑎𝑣𝑔 32 (𝑉 𝑉)𝐷𝑃=𝑝𝑜𝑜𝑙 𝑖𝑛𝑔𝑠𝑡𝑑 32 (𝑉 𝑉) 𝐸𝑃and 𝐷𝑃are computed using average pooling and pooled standard deviation, respectively, of VV values with a kernel size and a stride of 32. This results in a first approximation of the desired surface. After this operation, the dimensions of the input image are reduced by a factor of 32 in both axes. 𝐸𝐵=𝑏𝑙 𝑢𝑟𝑔 𝑎𝑢𝑠𝑠𝑖𝑎𝑛 7(𝐸𝑃)𝐷𝐵=𝑏𝑙 𝑢𝑟𝑔 𝑎𝑢𝑠𝑠𝑖𝑎𝑛 7(𝐷𝑃) The pooled values are then smoothed using Gaussian blur with a kernel size of 7. Smoothing reduces the impact of outliers caused by the possible presence of oil spills in pooled pixels. 𝐸𝑅=𝑟𝑒𝑠𝑖𝑧𝑒𝑙 𝑖𝑛𝑒𝑎𝑟 𝑉 𝑉(𝐸𝐵)𝐷𝑅=𝑟𝑒𝑠𝑖𝑧𝑒𝑙 𝑖𝑛𝑒𝑎𝑟 𝑉 𝑉(𝐷𝐵) Then we achieve the final approximation by resizing the smoothed values to the original VV image dimensions using linear interpolation. 𝑏∗=(𝑉 𝑉−𝐸𝑅)∕𝐷𝑅 Finally, we obtain the blue channel by normalizing VV values based on approximated local water brightness (𝐸𝑅) and local standard deviation (𝐷𝑅). For scaling we use arctangent function: 𝑏=1 𝜋ar ct an (𝑏∗)+ 0.5 ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 5 N. Kussul et al. Table 1 Quantitative results of trained models sorted by weighted f1 score. Architecture Encoder Batch Training time Train Validation Test Weighted size (minutes) F1 F1 F1 F1 LinkNet efficientnet-b4 4 25 0.774 0.703 0.765 0.746 DeepLabV3Plus timm-mobilenetv3_small_100 16 12 0.765 0.676 0.777 0.741 MAnet timm-mobilenetv3_small_100 16 5 0.780 0.642 0.787 0.737 FPN efficientnet-b4 4 33 0.777 0.658 0.768 0.733 LinkNet timm-mobilenetv3_small_100 16 6 0.693 0.675 0.785 0.728 UnetPlusPlus efficientnet-b4 2 35 0.758 0.670 0.753 0.726 PSPNet mobileone_s4 4 18 0.719 0.672 0.750 0.717 UnetPlusPlus timm-mobilenetv3_small_100 12 8 0.731 0.649 0.741 0.708 Unet efficientnet-b4 4 26 0.705 0.679 0.726 0.706 DeepLabV3 timm-mobilenetv3_small_100 12 19 0.768 0.557 0.778 0.702 FPN timm-mobilenetv3_small_100 16 8 0.731 0.614 0.740 0.696 LinkNet mobileone_s4 4 33 0.667 0.710 0.690 0.691 Unet mobileone_s4 4 32 0.703 0.696 0.669 0.686 PSPNet efficientnet-b4 4 13 0.723 0.630 0.692 0.678 Unet timm-mobilenetv3_small_100 16 6 0.746 0.637 0.655 0.669 MAnet efficientnet-b4 4 24 0.690 0.633 0.681 0.667 FPN mobileone_s4 4 14 0.650 0.653 0.629 0.641 PSPNet timm-mobilenetv3_small_100 16 3 0.588 0.584 0.642 0.611 UnetPlusPlus mobileone_s4 2 45 0.375 0.693 0.557 0.562 DeepLabV3Plus efficientnet-b4 2 61 0.572 0.620 0.498 0.555 MAnet mobileone_s4 4 23 0.496 0.534 0.541 0.529 Prior to assigning the obtained r, g, b values to the R, G, B color channels, we apply min–max scaling to r to fit the range [0, 255], while g and b are simply multiplied by 255. This preprocessing pipeline aims to enhance the visibility of oil spills while normalizing for both global and local brightness variations, improving the model’s ability to distinguish between oil spills and other dark surfaces in satellite images. 2.9. Experiment I: Architecture and encoder selection First of all, we want to know which combination of architecture and encoder is best suited for building a model to solve the oil slick detection problem. We trained and evaluated 24 distinct models, comprising combinations of 8 architectures (DeepLabV3, DeepLabV3+, FPN, LinkNet, MAnet, PSPNet, Unet, Unet++) and 3 encoders (EfficientNet-B4, MobileOne-S4, Timm-MobileNetV3-Small-100). This allows for a thorough assessment of various state-of-the-art segmentation architectures and efficient encoder backbones. All models will be trained with a learning rate of 10−3. Batch size will vary due to different requirements for RAM of each model, but will not exceed 16. Training time is limited to 300 epochs, however, if the model is not improving on validation for 10 epochs then training will be stopped earlier. All models in this experiment utilized the original VV to RGB conversion approach as described in Section 2.8.1 2.10. Experiment II: Model refinement and Mediterranean Sea evaluation Building upon the results of Experiment I, our second experiment aims to maximize model performance and assess generalization capabilities to the Mediterranean Sea context. We focused on the most effective architecture-encoder combination identified in the first experiment. Two variants of this model were trained: one using the original VV to RGB conversion approach, and another using the proposed preprocessing approach described in Section 2.8.2 For this fine-tuning phase, we reduced the learning rate to 10−4 while keeping the other training parameters the same as in Experiment I. This allows a more precise adjustment of the model weights. 3. Results 3.1. Experiment I Quantitative performance of trained models alongside with training time and batch size can be seen in Table 1. Due to the high spread of score between validation and test, we decided to add weighted average score. This score was constructed by taking train, validation and test scores in a ratio of 2:3:4. Such ratio was chosen to reward high test scores and punish low validation while taking into account the influence of corresponding subsets to the training process. From Table 1we can see that only 21 models was trained while in Section 2.9 was stated that 24 models would be trained. This mismatch was caused due to the fact that DeepLabV3 with efficientnet-b4 and mobileone_s4 encoder was unable to fit into available VRAM with a batch size of 2. The training of the last missing model (DeepLabV3Plus with mobileone_s4 encoder) was canceled due to poor performance and the long training time of the model with efficientnet-b4 encoder. As shown in Table 1, the best results were achieved by the LinkNet and FPN models with the EfficientNet-B4 encoder, as well as by DeepLabV3Plus and MAnet with the timm-mobilenetv3_small_100 encoder. However, while these models performed similarly on the training and test sets, only the LinkNet model demonstrated good results on the validation set. DeepLabV3Plus and MAnet exhibited interesting behavior, performing better with smaller encoders. For instance, both models achieved a weighted score of approximately 0.74 when using the timmmobilenetv3_small_100 encoder, which has 0.93 million parameters. In contrast, their performance dropped to 0.67 when utilizing the EfficientNet-B4 encoder, which has 17 million parameters. The overall conclusion from this experiment is that the combination of LinkNet and EfficientNet-B4 shows the highest potential for effectively detecting oil spills. 3.2. Experiment II Since, according to the results of the previous experiment, the best performance was shown by LinkNet in combination with the efficient-b4 encoder, we will use it for training. First of all, it is worth noting that the results obtained for the original approach (Table 2) show that reducing the learning rate to 10−4helped to significantly improve the model’s performance. ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 6 N. Kussul et al. Fig. 3. Segmentation of Mediterranean test site using the original approach of conversion from VV to RGB. (a) - RGB image; (b) - segmentation result. Fig. 4. Segmentation of Mediterranean test site using the proposed approach of conversion from VV to RGB. (a) - RGB image; (b) - segmentation result. Table 2 Quantitative results of trained LinkNet models with efficientnet-b4 encoder using different VV to RGB approaches. Approach Batch Training time Train Validation Test Weighted size (minutes) F1 F1 F1 F1 Proposed 4 55 0.875 0.748 0.831 0.813 Original 4 52 0.850 0.728 0.767 0.773 A comparison of the results using the original approach from this experiment with those from the previous one (Table 1) clearly shows an increase in both training (by 0.076) and validation (by 0.025) scores. However, the increase in test scores is negligible. This suggests that the model may have become more prone to overfitting. Regarding the proposed approach, the data in Table 2indicate a significant improvement in model accuracy across all cases. For instance, the new approach enhanced the performance on test by 0.064, and on training and validation by 0.025 and 0.020, respectively. This indicates a better generalization ability of the model with the new approach. After training the models, we applied them to data from the Mediterranean Sea. As can be seen from Fig. 3, the model using the original approach confidently identifies the main part of the oil slick, but misses the thinner tail of it. Additionally, it interprets a dark patch to the right of the slick as oil. The model using the proposed approach, as seen in Fig. 4, more confidently recognized the main part of the slick and also highlighted part of its tail. This model also highlighted a dark patch to the right of the slick as oil, increasing the likelihood that it was, although the lack of ground truth prevents us from verifying this. However, the model is much more sensitive, which improves its ability to detect oil spills, but leads to an increased number of false positives, especially near the no-data regions. Investigation into the reasons for this behavior shows that the most likely cause is the blue channel of the preprocessed image. Due to the use of convolutions, it is sensitive to how no-data is processed. In our case, the no-data value is replaced by the maximum valid value of VV, leading to an overestimation of background water brightness near nodata regions. A solution to this problem could be an adaptive selection of the value to fill no-data instead of using maximum. The appearance of other false positives can be explained by a small value of local standard deviation leading to dark spots appearing in regions with mostly constant water brightness. A possible solution to this problem would be a limit on the minimum value of the local standard deviation. Our experiments with varying these factors have confirmed that this can reduce the number of false positives. However, finding the balance between detection capabilities and false positive rate by these factors requires a separate investigation. 4. Discussion This study has confirmed the effectiveness of using pretrained models on ImageNet for segmentation tasks in other domains. The conducted experiments demonstrated that among the architectures and encoders examined, the LinkNet model with an EfficientNet-B4 encoder produced the best results for detecting oil spills based on VV-polarized SAR imagery. An unexpected finding was that the DeepLabV3+and MAnet models performed better with smaller encoders. Considering their training speed and inference efficiency, these models could potentially be used ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 7 N. Kussul et al. for preliminary and/or ensemble segmentation methods. However, this possibility was not explored within the scope of this work and warrants further investigation. The experiment involving different approaches to converting VV images to RGB showed that engineering within this step can significantly enhance model performance. Our proposed method for this conversion significantly improved the model’s quantitative metrics. When the model was transferred to a different geographical area, the model show significantly better capability to detect thinner parts of oil spills, however it also exhibited a higher level of false positives. Our experiments outside the article’s scope show that replacing no-data values with the mean value of valid pixels instead of the maximum leads to solving the problem of false positives at the border with no-data. Limiting the minimal value of local standard deviation in blue channel calculation helps decrease false positive rates. However, finding optimal values requires separate studies. Apart from the no-data cases, the only known limitation of the proposed approach lies in coastal areas with a significant amount of land, where, due to the peculiarities of the calculations, the 𝐵and 𝐺channels of the preprocessed image tend to be darkened or brightened, thus becoming less informative. This limitation can be partially overcome by masking land with no-data values. 5. Conclusion This study presents a significant step forward in the application of deep learning for oil spill detection using Sentinel-1 VV-polarized SAR data. Our comprehensive evaluation of various architectures and encoders identified the LinkNet architecture with an EfficientNet-B4 encoder as a particularly effective combination for this task, especially in scenarios with limited training data. A key contribution of this work is our novel preprocessing approach for converting SAR data to RGB representations. This method demonstrated notable improvements, increasing the F1 score by 0.064 on our test dataset compared to the traditional dB-scale transformation. The transferability of our model from U.S. coastal waters to the Mediterranean Sea showed promising results. The model demonstrated an improved ability to detect oil spills in a new geographic context, but the model’s higher sensitivity increases the number of false positives. The results obtained within the framework of VV to RGB conversion engineering can be used to solve other SAR-based segmentation tasks. And with some modifications, it can be adapted to other types of satellite data. This work aligns well with the objectives of the HORIZON Europe iMERMAID project, offering potential enhancements to monitoring capabilities for the Mediterranean ecosystem. The findings open up new avenues for research in SAR image preprocessing and segmentation, with potential applications extending beyond oil spill detection to other areas of remote sensing. CRediT authorship contribution statement Nataliia Kussul: Writing – review & editing, Supervision, Project administration, Funding acquisition. Yevhenii Salii: Writing – original draft, Validation, Software, Methodology, Investigation, Conceptualization. Volodymyr Kuzin: Writing – review & editing, Writing – original draft, Methodology, Data curation. Bohdan Yailymov: Writing – original draft, Data curation. Andrii Shelestov: Supervision. Declaration of Generative AI and AI-assisted technologies in the writing process Statement: During the preparation of this work, the authors used ChatGPT service, in order to assist with the drafting and refinement of certain sections of the manuscript. After using this tool, the authors thoroughly reviewed and edited the content as needed and take full responsibility for the content of the publication. The use of this AI tool was solely for language refinement and organization purposes, and did not contribute to the scientific content, analysis, or conclusions of the study. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments The study is supported by the European Commission through the HORIZON Europe project iMERMAID ‘‘Innovative solutions for Mediterranean Ecosystem Remediation via Monitoring and Decontamination from Chemical Pollution’’ (101112824). (https://imermaid.eu/). Data availability Data are available on request, please get in touch with e-mail: [email protected]. References Blondeau-Patissier, D., Schroeder, T., Suresh, G., Li, Z., Diakogiannis, F.I., Irving, P., Witte, C., Steven, A.D., 2023. Detection of marine oil-like features in sentinel-1 SAR images by supplementary use of deep learning and empirical methods: Performance assessment for the Great Barrier Reef marine park. Marine Poll. Bull. 188, 114598. http://dx.doi.org/10.1016/j.marpolbul.2023.114598. Carpenter, A., 2015. European maritime safety agency CleanSeaNet activities in the north sea. In: Oil Pollution in the North Sea. Springer International Publishing, pp. 33–47. http://dx.doi.org/10.1007/698_2015_429. Chen, Y.T., Chang, L., Wang, J.H., 2024. Full-scale aggregated MobileUNet: An improved U-net architecture for SAR oil spill detection. Sensors 24 (12), 3724. http://dx.doi.org/10.3390/s24123724. Chen, S., Lei, F., Zang, Z., Zhang, M., 2023. Forest mapping using a VGG16-UNet++&; stacking model based on google earth engine in the urban area. IEEE Geosci. Remote Sens. Lett. 20, 1–5. http://dx.doi.org/10.1109/lgrs.2023.3274128. Cristea, A., van Houtte, J., Doulgeris, A.P., 2020. Integrating incidence angle dependencies into the clustering-based segmentation of SAR images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 13, 2925–2939. http://dx.doi.org/10.1109/jstars. 2020.2993067. de Moura, N.V.A., de Carvalho, O.L.F., Gomes, R.A.T., Guimarães, R.F., de Carvalho Júnior, O.A., 2022. Deep-water oil-spill monitoring and recurrence analysis in the Brazilian territory using sentinel-1 time series and deep learning. Int. J. Appl. Earth Obs. Geoinf. 107, 102695. http://dx.doi.org/10.1016/j.jag.2022.102695. Ding, Y., Wu, M., Xu, Y., Duan, S., 2020. P-linknet: Linknet with spatial pyramid pooling for high-resolution satellite imagery. Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci. XLIII-B3-2020, 35–40. http://dx.doi.org/10.5194/isprs-archives-xliii-b32020-35-2020. Dong, X., Li, J., Li, B., Jin, Y., Miao, S., 2023. Marine oil spill detection from low-quality SAR remote sensing images. J. Mar. Sci. Eng. 11 (8), 1552. http: //dx.doi.org/10.3390/jmse11081552. Erten, H., Bostanci, E., Acici, K., Guzel, M.S., Asuroglu, T., Aydin, A., 2023. Semantic segmentation with high-resolution sentinel-1 SAR data. Appl. Sci. 13 (10), 6025. http://dx.doi.org/10.3390/app13106025. Hernández-Hamón, H., Ramírez, P.Z., Zaraza, M., Micallef, A., 2023. Google earth engine app using sentinel 1 SAR and deep learning for ocean seep methane detection and monitoring. Remote Sens. Appl. Soc. Environ. 32, 101036. http: //dx.doi.org/10.1016/j.rsase.2023.101036. Hong, S.J., Baek, W.K., Jung, H.S., 2020. Ship detection from X-Band SAR images using M2det deep learning model. Appl. Sci. 10 (21), 7751. http://dx.doi.org/10.3390/ app10217751. Huang, L., Tian, Q., Tang, B.-H., Le, W., Wang, M., Ma, X., 2023. Siam-EMNet: A siamese EfficientNet–manet network for building change detection in very high resolution images. Remote Sens. 15 (16), 3972. http://dx.doi.org/10.3390/ rs15163972. Kolosov, D., Kelefouras, V., Kourtessis, P., Mporas, I., 2022. Anatomy of deep learning image classification and object detection on commercial edge devices: A case study on face mask detection. IEEE Access 10, 109167–109186. http://dx.doi.org/10. 1109/access.2022.3214214. Kong, Y., Liu, Y., Yan, B., Leung, H., Peng, X., 2021. A novel Deeplabv3+ network for SAR imagery semantic segmentation based on the potential energy loss function of gibbs distribution. Remote Sens. 13 (3), 454. http://dx.doi.org/10.3390/ rs13030454. Krestenitis, M., Orfanidis, G., Ioannidis, K., Avgerinakis, K., Vrochidis, S., Kompatsiaris, I., 2019. Oil spill identification from satellite images using deep neural networks. Remote Sens. 11 (15), 1762. http://dx.doi.org/10.3390/rs11151762. ISPRS Open Journal of Photogrammetry and Remote Sensing 15 (2025) 100081 8 N. Kussul et al. Kuzin, V., Musial, J., Shelestov, A., 2022. EO4UA initiative: Scientific European support of ukrainian scientific community. In: 2022 12th International Conference on Dependable Systems, Services and Technologies. DESSERT, IEEE, http://dx.doi.org/ 10.1109/dessert58054.2022.10018706. Li, Q., Kong, Y., 2023. An improved SAR image semantic segmentation Deeplabv3+ network based on the feature post-processing module. Remote Sens. 15 (8), 2153. http://dx.doi.org/10.3390/rs15082153. Li, S., Tan, M., Pang, R., Li, A., Cheng, L., Le, Q.V., Jouppi, N.P., 2021. Searching for fast model families on datacenter accelerators. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. CVPR, pp. 8085–8095. Li, J., Xiu, J., Yang, Z., Liu, C., 2020. Dual path attention net for remote sensing semantic image segmentation. ISPRS Int. J. Geo-Inf. 9 (10), 571. http://dx.doi. org/10.3390/ijgi9100571. Lin, S.Y., Lin, C.W., van Gasselt, S., 2021. Processing framework for landslide detection based on synthetic aperture radar (SAR) intensity-image analysis. Remote Sens. 13 (4), 644. http://dx.doi.org/10.3390/rs13040644. Mahmoud, A.S., Mohamed, S.A., El-Khoriby, R.A., AbdelSalam, H.M., El-Khodary, I.A., 2022. Oil spill identification based on dual attention UNet model using synthetic aperture radar images. J. Indian Soc. Remote Sens. 51 (1), 121–133. http://dx.doi. org/10.1007/s12524-022-01624-6. NOAA, 2022. Marine pollution surveillance reports. URL: https://ospo.noaa.gov/ Products/ocean/marinepollution/2022_archive.html. Shazif, M., 2022. Oil spill clean-up methods; state of the art. SkyTruth, 2024. Mediterranean Sea Chronic Oil Pollution Analysis. Technical Report, SkyTruth, URL: . Analysis of oil slicks from July 2020 to January 2024. Sun, Z., Meng, C., Cheng, J., Zhang, Z., Chang, S., 2022. A multi-scale feature pyramid network for detection and instance segmentation of marine ships in SAR images. Remote Sens. 14 (24), 6312. http://dx.doi.org/10.3390/rs14246312. Vasu, P.K.A., Gabriel, J., Zhu, J., Tuzel, O., Ranjan, A., 2023. MobileOne: An improved one millisecond mobile backbone. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. CVPR, pp. 7907–7917. Vrînceanu, C.A., Grebby, S., Marsh, S., 2023. The performance of speckle filters on copernicus sentinel-1 SAR images containing natural oil slicks. Q. J. Eng. Geol. Hydrogeol. 56 (3), http://dx.doi.org/10.1144/qjegh2022-046. Wu, K., Cai, F., Wang, H., 2024a. CTMANet: A CNN-transformer hybrid semantic segmentation network for fine-grained airport extraction in complex SAR scenes. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 17, 4689–4704. http://dx.doi.org/ 10.1109/jstars.2024.3361927. Wu, W., Wong, M.S., Yu, X., Shi, G., Kwok, C.Y.T., Zou, K., 2024b. Compositional oil spill detection based on object detector and adapted segment anything model from SAR images. IEEE Geosci. Remote Sens. Lett. 21, 1–5. http://dx.doi.org/10.1109/ lgrs.2024.3382970. Yan, Q., Chen, Y., Jin, S., Liu, S., Jia, Y., Zhen, Y., Chen, T., Huang, W., 2023. Inland water mapping based on GA-LinkNet from CyGNSS data. IEEE Geosci. Remote Sens. Lett. 20, 1–5. http://dx.doi.org/10.1109/lgrs.2022.3227596. Yu, J., He, X., Yang, P., Motagh, M., Xu, J., Xiong, J., 2023. Coastal aquaculture extraction using GF-3 fully polarimetric SAR imagery: A framework integrating UNet++ with marker-controlled watershed segmentation. Remote Sens. 15 (9), 2246. http://dx.doi.org/10.3390/rs15092246. Zhang, T., Zhang, X., Ke, X., 2021. Quad-FPN: A novel quad feature pyramid network for SAR ship detection. Remote Sens. 13 (14), 2771. http://dx.doi.org/10.3390/ rs13142771.