Scholarlink

<p>Oil spill detection in satellite imagery is of high significance for marine environment protection; however, it remains a very challenging task because of problems such as class imbalance and the visual similarities of oil slicks with natural phenomena, such as lookalikes or wave shadows. Considering these challenges, a custom convolutional neural networks (CNNs) framework empowered by explainable artificial intelligence (XAI) is proposed in this study for segmenting oil spills. Spatial attention mechanisms and weighted loss functions were employed in this model to improve the feature extraction of minority classes. A powerful data augmentation pipeline can be generalized across diverse real-world conditions. The proposed model achieved a test accuracy of 94.85% and an average recall of 95.00%, significantly outperforming state-of-the-art methods by 32.09%. The performance highlights included 91.00% recall for the category “Oil Spill” and 68.00% for the “Ship” category, demonstrating its strength in challenging segmentation tasks. Gradient-Weighted Class Activation Mapping (Grad-CAM) visualizations enhance transparency, making the model suitable for real-world applications. This study established a new benchmark for oil spill segmentation by combining accuracy with interpretability. Future work will focus on improving the minority class performance, increasing the number of datasets, and deploying maritime safety and environmental monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A convolutional neural network-based segmentation approach for oil spill detection in satellite imagery with explainable AI

  • 期刊论文

  • 发表日期：

    2025年8月16日

查看全文
  • Arjun Kumar Bose Arnob,
  • M. F. Mridha,
  • Meshal Alfarhood,
  • Dunren Che,
  • Mejdl Safran

摘要

Oil spill detection in satellite imagery is of high significance for marine environment protection; however, it remains a very challenging task because of problems such as class imbalance and the visual similarities of oil slicks with natural phenomena, such as lookalikes or wave shadows. Considering these challenges, a custom convolutional neural networks (CNNs) framework empowered by explainable artificial intelligence (XAI) is proposed in this study for segmenting oil spills. Spatial attention mechanisms and weighted loss functions were employed in this model to improve the feature extraction of minority classes. A powerful data augmentation pipeline can be generalized across diverse real-world conditions. The proposed model achieved a test accuracy of 94.85% and an average recall of 95.00%, significantly outperforming state-of-the-art methods by 32.09%. The performance highlights included 91.00% recall for the category “Oil Spill” and 68.00% for the “Ship” category, demonstrating its strength in challenging segmentation tasks. Gradient-Weighted Class Activation Mapping (Grad-CAM) visualizations enhance transparency, making the model suitable for real-world applications. This study established a new benchmark for oil spill segmentation by combining accuracy with interpretability. Future work will focus on improving the minority class performance, increasing the number of datasets, and deploying maritime safety and environmental monitoring.