School of Systems Science and Engineering, Sun Yat-sen University, Guangzhou, Guangdong, 510006, China.
School of Marine Sciences, Sun Yat-sen University, Zhuhai, Guangdong, 519082, China.
Mar Pollut Bull. 2024 Jun;203:116475. doi: 10.1016/j.marpolbul.2024.116475. Epub 2024 May 17.
As marine resources and transportation develop, oil spill incidents are increasing, endangering marine ecosystems and human lives. Rapidly and accurately identifying marine oil spill is of utmost importance in protecting marine ecosystems. Marine oil spill detection methods based on deep learning and computer vision have the great potential significantly enhance detection efficiency and accuracy, but their performance is often limited by the scarcity of real oil spill samples, posing a challenging to train a precise detection model. This study introduces a detection method specifically designed for scenarios with limited sample sizes. First, the small sample dataset of marine oil spill taken by Landsat-8 satellite is used as the training set. Then, a single image generative adversarial network (SinGAN) capable of training with a single oil spill image is constructed for expanding the dataset, generating diverse marine oil spill samples with different shapes. Second, a YOLO-v8 model is pretrained via the method of transfer learning and then trained with dataset before and after augmentation separately for real-time and efficient oil spill detection. Experimental results have demonstrated that the YOLO-v8 model, trained on an expanded dataset, exhibits notable enhancements in recall, precision, and average precision, with improvements of 12.3 %, 6.3 %, and 11.3 % respectively, compared to the unexpanded dataset. It reveals that our marine oil spill detection model based on YOLO-v8 exhibits leading or comparable performance in terms of recall, precision, and AP metrics. The data augmentation technique based on SinGAN contributes to the performance of other popular object detection algorithms as well.
随着海洋资源和运输的发展，石油泄漏事件不断增加，危及海洋生态系统和人类生命。快速准确地识别海洋石油泄漏对于保护海洋生态系统至关重要。基于深度学习和计算机视觉的海洋石油泄漏检测方法在提高检测效率和准确性方面具有巨大的潜力，但它们的性能往往受到真实石油泄漏样本稀缺的限制，这对训练精确的检测模型提出了挑战。本研究介绍了一种专门针对样本量有限情况的检测方法。首先，使用 Landsat-8 卫星拍摄的小样本数据集作为训练集。然后，构建了一个能够仅使用一张石油泄漏图像进行训练的单图像生成对抗网络（SinGAN），以扩展数据集，生成具有不同形状的多样化海洋石油泄漏样本。其次，通过迁移学习的方法对 YOLO-v8 模型进行预训练，然后分别使用扩充前后的数据集进行训练，以实现实时高效的石油泄漏检测。实验结果表明，在扩充数据集上训练的 YOLO-v8 模型在召回率、精度和平均精度方面均有显著提高，分别提高了 12.3%、6.3%和 11.3%，与未扩充数据集相比。这表明，我们基于 YOLO-v8 的海洋石油泄漏检测模型在召回率、精度和 AP 指标方面表现出色或具有竞争力。基于 SinGAN 的数据扩充技术对其他流行的目标检测算法的性能也有贡献。
