Study On Detection Method Of Marine Oil Spill Areas Based On SAR Image

With the development of technology and industrial progress,China’s shipping industry and offshore oil extraction projects have also developed rapidly.However,at the same time,oil spill accidents caused by ship accidents and operational errors in offshore oil development and transportation are relatively frequent.Oil spills floating on the sea seriously threaten maritime traffic safety and marine ecological environment.Synthetic Aperture Radar(SAR)is one of the main technologies for marine environmental monitoring.By processing and analyzing the image information captured by SAR,it can effectively improve the accuracy of offshore oil spill detection and recognition and emergency response capabilities.With the continuous development and application of SAR systems in marine environmental monitoring,SAR image data has not only gained rich accumulation,but also contains more polarization scattering information,which is conducive to the extraction and detection of offshore oil spill areas.However,there are still differences in image information and performance parameters obtained by different SAR systems for SAR oil spill area detection that need to be studied.There is insufficient research on the correlation between SAR image oil spill area detection and the physical characteristics of the oil spill itself,and the accuracy of oil spill detection and recognition needs to be further improved.This article focuses on the characteristics of high noise,blurred boundaries,and uneven intensity in SAR images.Based on deep learning algorithms,oil spill area recognition and segmentation work is carried out on image information collected by different SAR systems;Based on the physical characteristics of oil spills and the prediction model of oil particle drift diffusion,a reference deployment method for oil particles is constructed.The main research work of this article is as follows:(1)A Dual Attention Encoding Network(DAENet)is proposed to address the issue of differences in image data and performance parameters among different SAR systems.The network adopts a U-shaped encoder decoder network structure,and adopts a dual attention encoder to adaptively capture the correlation between local and global features;The combination of GP(Gradient Profile)Loss and DAENet is added to the loss function.Furthermore,DAENet was trained and tested using the open-source dataset SOS(Deep SAR Oil Spill dataset),and achieved 85% m Io U and 86.1% F1 scores on the PALSAR test set in SOS,and 85.7% m Io U and 89.8% F1 scores on the Sentinel-1test set.This article also used the self-made Gaofen-3 test set to test the experimental results.On this dataset,92.2% of the m Io U score and 95.06% of the F1 score were obtained.The experimental results showed that both were the highest scores of similar semantic segmentation models.(2)In response to the low accuracy of deep learning semantic segmentation models in distinguishing oil spill areas and oil film like areas,this paper is based on the multiscale SAR image oil spill dataset from the MKLab(Multimedia Knowledge and Social Media Analytics Laboratory)laboratory.Experiments are conducted based on the Trans Unet semantic segmentation model,and an improved YOLOv5 object detection model is proposed to incorporate attention mechanism,Used to distinguish oil spill areas and oil like areas in SAR images,and further developed an oil spill target detection dataset based on the MKLab dataset for training and testing the oil spill target detection algorithm model.The results showed that the improved YOLOv5 model had an average accuracy(m AP)of 88%,achieving more accurate extraction of oil spill areas in large-scale SAR images.(3)Aiming at the problem that the oil particle model lacks real-time oil spill motion data to correct the prediction,this paper proposes an oil particle deployment algorithm based on SAR image and oil particle model,combining the physical characteristics of oil spill itself,and combining the segmentation results of oil spill areas in SAR image,which combines the near normal distribution matrix and SAR image matrix.The near normal distribution matrix is used to describe the physical characteristics of oil spill itself,The SAR image matrix is directly obtained from the grayscale values of the image,and the two are weighted to describe the distribution of oil particles.This algorithm can make the distribution of oil particles closer to the real situation,providing a new reference method for deploying oil particle models,and further improving the prediction accuracy of offshore oil spill drift diffusion.