【作者】 曹文俊；

【导师】 盛辉；

【作者基本信息】 中国石油大学(华东) ， 测绘工程（专业学位）， 2023， 硕士

【摘要】 随着海洋资源的不断开发与利用,海上活动愈加频繁,船舶与海上石油平台溢油事件频发,给海洋环境带来了巨大的破坏,因此及时准确的进行溢油检测刻不容缓。SAR因其具有全天时、全天候的能力而成为溢油检测的主要传感器。本文基于海丝一号(HISEA-1)数据、Radarsat-2数据以及Sentinel-1数据集,分别提取纹理特征,利用极度梯度提升树算法(XGboost)进行纹理特征选择,纹理特征选择结果与SAR灰度图像组合构建溢油数据集。发展了一种基于U-NET改进的像素级溢油检测模型,并将其与Faster-RCNN目标检测算法相结合,发展了一种海上溢油检测框架。本文主要研究内容包括:(1)开展了溢油SAR图像的纹理特征提取与选择。针对HISEA-1、Radarsat-2以及Sentinel-1数据集,分别提取了8种纹理特征。利用XGboost算法对各特征进行重要性排序,遴选了相关性等三种纹理特征,将其与SAR灰度特征组成溢油组合特征,作为溢油检测的输入数据。(2)针对纹理特征信息提取能力弱的问题,发展了一种基于U-NET改进的双分支像素级溢油检测模型(AW-net)。采用双分支编码器代替U-NET原本的编码器,并采用注意力门对双分支编码器中提取的多尺度纹理信息与灰度信息进一步捕获上下文信息,将捕获到的信息进行融合,获取更加显著的特征信息。利用三种数据分别对改进的模型进行测试,并与U-NET、Attention U-NET和FCN进行对比,结果表明改进的模型能够有效进行像素级溢油检测,且应用在不同数据集中仍能得到较好的检测结果,具有强鲁棒性。(3)针对溢油目标检测误检率高的问题,发展了一种Faster-RCNN与改进的U-NET模型相结合的溢油检测框架。框架首先利用Faster-RCNN对图像中的溢油进行初步目标检测,再利用改进的U-NET对目标检测框内区域进行像素级溢油检测,判别框内区域是否为溢油。结果与Faster-RCNN、YOLOv4、SSD、AW-net、U-NET、Attention U-NET和FCN等模型进行对比。结果表明本文框架在目标检测中能够有效去除溢油误检部分,获得更高的目标检测精度,且像素级溢油检测同样能够达到较好的精度。更多还原

【Abstract】 With the continuous development and utilization of Marine resources,Marine activities become more and more frequent,and ships and offshore oil platforms spill oil frequently,which brings great damage to the Marine environment.Therefore,timely and accurate oil spill detection is urgent.SAR has become the primary sensor for oil spill detection because of its all-day,all-weather capability.In this paper,based on HISEA-1 data,Radarsat-2 data and Sentinel-1 data set,texture features were extracted respectively,and extreme gradient lift tree algorithm(XGboost)was used to select texture features.Texture feature selection results were combined with SAR gray image to construct oil spill data set.An improved pixel level oil spill detection model based on U-NET was developed,and combined with Faster-RCNN target detection algorithm,a framework for offshore oil spill detection was developed.The main research contents of this paper include:(1)The texture feature extraction and selection of oil spill SAR images are carried out.Eight texture features were extracted from HISEA-1,Radarsat-2 and Sentinel-1 data sets,respectively.XGboost algorithm was used to rank the importance of each feature,and three texture features,such as correlation,were selected,which were combined with SAR gray feature to form the oil spill combination feature as the input data of oil spill detection.(2)To solve the problem of weak extraction ability of texture feature information,an improved two-branch pixel level oil spill detection model(AW-net)based on U-NET was developed.A double-branch encoder is used to replace the original U-NET encoder,and an attention gate is used to further capture the context information of the multi-scale texture information and gray level information extracted from the double-branch encoder,and the captured information is fused to obtain more significant feature information.The improved model was tested with three kinds of data respectively,and compared with U-NET,Attention U-NET and FCN.The results show that the improved model can effectively detect oil spills at pixel level,and can still get good detection results in different data sets,with strong robustness.(3)Aiming at the problem of high false detection rate of oil spill target,an oil spill detection framework based on Faster-RCNN and improved U-NET model is developed.Combining the target detection with the pixel level oil spill detection,the initial detection of oil spill in the image is carried out using the speed-r CNN,and then the improved U-NET is used to detect the pixel level oil spill in the area of the target detection box to determine whether the area in the box is oil spill.Comparison with Faster-RCNN,YOLOv4,SSD,AW-net,U-NET,Attention U-NET,FCN and other models.It shows that the framework in this paper can effectively remove the false oil spill detection part in the target detection,and obtain higher target detection accuracy,and the pixel level oil spill detection can also achieve better accuracy.更多还原