Study On Oil Spill Detection In SAR Images Based On Deep Learning And Decision Level Information Fusion

With the rapid development of global economy,the offshore oil and gas exploitation and ship transportation are increasing rapidly,which has increased the possibility of oil spill accidents gradually.Oil spill pollution caused by oil spill accidents has the characteristics of the large acreage,lasted long and various forms and it has seriously damaged the marine ecological environment and restricted the development of marine fishery,tourism and other economic industries.The United States National Academy of Sciences reported that oil spill pollution remains one of the 32 global marine environmental problems by 2030.It is an important task for us to discriminate the oil spills,tarce back to the source,monitor the drift and diffusion direction and estimate the volume in an effective and timely manner.Efficient and effective oil spill monitoring can provide data support for oil spills cleaning,oil spill disaster assessment and ecological restoration.With the rapid development of remote sensing technology,synthetic aperture radar(SAR)has become one of the most important methods for oil spill monitoring due to its all-time,all-weather and continuous observation capability.The key and difficult point of oil spill detection in SAR images is to distinguish oil spills and look-alikes accurately.Aiming at the above key and difficult point,the oil spills and look-alikes are selected as the study object and the study on oil spill detection based on deep learning and decision level information fusion is carried out.The main contents of this thesis are as follows.(1)The oil spill and look-alike sample database was established with oil spill and look-alike samples which interpreted from SAR images with the assistance of the prior knowledge of experts and multi-source remote sensing data.(2)The Subjective Bayesian method was used to model the process of oil spill detection with environmental factors and the oil spill detection method driven by environmental factors was established in this paper.(3)The Faster-RCNN(Faster Region-based Convolutional Neural Network)model was used to highly automate the process of feature extraction and feature selection.The oil spill detection method based on Faster-RCNN model was established from the perspective of image features and the Faster-RCNN model was optimized by the Soft Non-Maximum Suppression algorithm.(4)The D-S evidence theory(Dempster-Shafer evidence theory)improved by fuzzy theory was used to fuse the oil spill detection results of the subjective Bayesian method and Faster-RCNN model at decision level.An oil spill detection method based on decision level information fusion was established in this paper.Based on 2100 samples from the oil spill and look-alike sample database and the corresponding wind speed,the distance between dark spots and oil platforms and the distance between dark spots and waterways,the above three methods were verified in this paper.The experimental results show that the sample identification rate(SIR),oil spill detection rate(ODR)and oil spill false alarm rate(OFAR)of the subjective Bayesian method are 67.70%,59.87% and 37.24% respectively.Compared with the subjective Bayesian method,the ODR and SIR of Faster-RCNN method are increased by about 14% and 10%,respectively.The OFAR is reduced by about 10%.Compared with the Faster-RCNN method,the ODR of the decision level information fusion method reduced by 1.32%,but the SIR is increased by 1.04% and the OFAR is reduced by 2.39%.Seven oil spill SAR images are used to verify the effectiveness of the decision level information fusion method.The verification result shows that the ODR and SIR reach 80% under simple sea state.Meanwhile,the ODR and SIR are reach 60.86% and 73.68% respectively under complex sea state.The detection result demonstrates that the decision level information fusion oil spill detection method can effectively detect oil spills in SAR images and it has strong potential for business application.