Oil Spill Detection In SAR Images Based On Deep Learning Algorithm

Oil spill accidents happened more and more frequent with the rapid development of oil exploitation and marine transportation.Once the oil spill occurs,the crude oil layer will cause persistent damage to marine environment and economic.Therefore,the accurate detection of oil spill area plays an important role in marine environmental protection.Synthetic Aperture Radar(SAR)sensor is widely used in oil spill detection because of its wide imaging range and all-weather operation.This paper studies oil spill detection based on polarimetric SAR data and the main research contents includes:(1)An Refined Convolutional Neural Network(RCNN)oil spill detection method is proposed.The global average pooling(GAP)layer is used to replace the fully connected layer for oil spill area classification,which can reduce the number of parameters.Experiments was carried out on Radarsat-2 full-polarized data,the highest accuracy of RCNN achieves 93.7%,while the number of parameters of GAP reduced from 76800 to 16 compared to fully connected layer.The experiments also proved that polarized parameters has the ability to distinguish oil spill area from other similar areas.(2)Proposed an oil spill detection method combining semantic segmentation and SLIC superpixel segmentation to consider the context of pixels.This part designed a semantic segmentation model independently and introduced superpixel algorithm into oil spill detection.Experiments extracted five groups of polarized parameters from Radarsat-2 and SIR-C/X-SAR data,and SLIC superpixel could provide the context information of SAR image,improve the accuracy of oil spill detection.Taking Yamaguchi parameters as example,the MIo U rised from 86.4% to 90.5% by superpixel.(3)An adaptive threshold calculation and pixel value scaling method based on SLIC superpixel segmentation is proposed,in order to overcome the imaging differences between different SAR data.Based on SLIC superpixel,this method can adaptively calculate the threshold value and scale the pixel value,and detect oil spill area by semantic segmentation model.Experiments are carried out on the data set formed by eight Sentinel-1 dual-polarized data.The maximum Miou value of the dataset is only 69.7% when polarized parameters and SLIC are used,and it increased to 87.4% after pixel scaling.