| Marine oil spills will cause serious damage and huge losses to marine ecosystems and coastal economies.Synthetic Aperture Radar(SAR)has become an important means of marine oil spill detection due to its all-weather and all-day imaging capabilities.However,some "look-alikes" phenomena which have similar characterizations to oil slicks on SAR images will improves the false alarm and difficulty of oil spill detection.In recent years,polarimetric SAR has been gradually applied to the research of marine oil spill detection.Compared with SAR,the polarization information unique to polarimetric SAR is beneficial to distinguish oil slicks and look-alikes.In this thesis,oil spill detections were studied from polarimetric features of polarimetric SAR data.1)In order to improve the discrimination between oil slicks and look-alikes,based on the difference of scattering mechanism between oil slicks and look-alikes,the selfsimilarity parameter which is sensitive to the randomness of scattering target was introduced to improve the distinguish capability between oil slicks and look-alikes.The self-similarity parameter can well reflect the degree of scattering randomness of the target.Compared with the look-alikes and water,the scattering mechanism of the oil slicks is more complicated,and the degree of depolarization and randomness of scattering are higher.The RADARSAT-2 full-polarization SAR data which obtained by man-made oil spill experiment was used to analyze the difference of oil spill detection capability between the self-similarity parameter and other polarimetric features of oil spill detection.The qualitative and quantification analysis show that the self-similarity parameter has the better ability to distinguish between oil slicks and look-alikes than other polarimetric features.The UAVSAR full-polarization SAR data which obtained by man-made oil spill experiment was used to analyze the applicability of self-similarity parameter and other polarimetric features to distinguish oil slicks and look-alikes in different sea surface relative wind direction and sensor incident angle scene.The results show that the selfsimilarity parameter is more suitable for working under downwind conditions when the sea surface wind speed is large,and it performs well in the scene with incidence angle range from 29.7° to 43.5°.2)In order to improve the accuracy of oil spill detection,this paper proposes a novel multi-feature based method using random forest to detect oil spill,which including selfsimilarity parameter,Bragg energy proportion,geometric intensity,conformity coefficient,the real part of the co-polarization cross product,modified anisotropy coefficient,degree of polarization,pedestal height.The random forest with better classification performance,can effectively suppress noise and can better suppress errors caused by imbalance between classes was used to detect oil spill with multi-feature.Experiments result which using three fully polarimetric SAR data(RADARSAT-2,UAVSAR,SIR-C/X-SAR)show that compared with other three kinds of polarimetric SAR oil detection methods,the proposed method can improve the accuracy of oil spill detection.The importance of the feature based on random forest feedback once again proves that the self-similarity parameter has a good oil detection capability. |
