We propose a deep-learning framework to detect and categorize oil spills in synthetic aperture radar (SAR) images at a large scale. Through a carefully designed neural network model for image segmentation trained on an extensive dataset, we obtain state-of-the-art performance in oil spill detection, achieving results that are comparable to results produced by human operators. We also introduce a classification task, which is novel in the context of oil spill detection in SAR. Specifically, after being detected, each oil spill is also classified according to different categories of its shape and texture characteristics. The classification results provide valuable insights for improving the design of services for oil spill monitoring by world-leading providers. Finally, we present our operational pipeline and a visualization tool for large-scale data, which allows detection and analysis of the historical occurrence of oil spills worldwide.
25 April 2017 in IEEE Transactions on Geoscience and Remote Sensing
Oil spill detection using a time series of images acquired off Norway in June 2015 with the uninhabited aerial vehicle synthetic aperture radar is examined. The relative performance of a set of features derived from quad-polarization versus hybrid-polarity (HP) modes in detection of various types of slicks as they evolve on a high wind driven sea surface is evaluated. It is shown that the HP mode is comparable with the full-polarimetric mode in its ability to distinguish the various slicks from open water (OW) for challenging conditions of high winds (9-12 m/s), small release volumes (0.2-0.5 m 3 ), and during the period 0-9 h following release. The features that contain the cross-polarization component are better for distinguishing the various slicks from open water at later and more developed stages. Although these features are not available in the HP mode, we identify alternative features to achieve similar results. In addition, a clear correlation between the results of individual features and their dependence on particular components within the two-scale Bragg scattering theory is identified. The features that show poor detectability of the oil slicks are those that are independent of the small-scale roughness, while the features resulting in good separability were dependent on several factors in the two-scale Bragg scattering model. We conclude that the HP mode is a viable alternative for SAR-based oil spill detection and monitoring that provides comparable results to those from the quad-polarimetric SAR.
New quantitative and semiautomated methods for analyzing oil slick evolution using a time series of $L$ -band synthetic aperture radar (SAR) images with short repeat time are developed and explored. In this study, two methods that are complementary in terms of identifying temporal changes within an oil slick are presented. The two methods reflect two ways of evaluating the oil slicks. The first method identifies regions within the slick that show persistently high damping ratio (the contrast between clean sea and oil intensity), using higher damping values as a proxy for increasing oil thickness. This method also weights the age of the scenes as the algorithm incorporates new images. The second method outputs the short-term drift pattern and the changes in the damping ratios and copolarization ratios between two scenes, proxies for thickness, and emulsification. Both methods can aid in identifying regions of high priority for oil recovery. Due to the simplicity of the methods, they can be adapted to time-series data from different types of sensors, e.g., optical and SAR imagery. The methods are demonstrated on three $L$ -band uninhabited aerial vehicle SAR UAVSAR time series acquired in November 2016 over a persistent seep in the Mississippi Canyon Block 20 of the Gulf of Mexico. The results of the two methods clearly show the movement and the weathering of the oil as a function of both time and location.
1 June 2020 in IEEE Transactions on Geoscience and Remote Sensing
Utilizing several Synthetic Aperture Radar (SAR) missions will provide a data set with higher temporal resolution. It is of great importance to understand the difference between various available sensors and polarization modes and to consider how to homogenize the data sets for a following combined analysis. In this study, a uniform and consistent analysis across different SAR missions is carried out. Three pairs of overlapping hybrid- and full-polarimetric C-band SAR scenes from the Radar Imaging Satellite-1 (RISAT-1) and Radarsat-2 satellites are used. The overlapping Radarsat-2 and RISAT-1 scenes are taken close in time, with a relatively similar incidence angle covering sea ice in the Fram Strait and Northeast Greenland in September 2015. The main objective of this study is to identify the similarities and dissimilarities between a simulated and a real hybrid-polarity (HP) SAR system. The similarities and dissimilarities between the two sensors are evaluated using 13 HP features. The results indicate a similar separability between the sea ice types identified within the real HP system in RISAT-1 and the simulated HP system from Radarsat-2. The HP features that are sensitive to surface scattering and depolarization due to volume scattering showed great potential for separating various sea ice types. A subset of features (the second parameter in the Stokes vector, the ratio between the HP intensity coefficients, and the angle) were affected by the non-circularity property of the transmitted wave in the simulated HP system across all the scene pairs. Overall, the best features, showing high separability between various sea ice types and which are invariant to the non-circularity property of the transmitted wave, are the intensity coefficients from the right-hand circular transmit and the linear horizontal receive channel and the right-hand circular on both the transmit and the receive channel, and the first parameter in the Stokes vector.