On the Features Extracted From Dual-Polarized Sentinel-1 Images for Deep-Learning-Based Sea Surface Oil-Spill Detection

This study comprehensively investigates the effects of the oil-spill features extracted from dual-polarized Sentinel-1 images on oil-spill detection by using the DeepLabv3+ model, aiming at providing a constructive reference in the selection of features used for oil-spill detection based on dual-polarized SAR images. First, several commonly used features were extracted from Sentinel-1 SAR images, and corresponding oil-spill datasets were constructed based on these features. Next, the effects of individual features and their combinations on the oil-spill detection were analyzed separately within the DeepLabv3+ framework. Based on this analysis, a new synthetic feature was constructed using weighted bilinear pooling, derived from the feature combination that achieved optimal oil-spill detection performance. In addition, the original DeepLabV3+ model was modified by adding a specialized module to process the wind speed information. Finally, the performances of these features across different deep learning models and across another newly created datasets are further discussed. Experimental results showed that 1) the accuracy is improved with the increase of features but the degree of improvement varies; 2) the combination of polarization, texture, and intensity features (i.e., P(A, H)+T(V, E)+C<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">11</sub>)achieves the optimal detection accuracy with the mean Intersection over Union (mIoU) of 90.43%. Furthermore, the synthetic feature based on the optimal features can also achieve a high detection performance with the mIoU of 88.21% more efficiently; 3) the modified DeepLabV3+ model can capture the wind speed feature of oil spills and make the mIoU under the combination of P(A, H) + T(V, E) + C<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">11</sub> + WS reach to 90.89%; 4) these features or their combinations have the similar patterns on the oil-spill detection for different models and dataset. Overall, this study provides a valuable insight into the SAR-based oil-spill features, which can further serve for the construction of oil-spill features and the optimizations of deep learning networks for more effective oil-spill detect.