Adaptive oil spill detection network for scene-based PolSAR data using dynamic convolution and boundary constraints

Abstract

Remote sensing monitoring of oil spills is essential for ecological and environmental management. Polarimetric synthetic aperture radar (PolSAR) data have been extensively utilized for oil spill detection owing to the advantages of multi-polarization channels and all-time, all-weather observation capability. However, suspected oil spill phenomenon (algal blooms or wave shadows) and wind-caused oil spill drifting can lead to variable textures and irregular boundaries of oil spill coverage. These increase the complexity of oil spill scenes and pose challenges for PolSAR oil spill detection. Convolutional neural network (CNN) provides the potential to cope with such problems. However, it’s challenging for the existing CNN methods to extract high-order semantic features and attach to accurate boundaries. In this paper, a scene-adaptive oil spill detection network is proposed with a dynamic convolution and boundary constraint. First, aiming to improve the recognition ability of semantic features, a dynamic convolution module is introduced to replace the traditional convolutions and maximum pooling operations. Such module combines the conditionally parameterized convolutions (CondConv) layer with two-dimensional discrete wavelet transform (2D-DWT) layer. Second, a boundary constraint module is proposed by using the multitasking strategy to extract boundaries of oil spill for guiding the network training with the principle of the minimum deviation between the extracted and real boundaries. To sum up, we propose an oil spill detecting network, which fuses semantic features and boundary information under the attention mechanism. Radarsat-2 PolSAR data from the Gulf of Mexico and the European North Sea region are used for demonstrating the above improvements. In contrast to the current oil spill detection methods, our method improves the overall accuracy by up to 3.4% and 2.3% in two study regions, while the Macro-F1-scores are improved by 5.62% and 4.62% respectively. Besides, this article further discusses the influence of feature selection on the oil spill detection.

Keywords

• Polarimetric SAR
• Oil spill detection
• Convolutional neural network
• Conditionally parameterized convolutions
• Boundary constraint