Transfer Learning Between Sentinel-1 Acquisition Modes Enhances the Few-Shot Segmentation of Natural Oil Slicks in the Arctic

Abstract

Natural seepage is a significant contributor to marine hydrocarbon inputs. Remote and intermittent seeps are difficult to monitor in the field, yet surface oil slicks can be observed by spaceborne synthetic aperture radar (SAR) because they reduce backscatter, creating potential for automatic mapping. In mapping tasks like segmentation, deep learning models excel, albeit needing large amounts of labeled images. To deal with the scarcity of labeled images, transfer learning is an approach which makes use of knowledge from related domains. In the case of oil slicks, differences between Sentinel-1 acquisition modes, such as the interferometric wide (IW) in the North Sea and extra wide (EW) in the Arctic, complicate direct model transfer. Here, we present a use case where transfer learning enhances the few-shot segmentation of natural oil slicks. We used labeled slicks in IW images in the North Sea to pretrain a series of DeepLabv3 and segment anything models (SAMs). These models were then fine-tuned on EW-labeled slicks from two documented Arctic seeps on which we have only limited observations. Our results show clear evidence that transfer learning improves segmentation, notably in challenging and noisy images. Few studies, if any, have addressed transfer learning between SAR acquisition modes. This work contributes to improved monitoring of poorly understood or yet undiscovered hydrocarbon seeps.