Julien Vadnais

Deep transfer learning for mapping rare Arctic oil seeps observed from satellites

Julien Vadnais – Department of Earth Science, UiB

Abstract

Natural seepage is a significant contributor to marine oil inputs. Remote and intermittent seeps are difficult to monitor in the field, yet 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 scarcity of labeled images, transfer learning is an approach commonly used in computer vision, though still underutilized in remote sensing.

Different Sentinel-1 acquisition modes, in the Arctic compared to elsewhere, complicate direct model transfer for oil slick mapping. Here, we present a use-case where transfer learning enhances the segmentation of natural oil slicks, most notably in challenging and noisy images. Our method proves efficient even with limited training images, offering new prospects for studying poorly understood or yet undiscovered hydrocarbon seeps.