Semantic Segmentation of Metoceanic Processes Using SAR Observations and Deep Learning

Aurélien Colin,Nicolas Longépé,Alexis Mouche,Romain Husson,Ronan Fablet,Charles Peureux,Pierre Tandeo

doi:10.3390/rs14040851

Abstract

Through the Synthetic Aperture Radar (SAR) embarked on the satellites Sentinel-1A and Sentinel-1B of the Copernicus program, a large quantity of observations is routinely acquired over the oceans. A wide range of features from both oceanic (e.g., biological slicks, icebergs, etc.) and meteorologic origin (e.g., rain cells, wind streaks, etc.) are distinguishable on these acquisitions. This paper studies the semantic segmentation of ten metoceanic processes either in the context of a large quantity of image-level groundtruths (i.e., weakly-supervised framework) or of scarce pixel-level groundtruths (i.e., fully-supervised framework). Our main result is that a fully-supervised model outperforms any tested weakly-supervised algorithm. Adding more segmentation examples in the training set would further increase the precision of the predictions. Trained on 20 × 20 km imagettes acquired from the WV acquisition mode of the Sentinel-1 mission, the model is shown to generalize, under some assumptions, to wide-swath SAR data, which further extents its application domain to coastal areas.

Highlights

Since neural-networks-based algorithms have become the state-of-the-art framework for a wide range of image processing problems in the 2010s, the use of deep learning approaches has been extended to many kinds of remote sensing data, including for instance infrared imagery [1], land applications of Synthetic Aperture Radar (SAR) [2] and SAR-optical fusion [3]
The Dice index would always be equal to 0 in case of non-detection whereas the range of a mean square error or a crossentropy would depend on the surface covered by the phenomena
The extrapolation of our results suggests that building a fully groundtruthed dataset of dozen thousands segmented images may result in a significant gain in the segmentation performance

Summary

Introduction

Since neural-networks-based algorithms have become the state-of-the-art framework for a wide range of image processing problems in the 2010s, the use of deep learning approaches has been extended to many kinds of remote sensing data, including for instance infrared imagery [1], land applications of SAR [2] and SAR-optical fusion [3]. The SAR imagery is sensitive to the surface roughness which can be impacted by the wind, the waves, the presence of ships or icebergs, a surface viscosity difference caused by oil or biological slicks, the precipitations or by sea ice in polar regions [4]. SAR can be acquired in almost all conditions, especially over cloudy regions and at night. For these reasons, SAR images can be processed into a wide variety of geophysical products such as wind maps, wave spectra, surface currents, and its full potential still remains unexploited, especially in terms of scientific and operational services [5].

Objectives

Results

Discussion

Conclusion