Internal Solitary Waves characterization on Sentinel-1 observations

Internal Solitary Waves (ISW) are fundamental components of coastal ocean dynamics, playing a pivotal role in sediment transport, nutrient mixing, and the dissipation of tidal energy. These non-linear oscillations, driven by gravitational forces within stratified water columns, manifest on the ocean surface as curvilinear bands of varying roughness. This modulation of surface roughness makes them detectable by Synthetic Aperture Radar (SAR) sensors despite their negligible slope on the surface. However, while SAR offers high-resolution, all-weather monitoring capabilities, the automated quantification of ISW characteristics remains a complex challenge. Difficulties arise not only from signal contamination by atmospheric and oceanic features such as wind and rainfall, but also from the high geometric variability of the wavefronts themselves.

We propose a comprehensive analytical framework for detecting, segmenting, and characterizing internal waves using Sentinel-1 SAR observations. The methodology is developed and tested on a dataset of Level-1 Ground Range Detected Interferometric Wide Swath (GRD IW) images acquired along the northeast coast of South America. This region is oceanographically unique due to the intense stratification induced by massive coastal river discharge, specifically from the Orinoco and the Amazon rivers.

Our approach employs a two-stage process that synergizes deep learning with geometric modeling. In the first stage, a U-Net architecture segment the observation into two classes: the wave packets and the specific leading waves. The model is trained to predict a distance map relative to the feature boundaries rather than a simple binary mask. This pixel-wise regression, performed at a resolution of 50 m/px, is validated against manual annotations, providing a robust identification capability where no comparable high-resolution groundtruth exists.

Following segmentation, the second stage focuses on physical characterization. The leading wave of each detected packet is modeled using an adaptive polynomial function. Optimized via gradient descent, this function fits the curvilinear shape of the wavefront. This mathematical representation allows for the precise computation of wave orientation and propagation direction. Subsequently, the wave packet is projected onto a geometry orthogonal to the leading wave to obtain a curvature-independant representation of the wave packet. A Fourier Transform is applied to this projection to calculate the dominant wavelength. Furthermore, by analyzing the spacing and propagation direction, deducing the generation chronology of successive packets produced by tidal cycles.

Results demonstrate strong agreement between automated detections and regional dynamics. The spatial distribution reveal detection in the vicinity of . In the vicinity of Trinidad and Tobago, the spatial distribution highlights generation hotspots northwest of straits and continental shelf. Activity peaks during the autumn months, coinciding with the maximum discharge of the Orinoco River, inline with a strong modulation by stratification stability. Statistical analysis reveals a mode wavelength of approximately 350 meters, but is biased by the manual segmentation dataset. While challenges remain regarding overlapping wave fields, this tool provides a robust pathway for monitoring internal wave energetics, offering significant potential for synergy with altimetry missions, such as SWOT, and in-situ coastal management.