Sentinel-1 SAR Wave Mode vignettes acquired globally, covering up to 75 minutes of imagery per orbit. Data sourced from IFREMER's X-Waves platform.
IWS Validation Studio
Human-in-the-loop verification of machine learning predictions for the Internal Waves Service — building a global reference dataset from Sentinel-1 SAR imagery.
Internal Solitary Waves
Internal solitary waves (ISWs) are nonlinear disturbances that propagate along density interfaces within the stratified ocean. With amplitudes exceeding 100 metres and generating the highest vertical velocities observed in the deep ocean, these phenomena carry significant implications for underwater navigation, offshore infrastructure integrity, sediment resuspension, and deep-ocean mixing processes.
Synthetic Aperture Radar (SAR) remains the most effective modality for observing ISWs at global scale. Unlike optical sensors, SAR penetrates cloud cover and captures the surface roughness signatures that internal waves imprint on the sea surface. The Internal Waves Service ingests Sentinel-1 Wave Mode (WV) vignettes continuously from IFREMER's X-Waves platform, processing thousands of acquisitions per orbit to produce near real-time detection maps.
Detection Pipeline
A pre-trained neural network classifies each vignette as internal wave positive or negative, producing a confidence score that quantifies prediction certainty.
Domain experts review ML predictions in guided sessions, producing four-class verdicts (positive, negative, invalid, inconclusive) that form a validated reference dataset.
About This Application
The IWS Validation Studio is a hypermedia web application built with Julia and Datastar. All application state is owned by the server and projected to the browser as plain HTML fragments — no client-side JavaScript framework, no JSON APIs, no virtual DOM. The result is a lightweight, fast interface that stays in sync with the database at all times.
Reviewers authenticate, configure review sessions with filters for confidence range, ML label, and prior review count, then work through image queues at their own pace. Each verdict is recorded alongside the original ML classification, enabling systematic evaluation of model performance and the construction of high-quality training data for future iterations.
A collaboration between
AIR Centre · University of Porto · IFREMER · NERSC · NOC · WHOI · MIT · ESA · NASA JPL