Abstract

Wave breaking is a fundamental process in the nearshore zone. Yet, due to its chaotic nature, automatically identifying its spatio-temporal occurrence on a wave-by-wave basis remains a challenge. To address this, a generalized machine learning-based methodology is proposed for detecting wave-breaking events across a wide range of environmental and hydrodynamic conditions. The approach builds on a previously trained U-Net architecture using curated data for Duck (USA), which is extended using Transfer Learning (TL) and Human-in-the-Loop (HITL) techniques, enabling effective model adaptation to new coastal sites and conditions with minimal labeled data. The methodology is tested at three distinct locations: Duck, North Carolina (USA); the mouth of the Maipo River in San Antonio (Chile); and Saint-Pierre Beach, at Palavas-les-Flots (France). In addition, three types of users of varying proficiency are considered to assess how easy it would be to adopt the methodology. The resulting models achieve accuracies in wave breaking identification between 80% and 90%, depending on the site, with qualitative assessments confirming robust performance even under challenging conditions, such as variable lighting and wave-structure interaction. The open-source implementation aims to facilitate community use and adaptation. As a result, accurate identification of breaking patterns is expected to be a valuable tool for advancing our understanding of nearshore processes.