Training dataset RipAID is a dataset tailored to train Artificial Intelligence (AI) applications dedicated to automating rip current detection in RGB images. It comprises images captured by diverse acquisition systems (fixed stations, drones, satellites, smartphones), covering various beaches with varying fields of view, rip currents characteristics, and diverse meteoceanic and lighting conditions. The RipAID dataset contains three classes: ‘rip currents’, ‘sediment plume’, and ‘doubt’, labeled with oriented bounding boxes. Technical details RipAID v2.0.0 contains 6789 images distributed unevenly across various sites and imaging systems. For comprehensive information, users should consult the accompanying README file. Data preprocessing RipAID v2.0.0 builds upon version 1.0.0 with the addition of two new data sources (details in the Data Sources section). Pre-existing annotations from these sources were reviewed and modified to meet RipAID criteria. Specifically, ambiguous rip current instances were either relabeled as 'doubt' or removed. Where necessary, bounding boxes were converted to 'Oriented Bounding Boxes'. Furthermore, all time-averaged (Timex) and augmented images were excluded. For comprehensive details, consult the README file. Data splitting Researchers should consistently document their splitting method and rationale in publications to ensure reproducibility and facilitate comparisons. Classes, labels and annotations The RipAID dataset has been labelled manually using the 'Computer Vision Annotation Tool' (CVAT). In the RipAID dataset, three classes are differentiated: ‘rip_current’, ‘sediment’, and ‘doubt’. The README file contains further details on the criteria used to define bounding boxes. Label Description rip_current Clearly identifiable rip current, with defined lateral edges, and neck and/or head observable. doubt Plausible rip current, considering factors such as incoming wave patterns, disruption in wave breaking front, the presence of a defined neck, on other relevant hydrodynamic features. sediment Sediment plume that might relate to a rip current. RipAID v2.0.0 provides two main resources: oriented bounding box annotations in the Ultralytics YOLO-OBB format, and a CVAT backup file. The inclusion of the CVAT backup file is intended to facilitate easy data modification, expansion, and re-formatting. Parameters RGB values or any transformation in the colour space can be used as parameters. Data sources The training dataset was constructed from three distinct data sources:(i) RipAID V1.0.0, which includes images from the SIRENA videomonitoring system captured at two different Mediterranean microtidal beaches; (ii) The UFSC (2023) Roboflow public dataset, consisting of crowd-sourced smartphone imagery acquired by the Brazil CoastSnap stations at Moçambique and Santinho sandy beaches; (iii) The RipScout dataset (Khan et al., 2025), comprising aerial images collected from Google Earth and drone flights conducted along the shoreline. Data quality The uneven spatiotemporal distribution of images across sites and image acquisition systems, is influenced by the natural occurrence of rip currents and the data collection strategy. RipAID users should be aware of this variability to ensure accurate interpretation of results and to mitigate potential biases in their applications. In addition, the inherent variability and amorphous nature of rip currents can lead to challenges in achieving consistently accurate and unambiguous labeling. This should be considered when utilizing the RipAID dataset. Image resolution RipAID images vary in resolution. The smallest image is 640x640 px, and the largest is 1280x960 px. Contact information For further technical inquiries or additional information about the annotated dataset, please contact jsoriano@socib.es.

The RipAID dataset was made possible by the invaluable collaboration of the General Directorate of Emergencies and Home Affairs of the Balearic Islands ('Direcció General d’Emergències i Interior') and the Balearic Islands lifeguard teams, to whom we extend our sincere gratitude. We also appreciate the open-access image repositories provided by the authors of Khan et al (2025) and UFSC (2023), as these were instrumental in RipAID. We hope the result of this collaborative effort will be a significant contribution to beach safety and the creation of decision-making tools for effective beach management. Furthermore, we acknowledge the support and contributions of the iMagine project (https://www.imagine-ai.eu/) managers and partners in the area of AI-based image analysis data and services.