the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 Mar 2026
Oriented Object Detection for Complex Hydrodynamic Features: A Multi-Platform Rip Current Identification System
Abstract. Rip currents are hazardous, fast-moving seaward flows and remain one of the leading causes of rescues and drownings on surf beaches, yet their automated detection remains a significant challenge due to their amorphous, dynamic morphology and the environmental complexity of the surf zone. This study introduces a novel platform-agnostic deep learning–based framework for automated rip current detection from beach imaging platforms, integrating three core contributions: a diverse new dataset, a rigorous architectural benchmark, and a deployable operational tool. We first present RipAID, a comprehensive dataset enriched with multi-platform imagery and multiple viewing angles to ensure scale-invariant learning. Building on this resource, a systematic evaluation of state-of-the-art architectures demonstrates that geometric fidelity is critical; specifically Oriented Bounding Boxes (OBB) significantly outperform standard axis-aligned methods. Our optimized YOLOv11n-OBB model achieves robust performance (mAP50: 0.927), with inference speeds from 2.4 to 60 FPS on hardware ranging from edge devices to GPU workstations. To bridge the gap between research and practice, and ensure that the results are reusable and reproducible, the framework and model weights have been released as an open-source, containerized module (socib-rip-currents-detection), providing the coastal safety community with a scalable, ready-to-use and standardized tool for continuous, automated rip current monitoring.
- Preprint
(28612 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-1138: good idea but needs significant improvements', Anonymous Referee #1, 17 Apr 2026
-
AC1: 'Reply on RC1', Jesús Soriano-González, 15 May 2026
We thank the reviewer for the highly constructive feedback and meticulous review of our manuscript. The attention to detail, highlighting important methodological nuances and even taking the extra time to consult an additional domain expert, is highly valuable. We believe that the reviewer’s comments and corrections will significantly strengthen both the scientific rigor and the practical grounding of this paper. Below, we provide a detailed, point-by-point response to each of the comments.
- 1. Replacement of Custom Evaluation Metric with Standardized F2 Score
The rip_fit metric was initially introduced during the early experimental phases of this project, and we originally chose to include it in the manuscript for continuity. However, we completely agree with the reviewer’s assessment that introducing a custom metric risks fragmenting the literature. Consequently, we propose entirely removing the rip_fit metric from the manuscript to avoid unnecessary complexity and noise, and to align our methodology with community standards. We have recalculated our results using the F2 score, and all relevant text, tables, and figures will be updated accordingly in the revised document. To facilitate the review, we have attached the revised figures to this response. As these updated results demonstrate, adopting the F2 score does not alter the main conclusions of the models benchmarking.
- 2. Clarification on Dataset Accessibility and Justification of Novelty Claims
We would like to clarify that the CnE UFSC (2023) dataset remains publicly available. The URL appeared invalid in the preprint due to a PDF line-break formatting error. The correct, unbroken link is: https://universe.roboflow.com/cne-ufsc/rip-current-coastsnap-mocambique-santinho .
Regarding the size and diversity of the dataset, we agree that relying solely on frame count, our dataset is modest. Our intention behind claiming "diversity" and "novelty" was based on two main factors: (i) the extension of prior efforts with the addition of more imaging platforms, viewing geometries, and low-energy Mediterranean beaches. (ii) OBB re-annotation: Even for the images curated from previous standard datasets, images were re-annotated using OBBs, providing a dataset designed specifically for oriented detection. We have revised the text in both sections to be more precise. The proposed text explicitly states that the dataset’s main or differential contribution lies in the OBB annotations rather than its absolute size, and we have ensured that prior works are accurately contextualized.We propose modifying the first paragraph of the discussion (line 282) to read:
"The primary objective of this study was to advance the automation of rip current detection using beach imaging systems by delivering a platform-agnostic tool that integrates novel computer vision approaches. While previous studies have demonstrated the feasibility of deep learning for this task (de Silva et al., 2021; Zhu et al., 2022; Rashid et al., 2023), they have often relied on axis-aligned annotations and aerial imaging, which are difficult to obtain for near-real time applications. By curating RipAID v2.0.0 (Soriano-González et al., 2026), an open-access dataset enriched with multi-platform imagery and oblique viewing angles, we aim to provide a new perspective, using OBB to annotate rip currents. Beyond varied viewpoints and image geometries, the dataset also includes images from low-energy coastal environments in the Mediterranean, where rips often exhibit lower intensity and less visible structure, thereby broadening the variety of scenarios represented."
The revised Conclusion (line 287) will state:
"A central part of this work involves the development of RipAID v2.0.0, an updated dataset that builds upon previous efforts through the incorporation of OBB annotations. By reprocessing imagery from previous studies alongside new data, this version provides a multi-platform representation of rip current morphodynamics and aims to address specific camera angle limitations, while improving model adaptability to varied coastal environments."
- 3. Annotation Procedure and Justification of Figure 8c
(i) We agree with the reviewer that the annotation procedure required further elaboration in the text, and while we did not calculate a formal inter-annotator agreement coefficient (which we acknowledge as a limitation to be addressed in future iterations of the dataset), we did implement a multi-tiered review process to mitigate individual bias and uncertainty. The process was conducted by three researchers: two performed the initial annotations, while the third independently reviewed the entire dataset to ensure consistent criteria. To clarify the annotation procedure for the readers, we propose adding the following paragraph to the Training Dataset section:
"Annotations were performed following a predefined protocol: (1) OBBs must encompass both lateral boundaries of the rip current whenever possible; (2) rip currents located at the edge of the image are labeled if clearly identifiable; and (3) features exhibiting high uncertainty are labeled as ‘doubt’. Since rip currents are frequently misidentified (Pitman et al., 2021), ensuring annotation quality is inherently challenging. To minimize labeling errors and individual biases, the annotation process was conducted by three researchers: two performed the initial labeling, while a third independently reviewed the entire dataset to mitigate criteria heterogeneity and correct potential mistakes. While a formal inter-annotator agreement coefficient was not calculated for this iteration, this multi-tiered review process helped to establish a more homogeneous baseline. Further details regarding the annotation criteria can be found in the RipAID dataset repository (Soriano-González et al., 2026)."
(ii) Regarding Figure 8c, the human annotation on the left is indeed incomplete, while the YOLO detections on the right accurately identify the true rip currents. We deliberately selected and retained this specific image to illustrate this exact phenomenon. Because human annotators struggle to perfectly identify all rip currents (as noted by the literature cited by the reviewer), the dataset contains unavoidable human errors. However, Figure 8c demonstrates that the model has generalized the visual features of rip currents well, so it can detect hazards that the human annotator missed. We refer to this in the text as the "Ground Truth Paradox":
"This capacity of the model to occasionally outperform the human annotators highlights a ‘ground truth paradox’ (Plank, 2022). For instance, Figure 8c presents a scenario where the model detected two additional rip currents alongside the labeled instance; post-hoc visual inspection suggests these detections are likely valid."
While future versions of the RipAID dataset will correct these specific annotation errors, we believe it is important to keep Figure 8c as it is in this manuscript. It provides evidence that AI-assisted detection can help mitigate human limitations/mistakes, as outlined in lines 371-373 and 404-407 of the document.
- 4. Clarification on Data Splitting and Correction of K-Fold Cross-Validation Methodology
(i) Clarification on data splitting: We can confirm that the data partitioning was strictly performed at the image level, not the instance level. We mentioned the instance distribution in the original text simply to demonstrate that despite splitting by images, the resulting folds maintained a balanced ratio of target features.
(ii) Correction of the K-Fold Methodology: Following the reviewer's recommendation, we have completely re-run the 10-fold cross-validation isolating a fixed, independent test set (10% of the images), and performed the 10-fold split exclusively on the remaining 90% of the data. We have updated the corresponding figure and numbers (attached to this response). Aligning with this, we will rewrite the 2.4.5 cross-validation methodology section in the manuscript; to read:
“In order to obtain a more reliable estimate of model performance and stability, a 10-fold cross-validation procedure was implemented using the optimized model configuration. Data partitioning was performed at the image level. First, 10% of the total images were isolated as a fixed, independent test set, which remained unseen during the training and validation processes. The remaining 90% of the dataset was then partitioned into ten subsets (i.e., folds). The model was trained and evaluated ten times; in each iteration, one fold was used for validation while the remaining nine folds were used for training. While the data splits were image-based, the distribution of rip current instances remained highly consistent (Figure 2). Furthermore, background images lacking annotations comprised ~16% of the data across all splits, providing the model with negative examples to help minimize false detections. This methodology ensured that each run was executed using a balanced and representative data sample.”
- Updated crossvalidation results on the test set -
Precision Recall mAP50 mAP50-95 F2 mean±sd 0.875±0.01 0.841±0.01 0.921±0.00 0.669±0.00 0.848±001 - 5. Clarification on Single-Class Training and Inclusion of Background Images
We completely agree with the reviewer’s assessment: training a model to distinguish between "rip" and "doubt" as separate classes would cause severe class confusion due to their visual overlap; and introducing negative samples is critical to reduce false positives.
(i) Clarification on the 3-Class Training: We did not train the model to predict three separate classes. In all of our experiments, the models were trained using a single, unified "rip current" class. The "rip", "sediment", and "doubt" tags exist strictly as metadata within the open-source dataset. We included these tags to give future users and researchers the flexibility to filter the dataset for more permissive or conservative training regimens. For example, in our own experiments, the difference between the "permissive" and "conservative" configurations simply dictated whether the "doubt" and "sediment" annotations were merged into the single unified target class, or if they were ignored and treated as background. To prevent this misunderstanding for future readers, we propose adding clarifying paragraphs to the pertinent sections (detailed below).
(ii) Inclusion of background Images (without rip currents): As detailed in our response to Comment 5 (and in lines 110-111 of the original manuscript), our dataset and training splits do, in fact, include these. Approximately 16% of the images in every data split are pure background scenes containing no annotations. We have now made this explicit in the 2.4.5 methodology section, as noted in our previous response.
Proposed addition for the revised manuscript (2.1 RipAID training dataset):
“This three-label classification ('rip', 'sediment', 'doubt') is provided strictly as dataset metadata to grant future users the flexibility to apply permissive or conservative filtering. Training models to predict these classes separately is discouraged due to their high visual and conceptual similarity, which would likely induce severe class confusion”.Proposed addition for the revised manuscript (2.4.3 Annotation class configuration section):
“The models were never trained to predict multiple distinct classes; instead, all selected annotations were mapped into a single unified "rip current" class”.- 6. Mitigation of Deployment Risks and Clarification of Module Intent
Our primary goal in packaging this detector as an open-source module was to bridge the gap between academic research and deployment by providing an operative framework. However, as stated in our original manuscript, the system is explicitly designed to function as an advisory support tool—not as an autonomous "absolute truth" or a replacement for human judgment. We noted in the original text:
"Ultimately, its greatest potential lies in the adoption of the module within coastal safety frameworks as a support tool for lifeguards, emergency services, or applications that bridge beach monitoring cameras with end-users. In this role, the system could serve to reduce rip current risk by drawing attention to potential hazards, maximizing coverage while retaining human verification to ensure reliability."
However, we recognize that releasing the module inherently carries the risks highlighted by the reviewer. To ensure we are promoting responsible deployment, we have reviewed and tempered the language in both the Discussion and Conclusion sections. We will add explicit disclaimers emphasizing that this module represents a "first step" baseline requiring local calibration and human-in-the-loop verification, rather than a definitive, standalone safety guarantee.
Proposed text modification (line 382-384) to the Discussion:
“Despite these promising results, it is imperative to approach the operational deployment of this module with caution. In high-stakes domains, full automation is often not desirable due to safety, ethical, and legal concerns (Lai et al., 2023). Given the inherent complexities of coastal morphodynamics and the current volume of training data, the delivered system is not intended to function as an autonomous, standalone decision-making tool. Rather, it is designed strictly as a 'human-in-the-loop' advisory system. Its purpose is to flag potential hazards for human verification, rather than replacing human oversight. In this context its greatest potential lies in the adoption of the module within coastal safety frameworks as a support tool for lifeguards, emergency services, or applications that bridge beach monitoring cameras with end-users while retaining human verification to ensure reliability."
Proposed adjustment to the Conclusion (line 401):
"By releasing the complete codebase and pre-trained weights, we provide a foundational framework to democratize access to advanced safety tools. However, we emphasize that this module represents an initial baseline rather than a definitive fit-for-all solution. Safe operational deployment requires rigorous local benchmarking and strict integration into existing coastal safety protocols where human verification remains the ultimate authority."
- 7. Comparison with Existing Benchmark Datasets and Justification of Dataset Scale
We agree that a more explicit comparison with existing dataset-benchmark papers would strengthen the manuscript. However, direct comparison of model performance metrics is challenging due to fundamental differences in annotation types (axis-aligned bounding boxes, oriented bounding boxes, and segmentation masks) and evaluation protocols across datasets. To the best of our knowledge, no prior rip current dataset employs oriented bounding box annotations, further limiting direct comparison. However, to provide a more detailed introduction to the current dataset-benchmark papers, we propose the following changes:
Proposed addition to the Introduction (line 84):
“Several prior studies have also published their training datasets alongside their results, providing a basis for comparison. Maryan et al. (2019) constructed a dataset of 514 rip channel instances from time-averaged aerial imagery. De Silva et al. (2021) released a dataset of 2,440 images, the majority consisting of top-down Google Earth imagery, with 700 background images, and supplemented the test split with 23 video sequences. Building on this, Zhu et al. (2022) incorporated 1,352 additional real beach scene photographs (746 containing rip currents and 606 without), yielding a combined dataset of 3,792 images. All of these datasets employ axis-aligned bounding box annotations. More recently, Dumitriu et al. (2025) released a large-scale, multi-angle benchmark comprising 184 videos with 212,328 annotated frames (163,528 containing rip currents), using pixel-level segmentation masks.”
Proposed modification to the Discussion (line 355):
“Direct comparison between the results of the proposed model and existing state-of-the-art rip current detection models is challenging due to fundamental differences in annotation types and evaluation protocols across datasets. To the best of our knowledge, no prior rip current dataset employs oriented bounding box annotations, further limiting direct comparison. In terms of dataset scale, the RipAID v2.0.0 includes 6,789 training images, 2,815 of which are new images from 8 different camera angles of low-energy Mediterranean beaches, and 3,974 reprocessed images from previous datasets. The current version of RipAID includes more rip current instances than previously reported image-based datasets (Maryan et al., 2019, Zhu et al., 2022), but fewer than video-based datasets (De Silva et al., 2021, Dumitriu et al., 2025). However, by sampling independent images (at minimum one-hour intervals), we mitigate the temporal correlation and information redundancy that typically exist between consecutive frames in large video-based datasets.
Our model generalizes to the oblique viewing angles typical of shore-based cameras, contributing to address previous concerns in the literature (de Silva et al., 2021; Zhu et al., 2022; Rampal et al., 2022; Rashid et al., 2023), and aligning with most recent developments in the field (Khan et al., 2025a, Dumitriu et al., 2025). Furthermore, by opting for OBBs, the system captures orientation effectively without the cost of pixel-level masking as in instance segmentation approaches (Dumitriu et al., 2025), easing scalability for future dataset generation.”- 8. Revision of "Advanced Modalities" Terminology
We agree that the term used is inaccurate. We propose rephrasing as:
"Most recently, research has focused on increasing precision through dense mask annotation."- 9. Addressing General Comments: Target Audience, Level of Detail, and Adjustment of Claims
We sincerely thank the reviewer for this comprehensive and constructive overview of our manuscript. We align with the framing of this work as a "valuable checkpoint" rather than a final result; this captures our true intent, and we have adjusted the tone of the manuscript to better reflect this reality.
Regarding the omission of the annotation procedure, we have addressed this by including the detailed protocol in the text, as outlined in our response to previous comments.
Regarding the inclusion of basic deep learning concepts (such as explaining True Positives, False Positives, etc.): this was a deliberate choice tailored to the wide readership we expect from NHESS (including marine researchers, geomorphologists, and coastal emergency managers). Because a core contribution of this paper is the release of an open-source, Dockerized inference module meant for public use, we felt it was important to ensure the evaluation metrics were accessible to non-experts in machine learning.
We acknowledge the reviewer’s critique that the manuscript aimed for claims that were too large for the current volume of data and the stage of the research. Our core novelty lies in taking the first steps toward operationalization, providing the first OBB dataset, open weights, and a containerized API that the community can use and expand upon over time. To ensure the manuscript reflects this "checkpoint" reality and avoids overstating our achievements, we have reviewed the Discussion and Conclusion sections and removed or toned down any overly strong claims. Besides other changes addressed in previous comments in this line, some other modifications include: (i) removing the phrase “(in line 375)......establishing a new baseline for operational readiness” and simply stating that the work "provides an open-source foundational framework for future operational development."; (ii) modifying the phrase "(in line 376)...designed for immediate integration into multi-platform observation networks." by removing "immediate”, acknowledging that while the Dockerized packaging makes it available, actual integration into external operational networks might require site-specific calibration.
We believe these revisions, prompted by the reviewer’s valuable feedback, will result in a more scientifically rigorous, prudent, and well-grounded manuscript. We sincerely thank the reviewer again for the time, expertise, and guidance in improving this work.
-
AC1: 'Reply on RC1', Jesús Soriano-González, 15 May 2026
-
RC2: 'Comment on egusphere-2026-1138', Anonymous Referee #2, 30 Jun 2026
Overall Evaluation
This manuscript advances the field by offering a comprehensive dataset, an improved algorithm, and an open-source, containerized inference tool for automated rip current detection. However, it currently falls short in validating novelty quantitatively, performing detailed ablation analyses, interpreting hydrodynamic nuances, and delivering a polished presentation. Addressing the issues above will substantially enhance the scientific rigor and practical impact of the work.
I recommend major revision focused on these points before the manuscript is suitable for formal publication.
General Comments (Major Issues to Address)：
- Insufficient Quantitative Comparison with State-of-the-Art (SOTA) Models
The manuscript acknowledges limitations of prior rip current detection methods but does not provide a direct, quantitative benchmark against published SOTA models (e.g., RipFinder, RipScout, YOLO-Rip, segmentation-based RipVIS) on the unified RipAID dataset. Without these comparisons, the actual performance gains offered by the proposed YOLOv11n-OBB model remain unclear. I recommend implementing or re-evaluating existing SOTA approaches under the same protocol to provide fair and transparent performance metrics. - Experimental Design & Methodology Deficiencies
- Class configuration ablation lacks depth. The paper compares permissive vs conservative labeling but does not test a multi-class classification setup (separately predicting rip current, sediment, doubt as independent target classes). The authors only justify single-class training for safety-oriented high recall, but no quantitative evidence is provided to prove multi-class detection would degrade hazard warning performance.
- Hyperparameter optimization yields negligible performance improvement (<1%). The discussion of Optuna tuning is superficial; the authors do not analyze why default YOLO hyperparameters are nearly optimal for rip imagery, nor explain the physical meaning of key optimized augmentation parameters (e.g., reduced scale, adjusted HSV brightness) for coastal lighting variations.
- Cross-validation only reports aggregated average metrics without stratified analysis by imaging platform (UAV, shore camera, smartphone, satellite). It remains unknown whether the model generalizes equally well across all four data sources, which is the core selling point of the multi-platform dataset RipAID v2.0.0.
- Inference hardware tests only report FPS values without latency breakdown (preprocessing / model inference / post-processing). For edge coastal monitoring practitioners, separate latency components are critical for system scheduling design.
- Structural & Manuscript Organization Problems
- Severe content redundancy: Background introduction repeatedly restates the limitations of traditional image processing and early deep learning methods; the Discussion section repeats many quantitative results already presented in Section 3 without deeper mechanistic analysis.
- Imbalanced section length: The Methods chapter is overly bloated with detailed descriptions of standard off-the-shelf networks (Faster R-CNN, RT-DETR, FCOS, RetinaNet) that are widely known in computer vision. These lengthy architecture introductions waste space that could be allocated to hydrodynamic analysis, SOTA comparison, and ablation studies specific to rip currents.
- Practical Limitations & Future Work Are Underdeveloped
- The paper acknowledges optical observation limitations (low visibility, lack of surface sediment plumes) but proposes no mitigation strategies. The authors should discuss multi-modal fusion opportunities (combining wave sensor data, bathymetry, and optical imagery) to overcome pure RGB detection drawbacks.
- Dataset bias is insufficiently discussed: RipAID v2.0.0 heavily samples Mediterranean low-energy beaches; the generalization ability to high-energy rip systems (Australian, American surf zones) is not evaluated or discussed.
- The self-refinement loop (model-assisted semi-automatic annotation) is mentioned briefly but lacks a concrete workflow design and preliminary experimental proof-of-concept results.
- Language, Formatting, and Terminology Consistency
- There are minor but consistent issues with tense usage, wordiness, article/preposition errors, and inconsistent citation styles that require thorough copyediting.
- Terminology such as “rip current” vs. “rips” and phrases like “seaward flows” should be standardized and aligned with coastal engineering conventions.
- Long, complex sentence structures should be simplified for clarity, benefiting interdisciplinary readers.
-
AC2: 'Reply on RC2', Jesús Soriano-González, 03 Aug 2026
We thank the reviewer for this thorough and constructive evaluation of our manuscript, and for the detailed attention given to both the model's scientific contribution and its practical deployment aspects. The comments have helped us identify several areas where additional analysis, clarification, and rephrasing can improve the manuscript. Below, we provide a detailed, point-by-point response to each of the comments raised.
- 1. Comparison with SOTA models:
We agree that a direct, quantitative comparison against published SOTA models would provide additional context for interpreting the reported performance metrics. We would like to clarify, however, the specific benchmarking scope of this study: the architectural comparison performed here (Section 2.4.1–2.4.2) is designed to isolate the effect of annotation geometry (OBB vs. axis-aligned) under a common protocol (dataset, architectures). Most published rip current detection models (e.g., RipFinder, RipScout, YOLO-Rip) use axis-aligned annotations, on different datasets and under different evaluation protocols; therefore a direct numerical comparison against these, would reflect differences in data, protocol, and annotation geometry simultaneously.
As noted in our response to RC1, we intend this framework to serve as an evolving, open tool rather than a static benchmark. To address this point directly, we have revised the Discussion to acknowledge this as an open direction.
Proposed addition to the Discussion (line 374):
"This study was designed to isolate the effect of annotation geometry under a unified evaluation protocol, rather than to directly reproduce the datasets and conditions of previously published rip current detection models. Nevertheless, direct quantitative comparisons against established state-of-the-art approaches represent a logical next step for future research and module evolution, particularly as standardized, multi-site open datasets expand in scale and diversity."
- 2. Experimental Design:
(i) Class configuration:
We appreciate the reviewer's observation regarding the “class-configuration” experimental design (conservative vs permissive). The fact that both reviewers (RC1, Comment 5, and RC2) raised questions surrounding the boundary dynamics between 'rip' and 'doubt' demonstrates the importance of this topic and highlighted the need for us to clarify our underlying methodological rationale.As defined in Section 2.1, the 'doubt' category does not represent a visually distinct, well-defined target class; rather, it reflects explicit label ambiguity at the decision boundary where features cannot be confidently separated by expert annotators. Training a deep learning model to treat 'doubt' as an independent target class forces the algorithm to construct artificial decision boundaries around label uncertainty. Treating boundary noise as a learnable concept inherently degrades feature separability for the primary class and undermines model generalization (Almeida et al., 2021).
This reasoning is consistent with the concern raised in RC1, and commented on point 5 of the RC1-Response, which identified the visual similarity between 'rip' and 'doubt' as a training risk. The doubt and rip class can be visually similar; the model understands "doubt" as a completely different class and thus may be confused by the visual similarities. For these reasons, we consider the single-class strategy with a safety-oriented threshold a theoretically and empirically justified design choice rather than an open experimental question. To clarify this methodological rationale further, we propose expanding Section 2.1 of the manuscript, following the specific text additions proposed in our response to Comment 5 of RC1.
(ii) Hyperparameter optimization:
We agree that extending the discussion on hyperparameter optimization discussion enriches the manuscript's methodological insights, specifically by addressing: (i) why the optimization converged close to YOLOv11's defaults, and (ii) what the small deviations that did occur imply about the RipAID dataset's variability.
Proposed addition to the Discussion (line 336):
"Hyperparameter tuning yielded negligible performance variability (<1%) between the optimized and default configurations, as Optuna consistently converged close to YOLOv11's default settings. This suggests that these defaults, already tuned for general-purpose object detection using massive training datasets, transfer well to rip current detection without requiring task-specific adjustment. Among the evaluated parameters, the distribution focal loss (dfl) demonstrated the greatest influence on hyperparameter sensitivity (Table A5); its optimized value favoured a lighter localization penalty, which could relate to the inherent boundary ambiguity of rip current annotations (Section 2.1) rather than a need for finer box regression. Furthermore, parameters governing spatial and color augmentations (e.g., scale and saturation) either remained near default values or shifted toward lower intensities. This suggests that the native environmental diversity, multi-platform perspectives, and variable lighting already present within the RipAID dataset reduce the necessity for aggressive synthetic data augmentations. While these parameter shifts may offer methodological insights, the resulting gains are too marginal to justify the computational expense. Based on our experience, extensive hyperparameter tuning on highly pre–tuned architectures like YOLOv11 offers low investment return for this task unless substantially larger computational resources are allocated (e.g., hundreds of additional trials)."(iii) Validation by imaging platform:
To address the comment regarding model performance across different imaging platforms, we computed test-set performance separately per original data source, which serves as a direct proxy for the imaging platform, since each source corresponds to a distinct acquisition system or viewing geometry.
Proposed addition to the Results (Section 3.4, following Table 2):
“Performance variability across imaging platforms (F2 ranging 0.877 – 0.997; Table X) reinforces the generalization capabilities of the model. However, performance is consistently lower for oblique imagery (smartphones, shore cameras) than for top-down imagery (satellites, UAV), suggesting that it is primarily influenced by viewing geometry rather than platform type. This can be attributed to the greater background complexity inherent in oblique perspectives, the increased heterogeneity of oblique scenes, as well as the perspective distortion that causes target features to vary significantly in scale and appearance depending on their distance from the camera.
Table X. Test-set performance stratified by original data source (platform).”Source Platform Precision Recall mAP50 mAP50-95 F2 De Silva el al. (2021)
Satellite / aerial (top-down) 0.985 1.000 0.995 0.769 0.997 Khan et al. (2025b)
UAV (top-down) 0.980 0.982 0.993 0.853 0.982 CnE UFSC (2023) Smartphone (oblique)
0.861 0.881 0.948 0.616 0.877 Soriano-González et al. (2025) Coastal camera (oblique) 0.864 0.925 0.958 0.644 0.912 (iv) Inference hardware tests:
Our main rationale for prioritizing total inference time is that the socib-rip-currents-detection module is distributed as a self-contained, Dockerized inference module. From the user's perspective, the system is executed through a single API call, making the total processing time the most representative metric for operational deployment. Nevertheless, following the reviewer's suggestion, we profiled the complete processing pipeline by separating the execution into pre-processing, model inference, and post-processing stages on both CPU and GPU hardware in different machines. The results (Table below), confirm that model inference is by far the dominant computational component, indicating that the reported end-to-end FPS is largely governed by the neural network inference itself rather than auxiliary operations.System Hardware Breakdown (average per image) Laptop i7-13700H with RTX3050 cpu Pre-processing: 3.82 ms | 2.2%
Model Inference: 167.45 ms | 96.1%
Post-processing: 2.98 ms | 1.7%Laptop i7-13700H with RTX3050 gpu Pre-processing: 2.08 ms | 16.1%
Model Inference: 8.55 ms | 66.5%
Post-processing: 2.23 ms | 17.3%
Desktop i7 8700k with RTX5070 cpu Pre-processing: 1.27 ms | 2.6%
Model Inference: 46.40 ms | 95.7%
Post-processing: 0.81 ms | 1.7%Desktop i7 8700k with RTX5070 gpu Pre-processing: 1.08 ms | 10.9%
Model Inference: 7.23 ms | 73.0%
Post-processing: 1.60 ms | 16.1%Workstation Intel Xenon w5 with RTX5000 ADA cpu Pre-processing: 1.31 ms | 3.7%
Model Inference: 33.31 ms | 94.0%
Post-processing: 0.81 ms | 2.3%Workstation Intel Xenon w5 with RTX5000 ADA gpu Pre-processing: 1.23 ms | 17.2%
Model Inference: 4.60 ms | 64.3%
Post-processing: 1.32 ms | 18.5%We propose to add the following paragraph in the results section (section 3.5 in the original manuscript) informing the user of the computational profile of the proposed pipeline while maintaining end-to-end throughput as the primary deployment-oriented performance metric:
“[...] Furthermore, the architecture supports near real-time deployment, achieving inference speeds ranging from 2.4 to 60 FPS depending on the specific hardware configuration (Table~\ref{tab:speeds}). The reported throughput is largely dominated by neural network inference, which represents approximately 64–73% of the end-to-end runtime on GPU hardware and 94–96% on CPUs, with preprocessing and post-processing introducing comparatively small overheads (2–19%).”
- 3. Manuscript structure:
(i) Content redundancy:
Instances of limitations redundancy from Introduction in the Discussion were already indirectly corrected while addressing comments 2 and 7 of the RC1. Regarding the restatement of limitations in the Background/Introduction itself, we believe these passages describe distinct constraints specific to each method family, rather than repeating the same argument in different words. We have revised the text to eliminate identified redundancies.To address the remaining repetitions of Results in the Discussion section, the following changes are proposed:
Proposed sentences for removal (between lines 298-312):
“[...] FCOS, Faster R-CNN, and RetinaNet demonstrated clear underperformance compared to [...] with the latter two achieving comparable overall scores (Fig. 4). [...] used for YOLO and RT-DETR, they nonetheless align with findings in the broader literature [...] This geometric adaptation yielded an average improvement of 12.8% across all metrics compared to axis-aligned counterparts, independent of the YOLO model version or size. Notably, [...].”
Proposed modifications (line 339-346):
“To integrate the system into coastal monitoring networks, however, accuracy must be matched by computational efficiency. The developed module utilizes the lightweight YOLOv11n-OBB architecture (2.7 M parameters) maintaining real-time-capable inference speeds across all tested hardware configurations, from edge devices to GPU workstations (Table 3). This efficiency is maintained even in spatially complex scenes. Figure 8e shows the model resolving multiple, closely spaced rip current instances within a single frame, while Figure 8f demonstrates the correct detection of rip currents at varying distances from the camera. These results confirm that the module is sufficiently computationally lightweight to be deployed on edge devices and for incorporation into high-frequency alert systems without creating bottlenecks.“(ii) Imbalanced sections length:
We agree that the descriptions of the architectures in Section 2.2 are more detailed than necessary, particularly for readers already familiar with these widely used networks, while remaining somewhat inaccessible for readers who are not. We have compressed these descriptions accordingly, retaining only the single defining characteristic of each architecture, its relevant prior use in rip current detection where applicable, and its parameter count, which is directly relevant to the subsequent discussion on model size versus performance (Section 3.2, Fig. 6).
Proposed modified Section 2.2 (Deep learning methods):
"To detect rip currents, the following networks were evaluated:
- Faster Region-based Convolutional Neural Network (Faster R-CNN) (Ren et al., 2017) is a two-stage detector previously used in rip current detection (de Silva et al., 2021). Our implementation adopts the architectural enhancements proposed by Li et al. (2021), resulting in a model size of approximately 43.7M trainable parameters.
- The You Only Look Once (YOLO) family (Redmon et al., 2016) represents the state-of-the-art in real-time, one-stage object detection across various research domains such as medicine, agriculture, and industry (e.g., Khan et al., 2024; Jayagopal et al., 2024; Lv et al., 2023), and has also been used in the framework of rip current detection (Zhu et al., 2022; Rashid et al., 2023; Dumitriu et al., 2023). In this study, the recent OBB-supported iterations YOLOv8 (Jocher et al., 2023a), YOLOv11 (Jocher and Qiu, 2024), and YOLOv12 (Tian et al., 2025), implemented via the Ultralytics library (Jocher et al., 2023b), were evaluated across model scales ranging from Nano (n) to Extra-Large (x), with the smallest variants containing as few as 3.5M parameters.
- The Fully Convolutional One-Stage object detector (FCOS) (Tian et al., 2019) is an anchor-free detector that predicts bounding boxes directly at each feature-map location, eliminating the need for region proposals or predefined anchor boxes. This architecture contains 32.3M trainable parameters.
- The Real-Time Detection Transformer (RT-DETR) (Zhao et al., 2024) is a real-time, end-to-end object detector based on the Transformer architecture introduced in Carion et al. (2020). We utilized the Ultralytics implementation, evaluating the Large (RT-DETR-L, 42M parameters) and Extra-Large (RT-DETR-X, 74M parameters) variants (Jocher et al., 2023b).
- RetinaNet (Lin et al., 2020) is a single-stage detector that addresses the foreground-background class imbalance inherent to one-stage detection via Focal Loss. The version used in this work incorporates the improvements proposed by Zhang et al. (2020), resulting in a model with 38M parameters."- 4. Practical limitations and future work:
(i) Optical observation limitations and mitigation strategies:
We agree with the reviewer that discussing mitigation strategies for optical detection limitations would strengthen the manuscript's outlook. To address this, we have outlined a multi-modal fusion strategy that leverages the synergy between optical detection and modeling.
Proposed addition to the Discussion (around current line 364):
“To mitigate the inherent limitations of optical detection, future operational systems could explore multi-modal data fusion. Given that continuous in-situ sensor deployments (e.g., current meters, AWACs) are financially and logistically unscalable, we propose a synergistic approach using the image-based rip current detection module to generate a comprehensive database of rip current occurrences. This derived data can then serve as ground truth to train and calibrate statistical or machine-learning forecasting models driven by physical predictors, such as regional wave conditions, tidal elevations, and bathymetric data (Casper et al., 2024; Castelle et al., 2025). Because these forecast models would run entirely on physical parameters, this approach contributes to mitigate the inherent limitations of pure RGB imagery while capitalizing on its historical data. Simultaneously, it offers a scalable pathway to extend rip current hazard prediction to coastal areas lacking continuous camera monitoring."
(ii) Dataset bias discussion:
We acknowledge that generalization to high-energy rip systems (e.g., Australian or American Pacific surf zones) is not directly evaluated in the current manuscript. However, two of the four data sources composing RipAID do originate from comparatively higher-energy, non-Mediterranean environments: CnE UFSC (2023), sourced from Santinho and Moçambique, Atlantic-exposed surf beaches in Brazil, and Khan et al. (2025b), sourced from California. The per-source breakdown introduced in response to a previous comment (2-iii), shows that the model maintains strong F2 performance on both of these sources (0.877 and 0.982 respectively), providing partial evidence that detection performance is not strictly tied to the low-energy Mediterranean conditions that dominate the dataset by volume.However, we agree this evidence remains limited, as it does not extend to extreme high-energy conditions. Moreover, no site-level wave energy metadata is currently available in RipAID to support a more systematic energy-stratified analysis. Therefore we retain this as an explicitly acknowledged limitation rather than a validated capability.
Proposed addition to the Discussion (line 368):
"While RipAID v2.0.0 predominantly features imagery from low-energy beach conditions, the model maintains strong detection performance across geographically distinct coastal typologies, including sites in Brazil and California (Table X). This demonstrates that the model's applicability extends beyond a single environmental setting. However, its generalization to highly energetic or extreme surf conditions remains untested. Consequently, further testing by end-users and expanding the dataset to include imagery from high-energy environments is needed for deploying the model in such dynamic regimes."
(iii) Self-refinement loop workflow:
We appreciate that providing a conceptual workflow for model-assisted annotation can help readers better understand how the proposed detector could support future dataset expansion. Accordingly, we propose expanding the “Future Work” in the discussion section to describe a semi-automatic annotation strategy, as follows:
“[...] Future work could take advantage of the model to enhance and expand the training data, with a particular focus on prioritizing the inclusion of underrepresented and ambiguous cases. To achieve this, a human-in-the-loop annotation workflow is proposed utilizing self-hosted instances of CVAT, the same annotation platform employed to generate RipAID, which supports the integration of custom AI models. Under this framework, newly acquired coastal imagery is first processed in batch mode by the trained detector to generate candidate OBB annotations. These predictions are then reviewed, corrected, or rejected by expert annotators before being incorporated into the dataset. This strategy allows the annotation effort to focus on refining the generated labels rather than producing them manually from scratch. Alternatively, dataset expansion could be performed outside the annotation platform via a fully automated workflow, such as by accepting only high-confidence detections without manual verification. However, such an approach would inherently favor the most straightforward or predominant examples, rendering it less effective at capturing the underrepresented or ambiguous cases necessary to improve model robustness. By contrast, the proposed human-in-the-loop workflow preserves annotation quality while systematically expanding the diversity and representativeness of RipAID. Furthermore, this method can be readily adopted or adapted by external users wishing to extend the dataset with independent imagery.”
We believe the proposed framework and workflow provide a practical roadmap that the reader could leverage.
- 5. Language, Formatting, and Terminology Consistency
We thank the reviewer for the careful reading and detailed list of observations. This feedback helped us resolve several minor inconsistencies and formatting details across the manuscript. While the complete set of revisions will be included in the finel manuscript, some examples of these corrections are shown below:
(i) Minor but consistent issues with tense usage, wordiness, article/preposition errors, and inconsistent citation styles as in double parenthesis in “ (RipAID v2.0.0; (Soriano-González et al., 2026))”; chronological order in "(CnE UFSC, 2023; Soriano-González et al., 2025; Khan et al., 2025b)"; use of present tense in section 3.1 instead of the past tense used through the whole section; article/preposition errors such as in “ [...] with inference speeds; [...] where YOLO outperformed RT-DETR margin of [...]”; the mixed use of crossvalidation and cross-validation, etc.
(ii) Standardization of the informal shorthand "rips" to "rip currents" throughout the manuscript for terminological consistency. All instances of "rips" used as a standalone noun will be replaced with "rip currents". We also acknowledge that the description of rip currents as "fast-moving seaward flows" is inconsistent with the more detailed definition given in the Introduction (line 15-16). We propose the following modification to the Abstract (line 2):
[old] "Rip currents are hazardous, fast-moving seaward flows and remain one of the leading causes of rescues and drownings on surf beaches"
[new] "Rip currents are hazardous, fast-moving, jet-like flows of water extending from the shoreline to beyond the surf zone, and remain one of the leading causes of rescues and drownings on surf beaches"
(iii) Simplification of long, complex sentence structures:
We propose the following changes to break down such sentences into shorter, single-idea statements:
Proposed modifications to introducción (line 28-32):
"While lifeguards are trained to identify these hazards, their coverage is often limited to specific seasons, hours, and locations. This creates periods of vulnerability, especially outside of peak times or in less-frequented areas."
Proposed modifications to introducción (line 35-38)
"While effective, field deployment in the surf zone is logistically difficult, expensive, and often restricted in crowded beach environments. These methods also require significant technical expertise and are typically applicable only to highly localized instances in time and space (Pitman et al., 2016). As a result, they remain limited to experimental scenarios."
Proposed modifications to 2.4.4 Hyperparameter optimization (line 190-194):
"In this study, tuning was performed to maximize the rip_fit self-defined metric (Table 1) using Optuna, an open-source hyperparameter optimization framework (Akiba et al., 2019). Optuna uses a smart sampling strategy that analyzes past trials to discard unpromising ones early during training (i.e., pruning). This significantly reduces the overall optimization time while focusing resources on the most promising configurations."
Proposed modifications to Discussion (line 331-338):
"In such tasks, where human agreement can be low, the model may effectively outperform human annotators by learning more stable and consistent patterns than the human labels themselves. This underscores the potential to enhance ground truth quality. It also suggests that the deployed model can serve a dual purpose: not only as a detection system, but also as a semi-automated dataset refinement or expansion tool."
Overall, addressing RC2's comments has prompted several key analytical additions that substantially broaden the empirical scope of this work. Beyond clarifying our experimental design, we have expanded the manuscript to include a deeper discussion on hyperparameter tuning, a platform-stratified benchmark across diverse imaging geometries, the framework latency profile, and an explicit delineation of operational scope and future applications. We believe that these targeted revisions ensure a more rigorous, transparent, and comprehensive presentation, and we sincerely thank the reviewer for the rigorous and valuable feedback, which significantly elevated the quality of this manuscript.Citation: https://doi.org/10.5194/egusphere-2026-1138-AC2
- Insufficient Quantitative Comparison with State-of-the-Art (SOTA) Models
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 641 | 456 | 55 | 1,152 | 50 | 66 |
- HTML: 641
- PDF: 456
- XML: 55
- Total: 1,152
- BibTeX: 50
- EndNote: 66
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General comments:
The paper introduces a fair idea, arguing that OBB facilitates less background noise compared to axis-aligned bounding boxes, with some loss in accuracy but at a fraction of the cost compared to segmentation. It introduces 6,789 images with rip currents annotated with oriented bounding boxes.
Specific comments:
The paper's main argument is a reasonable one, rip currents benefiting from OBB detection. There are several strong critiques that need to be addressed:
- the evaluation is done on classic detection metrics and a custom metric, rip_fit. While this is a reasonable approach, as per their own introduction, rip currents are a leading cause of drownings globally. For such scenarios where a false positive is costly, the F2 is a standardized metric that can be used (alongside currently used ones). Introducing a custom metric risks fragmenting the research even more, with minimal benefits, if any;
- the dataset, with 6,789 is rather modest in size. One of the cites sources is an roboflow link which is no longer available (CnE UFSC (2023)) and another source, RipScout (2025), claims 2555 images from 73 videos; upon closer look, RipScout introduced 2555 images from 8 videos, with highly repetitive patterns, alongside 1767 images from de Silva (2021); this raises a doubt over the claimed source and diversity of the dataset, especially when compared to de Silva (2021). De Silva (2021) trained on 2440 images (1767 with rip currents and 673 without) and they tested on 23 videos. Dumitriu (2025) introduced 150 rip currents videos with 15,784 annotated frames, alongside 34 videos with no rip currents. The conclusion states that "a key contribution is the development [...] a novel dataset that significantly expands the diversity of the available training data [...]". Considering that more than 1/3 of the data is from previous papers and another source is not available anymore to analyze, this is questionable;
- there is little discussion on the annotation procedure. Subsequent studies (Ballantyne et al. (2005), Sherker et al. (2010), Caldwell et al. (2013), Brannstrom et al. (2014), Sotes et al. (2018), Pitman et al. (2021)) have found that while many people are confident that they can correctly identify rip currents, most of them cannot do so. While the surveyed people were not experts, such process needs to be thoroughly done and documented. If multiple annotators were used, an inter-annotator agreement coefficient is also recommended (such as Cohen-Kappa). For example, a rip current expert that we asked has said that at first glance, Figure 8.c is incorrectly annotated. The yolo detection on the right appear to be correct, while the annotation on the left is not clear. There are other instances from that camera that do have only rip current on the left, but these seem to be at a different time and date and do not necessarily generalize (based on simple visualization);
- 10-fold is indeed useful in scenarios with this volume of data. It is not clear, however, how it was done. The paper introduces 6,789 images with 10,131 annotations (I am assuming rip currents instances, which can be more per image) and proceed to do the 10-fold split on the rip currents and not on the images. It is both unclear how and why the authors have proceeded to do so;
- the way the authors approached k-fold cross-validation is incorrect. Please check https://scikit-learn.org/stable/modules/cross_validation.html for details. 10-fold should be done only on train and validation with a separate and fixed test data. The way the authors did it, the test data is unreliable and more akin to validation data. A correct way to do it is do an initial split between (train+val) and keep test fixed. Then you can do a k-fold split in order to find the best combination of train and val splits; once done so, you evaluate only once on test data and this becomes you final result. Modifying the training based on this result negates the value of the test data;
- splitting into 3 classes, "rip", "sediment" and "doubt" is risky from a training point of view. The doubt and rip class can be visually similar; the model understands "doubt" as a completely different class and thus may be confused by the visual similarities. I suggest an approach based on clear rip current classes (in this case, sediment is a large enough distribution to make sense) and judge the "doubt" or not based on model confidence, not as a separate class on its own. Especially since the annotations in itself are not discussed and can be questionable. This both confuses the model and risks introducing the strong bias of the annotators. They did compare 3 classes vs 1 class (line 317), which is reasonable, but in order to make an argument for the doubt class, it should be isolated in an experiment with 3 classes vs 2 classes and the results should be evaluated by a rip currents expert. The quantitative results are not enough in this context: dataset too small, no null (no rip) images in training and evaluation, class imbalance, annotation biases etc. They should also account for the class imbalance of the 3 classes and maybe use a different loss (such as weighted cross-entropy);
- training and evaluation in such cases needs to also consider images without rip currents, simiarly to previous rip current papers (deSilva (2021), Zhu (2023), Dumitriu (2025)) and to most papers in object detection field;
- considering the low volume of data and several questionable aspects, mentioned above, packaging a detector in "ocib-rip-currents-detection", while a noble and arguably required endeavor, poses multiple risks. Unknowing and non-technical people can use and rely on it with potentially grave consequences. A rip current detector, while extremely useful if done correctly, should be approached and published with meticulous detail, testing and care;
- the authors also do not compare directly (enough) to other dataset-benchmark papers. They do mention some of them, but a direct comparison (numbers, annotations, results etc) is warranted;
- the paper claims that recent advacements use "advancet modalities" (line 80). That is incorrect - all of the current papers use RGB images / videos. Advanced modalities in the context of computer vision implies non-traditional data sources beyond 2D RGB imaging;
- the paper is long and detailed, containing a bit of everything. In some places, the detailed approach is well done, while in others it either doesn't mention relevant information (such as the aforementioned annotation proceedure) or it goes into too much detail of basic information (such as explaining what TP, FP etc are). The paper should focus more on the novelty, the contribution and the results;
Overall, the paper starts researching in a direction with considerably potential, but aims for claims that are too big for the data, research and the effort that was put into it. I recommend this be a valuable checkpoint in the direction of such claims and target and not the final result.