| 06 Nov 2025
Coastal process understanding through automated identification of recurring surface dynamics in permanent laser scanning data of a sandy beach
Abstract. Four-dimensional (4D) topographic datasets are increasingly available at high spatial and temporal resolution, particularly from permanent terrestrial laser scanning (PLS) time series. These data offer unprecedented opportunities to analyse rapid and complex morphological processes occurring in sandy coastal environments, such as sandbar welding or bulldozer activity, as well as their longer-term impacts on sandy beaches. However, studying these processes requires the extraction and recognition of recurrent topographical surface dynamics across time, which in turn demands novel, automated methods. This study presents a novel workflow that combines 4D objects-by-change (4D-OBCs) with unsupervised classification using Self-Organizing Maps (SOMs) and hierarchical clustering. Applied to a three-year PLS time series comprising 21,194 hourly point clouds, the method identifies 4,412 instances of short-term surface dynamics. These are organized into two SOMs (64 nodes each) and further grouped into 31 clusters representing distinct dynamic types, such as berm deposition, large-scale backshore erosion, and human interventions (e.g., bulldozer activity). The classification results enable detailed spatiotemporal analyses of coastal morphodynamics. The SOM topology reveals seasonal patterns in surface activity, where, for example, winter is dominated by erosional activity over the whole beach but depositional activity mainly occurs in the intertidal area. The broader clusters facilitate interpretation of environmental responses and identification of changes in cross-shore zonation of types of dynamics, like berm formation. This approach demonstrates the potential of integrating PLS and unsupervised learning to characterize complex surface dynamics, and provides a framework for targeted, data-driven investigation and prediction of morphodynamic processes in high-resolution 4D remote sensing datasets.
This paper introduces a highly novel approach to disentangling the complex processes driving short-term beach changes through an unsupervised, data-driven method. This is impressive work, as it enables the grouping of similar processes across both space and time, allowing researchers to make effective use of the vast amounts of PLS data. The proof-of-concept presented here is both important and interesting, offering a significant step forward in how we interpret high-frequency coastal dynamics. However, while the data-driven methodology is a strong asset, the manuscript would benefit from a more rigorous geomorphological justification for the specific thresholds and technical choices made during the analysis.
One area for improvement concerns the integration of environmental variables. While the authors claim to combine these variables with the 4D-OBCs, the current analysis remains largely qualitative, relying on manual comparisons and interpretations for only eight selected clusters. To avoid giving readers the false impression of a fully automated or integrated analysis, the abstract and introduction should be rephrased to accurately reflect this level of manual intervention. Furthermore, the authors should explore or at least highlight a bit earlier whether/how a more data-driven comparison, such as calculating direct correlations between clusters and environmental drivers. At the moment some discussion is performed solely towards the end of the manuscript.
The spatial relevance of the environmental data requires further clarification. The monitoring stations used appear quite distant from the PLS site, raising questions about whether these data streams remain spatially correlated with the local beach forms. The authors should also address whether the model needs to account for sediment availability, specifically considering inputs from along-shore transport or foreshore regions, which are critical drivers of geomorphic change. Finally, regarding data management, the authors suggest partitioning large datasets based on existing gaps. However, for continuous datasets where gaps are few or non-existent, a clearer strategy is needed; the authors might consider discussing whether a sliding window approach would be more appropriate than arbitrary partitioning to avoid artificial hard breaks in the process analysis.
More specific comments:
L. 117-118: It needs to be clearly demonstrated how the proposed method is transferable to other monitoring platforms and different spatial scales.
Fig. 2: Please add a scale bar to the map so that the reader can assess the actual distances between the meteorological stations and the PLS site.
L. 177-178: Could you clarify the reasoning behind the assumption that smaller scale changes are not considered geomorphological changes?
L. 180: Regarding the choice of a one-week averaging window, please justify this specific timeframe and address whether there is a risk of smoothing out significant short-term signals.
L. 182: It would be helpful to know the magnitude of the data gaps encountered; please provide a summary or descriptive statistics for these gaps.
L. 187: Please provide specific numerical values to define what is considered too large in this context.
L. 188: How does the model account for geomorphic processes that do not return to the original elevation, such as permanent erosion or dune formation? It might be beneficial to mention this briefly here, even if the authors expanded upon this in the discussion, to resolve potential reader confusion early on.
L. 195: Please explicitly define the parameter used for the homogeneity criteria.
L. 200: Why was an area of 10 m² selected, and how does this specific size relate to the scale of the expected geomorphic processes?
L. 236-238: Would it be possible to consider filtering the 4D-OBCs individually, such as filtering the margins of an object rather than only using internal variability for noise identification and subsequent removal?
L. 259-260: Might a centroid be more suitable for the analysis.
L. 271-275: I am afraid that this paragraph is unclear to me; would it be possible to scale or normalize the volumeTS to improve interpretability?
L. 308: Please specify how the weight vectors are initialized, e.g., are they generated randomly?
L. 302: What specific metric is used for the maximum dissimilarity sampling? Is it, e.g., based on Euclidean distance.
L. 320: When you mention radius are you referring to sigma?
L. 335-336: Could you confirm if you are using average linkage for the clustering? If so, did you also consider the Ward linkage approach, which might provide more balanced clusters by accounting for cluster variance?
L. 351-357: I am afraid that this entire paragraph is difficult for me to follow.
L. 362: Why was only a single cluster selected for manual investigation, and what were the specific criteria used to choose this one?
L. 364: Please clarify what is meant by “relative frequency by season".