Abstract: Objective　Shallow water bathymetry is a critical research focus for coastal environment management and protection. However, traditional passive remote sensing bathymetry techniques rely heavily on in-situ depth control points to build mathematical models for inversion, which severely limits their application in regions where field data is unavailable. While ICESat-2, equipped with the Advanced Topographic Laser Altimeter System (ATLAS), has a distinct advantage in acquiring reference bathymetric control points, its raw data contains substantial noise from solar background and system electronics. In addition, in terms of photon signal processing, conventional clustering methods like standard density-based spatial clustering of applications with noise (DBSCAN) method, which use fixed parameters, fail to adapt to the dynamic photon density distributions caused by varying underwater topography and water conditions. Therefore, this study aims to propose an adaptive elliptical density-segmentation method to efficiently and accurately extract high-quality ICESat-2 reference bathymetry points for large-scale coastal applications. Methods　The proposed methodology follows a structured workflow designed for efficient and high-accuracy signal detection. Firstly, a shallow-water feature photon extraction strategy is implemented to avoid processing entire ICESat-2 laser trajectories, thereby reducing computational overhead. This step involves pre-processing Sentinel-2 L1C images using the Sen2Cor plugin to generate L2A surface reflectance products, and then these reflectance products are resampled to 10 m resolution. The Normalized Difference Water Index (NDWI) is utilized to delineate island and coastline boundaries. Based on the intersection of ICESat-2 trajectories and these boundaries, specific photon segments are extracted. Specifically, for the intersections with land, a segment from 1000 m inland to 5000 m offshore is defined to ensure comprehensive coverage of the nearshore zone. Secondly, the core of the study is the adaptive elliptical density-segmentation algorithm. Unlike standard DBSCAN, this method adaptively calculates the optimal neighborhood radius and minimum point threshold by evaluating the Euclidean distance matrix and the average count of signal versus noise photons within M frames of the data segment. This elliptical approach is specifically adjusted to the spatial distribution of lidar photons, allowing the algorithm to dynamically adjust to topographical changes. Following signal extraction, Parrish's refraction correction method is applied to account for the geometric displacement caused by the air-water interface. Finally, the dataset is refined through an anomaly rejection process using wavelet filtering and K-medoids clustering to eliminate outliers and redundant data points. For a wider range of applications, these discrete reference points are integrated with Sentinel-2 multispectral bands (B1–B7) into a 50-layer deep neural network (DNN) to generate continuous bathymetric maps. Results and Discussion　The proposed method was applied to the ICESat-2 data over five islands in the Virgin Islands, including Anegada, Anguilla, St. Thomas, St. Croix, and Basseterre, and highly accurate results were obtained, which significantly outperform traditional fixed-parameter methods. When using NOAA Continuously Updated Digital Elevation Model (CUDEM) as a reference for comparative analysis over St. Thomas, it was shown that the adaptive elliptical densitysegmentation DBSCAN effectively suppressed noise photons that standard DBSCAN failed to remove. Quantitative verification showed that the adaptive method achieved a determination coefficient (R2) of 0.99 and a root mean square error (RMSE) of 0.48 m, whereas the standard DBSCAN exhibited a much higher RMSE of 1.79 m due to its inability to adapt to photon density fluctuations. Furthermore, the joint inversion of ICESat-2 reference points and Sentinel-2 images through DNN produced a continuous bathymetric map for St. Thomas with an R2 of 0.96 and an RMSE of 1.41 m when compared to in-situ data. Compared to the discrete ICESat-2 points, there was a slight increase in error due to the spatial resolution differences during the transition from linear trajectories to continuous grids. Across the five study sites, a total of approximately 159,400 high-precision reference bathymetry points were generated. Statistical distribution analysis revealed that Anegada, possessing the largest water area, contributed 68% of the total points, with a maximum depth of 45.28 m. On the other hand, Anguilla showed the highest proportion of valid reference points in deep waters exceeding 25 meters, indicating its superior water clarity and the algorithm's robustness in detecting deep-seated subsea signals. These results demonstrate the potential of the proposed adaptive batch-processing framework for large-scale maritime mapping. Conclusions　This research demonstrates that ICESat-2 ATLAS, when processed with adaptive algorithms, can provide a highly reliable source of reference bathymetry control points, which can effectively fill the data gaps in global coastal mapping. The proposed adaptive elliptical density-segmentation method overcomes the limitations of traditional clustering by dynamically adjusting to adapt to complex topographical and environmental variations, achieving a vertical accuracy (RMSE) of less than 10% of the maximum water depth. This study marks the first large-scale application of such an adaptive batch-processing method to the Caribbean islands, successfully generating a massive dataset of 159,400 reference points. By eliminating the reliance on expensive and logistically challenging in-situ measurements, this framework offers a scalable solution for nearshore bathymetric monitoring. The successful fusion of active spaceborne lidar and passive multispectral imaging confirms the feasibility of constructing accurate, high-resolution global bathymetric maps. Future research will aim to incorporate advanced machine learning and multi-temporal stacking techniques to further enhance signal-to-noise ratio and extend detection depth in more turbid coastal waters.

Key words: hydrographic surveying, ICESat-2, Sentinel-2, lidar