Abstract

In the active–passive fusion-based bathymetry inversion method using single-temporal images, image data often suffer from errors due to inadequate atmospheric correction and interference from neighboring land and water pixels. This results in the generation of noise, making high-quality data difficult to obtain. To address this problem, this paper introduces a multi-temporal image fusion method. First, a median filter is applied to separate land and water pixels, eliminating the influence of adjacent land and water pixels. Next, multiple images captured at different times are fused to remove noise caused by water surface fluctuations and surface vessels. Finally, ICESat-2 laser altimeter data are fused with multi-temporal Sentinel-2 satellite data to construct a machine learning framework for coastal bathymetry. The bathymetric control points are extracted from ICESat-2 ATL03 products rather than from field measurements. A backpropagation (BP) neural network model is then used to incorporate the initial multispectral information of Sentinel-2 data at each bathymetric point and its surrounding area during the training process. Bathymetric maps of the study areas are generated based on the trained model. In the three study areas selected in the South China Sea (SCS), the validation is performed by comparing with the measurement data obtained using shipborne single-beam or multi-beam and airborne laser bathymetry systems. The root mean square errors (RMSEs) of the model using the band information after image fusion and median filter processing are better than 1.82 m, and the mean absolute errors (MAEs) are better than 1.63 m. The results show that the proposed method achieves good performance and can be applied for shallow-water terrain inversion.