Band Weight-Optimized BiGRU Model for Large-Area Bathymetry Inversion Using Satellite Images

Currently, using satellite images combined with deep learning models has become an efficient approach for bathymetry inversion. However, only limited bands are usually used for bathymetry inversion in most methods, and they rarely applied for large-area bathymetry inversion (it is important for methods to be used in operational environments). Aiming to utilize all band information of satellite optical image data, this paper first proposes the Band Weight-Optimized Bidirectional Gated Recurrent Unit (BWO_BiGRU) model for bathymetry inversion. To further improve the accuracy, the Stumpf model is incorporated into the BWO_BiGRU model to form another new model—Band Weight-Optimized and Stumpf’s Bidirectional Gated Recurrent Unit (BWOS_BiGRU). In addition, using RANSAC to accurately extract in situ water depth points from the ICESat-2 dataset can accelerate computation speed and improve convergence efficiency compared to DBSCAN. This study was conducted in the eastern bay of Shark Bay, Australia, covering an extensive shallow-water area of 1725 km². A series of experiments were performed using Stumpf, Band-Optimized Bidirectional LSTM (BoBiLSTM), BWO_BiGRU, and BWOS_BiGRU models to infer bathymetry from EnMAP, Sentinel-2, and Landsat 9 satellite images. The results show that when using EnMAP hyperspectral images, the bathymetry inversion of BWO_BiGRU and BWOS_BiGRU models outperform Stumpf and BoBiLSTM models, with RMSEs of 0.64 m and 0.63 m, respectively. Additionally, the BWOS_BiGRU model is particularly effective in nearshore water areas (depth between 0 and 5 m) of multispectral images. In general, comparing to multispectral satellite images, using the proposed BWO_BiGRU model to infer hyperspectral satellite images can achieve better bathymetry inversion results for large-area bathymetry maps.