Rational Utilization of Multitemporal Satellite Remote Sensing Data for Fully Satellite-Derived Bathymetry

Abstract

Existing multitemporal empirical satellite-derived bathymetry (SDB) methods, especially those that use both multitemporal multispectral satellite images and multiple-pass ICESat-2 global geolocated photon data (i.e., ATL03), either train the bathymetric model of each individual image in the multitemporal sequence in isolation, or generate digital depth models (DDMs) with unexplainable measurement times, or rely entirely on external tools to correct calibration data. To address such weaknesses, this study proposed a joint-training depth estimation method that can rationally and effectively utilize these two types of satellite remote sensing data for fully satellite-derived empirical bathymetry. The proposed method allows each individual image in the multitemporal sequence to have one bathymetric model for depth estimation, but all the models are jointly trained using the entire model calibration dataset. The joint training is implemented by obtaining least-squares estimates for unknown variables in all multitemporal model calibration equations. This method ensures that its trained models are all optimal in the least-squares sense, and that its resulting DDMs, including temporal DDMs and mean DDM, are temporally well-defined. It also automatically corrects ATL03 calibration depths to the in situ depth measurements during joint model training, representing a real novelty that, to the best of the authors’ knowledge, has never been reported in SDB studies. It also supports time-series bathymetric applications. This method was then validated in Oahu, Saint Thomas, and Saipan study areas using multitemporal Sentinel-2 MSI (multispectral instrument) images and ICESat-2 ATL03 data, resulting in root-mean-square error (RMSE) = 2.43 m and mean absolute error (MAE) = 1.49 m in Oahu, RMSE = 1.37 m and MAE = 0.97 m in Saint Thomas, and RMSE = 1.90 m and MAE = 1.26 m in Saipan. The validation also showed that the proposed method outperforms two typical existing multitemporal depth estimation methods, namely the straightforward separation method and the mean-fused method, in terms of bathymetric accuracy. Moreover, it remains preferred among the reported typical multitemporal empirical bathymetric methods when factors including temporal interpretability, time-series bathymetric ability, multitemporal collaborative utilization ability, and automatic calibration depth correction ability are comprehensively considered.