"Metadata for A Mixed Convolution Neural Network for Bathymetric Prediction from Multi-source Gravity Anomalies"

"Mapping seafloor topography from gravity anomalies is a highly nonlinear inverse problem. While Convolutional Neural Networks (CNNs) are powerful nonlinear tools, their fixed kernel sizes cannot handle the spatially variable scales in the gravity-topography relationship. To this end, we propose a Mixed Convolutional Neural Network (MCNN) for predicting regional bathymetry. MCNN uses satellite gravity anomalies and limited shipborne bathymetry to model the nonlinear mapping from sea-surface gravity to short-wavelength topography. Multi-scale truncated convolutional kernels are employed for feature extraction, demonstrating superior generalization compared to single-scale convolution. We validate MCNN's effectiveness for bathymetric reconstruction through training and testing in three geographically distinct regions, benchmarking performance against \\cite{smith1994bathymetric}'s method (SAS), Gravity Geologic Method (GGM), Deep Neural Networks (DNN), and standard CNNs. To assess MCNN's regional transfer capability, critical for predicting bathymetry in uncharted areas, we conduct cross-regional transfer experiments. MCNN outperformed SAS and CNN baselines, exhibiting significantly enhanced transfer stability. Finally, we systematically analyze the impact of the truncated convolutional kernel size (a key MCNN hyperparameter) on inversion accuracy, contrasting its design with CNN's adaptive kernel performance. Experimental results confirm MCNN's robustness in delivering stable, high-accuracy predictions under sparse shipborne data constraints."