A hybrid neural network and interpolation framework for volume estimation of submarine landslide scars

Submarine landslide reconstructions are essential for understanding failure mechanisms and assessing geohazards. A key factor in this context is the initial failure volume, which strongly influences the tsunami generation potential of landslides. Accurate volume estimates require knowledge of the pre-failure morphology; however, this surface is typically not available and often approximated through extrapolation of the surrounding seafloor. We propose a hybrid reconstruction approach that combines conventional interpolation techniques with a neural network designed to refine them. Instead of replacing classical methods, the neural network builds upon their outputs as initial estimates and iteratively improves them by learning spatial and geomorphological patterns from the surrounding bathymetry. The model is trained in a self-supervised manner using masked bathymetric patches, enabling it to infer realistic pre-failure morphologies and enhance the accuracy of traditional reconstruction workflows. The method was evaluated using bathymetric datasets from different regions and resolutions, including randomly sampled EMODnet patches, a regional EMODnet subset offshore Crete (Greece), and two high-resolution multibeam datasets collected around Crete. Moreover, we analyzed how the reconstruction error changes with the distance from each masked grid cell to the nearest unmasked region. This revealed a consistent pattern that can be used to create a look-up table for estimating reconstruction uncertainty based on the spatial arrangement of missing data. Our findings demonstrate that neural networks provide a powerful improvement for reconstructing missing bathymetric data, reducing reconstruction deviation by up to 5 percentage points compared to the standard method. The approach is particularly valuable for large-scale reconstructions where conventional techniques face limitations.