From Land to Ocean: Bathymetric Terrain Reconstruction via Conditional Generative Adversarial Network

Abstract Acquiring global ocean digital elevation model (DEM) is a forefront branch of marine geology and hydrographic survey that plays a crucial role in the study of the Earth's system and seafloor's structure. Due to limitations in technological capabilities and surveying costs, large-scale sampling of ocean depths is very coarse, making it challenging to directly create complete ocean DEM. Many traditional interpolation and deep learning methods have been applied to reconstruct ocean DEM images. However, the continuity and heterogeneity of ocean terrain data are too complex to be approximated effectively by traditional interpolation models. Meanwhile, due to the scarcity of available data, training an sufficient network directly with deep learning methods is difficult. In this work, we propose a conditional generative adversarial network (CGAN) based on transfer learning, which applies knowledge learned from land terrain to ocean terrain. We pre-train the model using land DEM data and fine-tune it using ocean DEM data. Specifically, we utilize randomly sampled ocean terrain data as network input, employ CGAN with U-Net architecture and residual blocks to capture terrain features of images through adversarial training, resulting in reconstructed bathymetric terrain images. The training process is constrained by the combined loss composed of adversarial loss, reconstruction loss, and perceptual loss. Experimental results demonstrate that our approach reduces the required amount of training data, and achieves better reconstruction accuracy compared to traditional methods.