OceanVerse: Evaluable 4D Ocean Element Reconstruction Dataset under Realistic Sparsity

Abstract

The vast oceans record the impacts of climate change and human activities on the Earth system. Over the past century, oceanographic scientists have collected extensive ocean profile data to reflect variations of oceanic elements, such as dissolved oxygen. However, due to the sophisticated measurements and high costs, historical ocean element observation data remains highly sparse and uneven across the global ocean, with the annual missing rate exceeding 90%. Thus, quantitatively understanding the four-dimensional (4D) spatiotemporal evolution of oceanic elements continues to pose a significant challenge. Machine learning (ML) techniques demonstrate superior capabilities in perceiving spatiotemporal variations within large-scale data, presenting promising opportunities to harness implicit correlations for global reconstruction. However, fragmented data and interdisciplinary differences create barriers to the availability of AI-ready open data, further hindering ML practitioners from designing specialized models. To solve this problem, we present the first oceanic 4D sparse observation reconstruction dataset, named OceanVerse. By integrating nearly 2 million real-world profiles since 1900 and three differentiated Earth system numerical simulation, we construct a comprehensively evaluable dataset with missing patterns that align with real-world conditions through a digital twin sampling. OceanVerse provides a novel large-scale (~100× nodes vs. existing datasets) dataset that meets the MNAR (Missing Not at Random) condition, supporting more effective model comparison, generalization evaluation and potential advancement of scientific reconstruction architectures. The OceanVerse dataset and codebase are publicly available. The OceanVerse resource is available at https://github.com/jingwei-sjtu/OceanVerse.