OceanDepths: A Global Dataset of Paired Subsurface and Surface Ocean Observations

The world's oceans cover more than seventy percent of the planet's surface, yet they remain one of the least understood parts of our own world. While satellites can watch the skin of the ocean with incredible detail, measuring the temperature and salinity of the water right at the surface, they cannot see what lies beneath. To truly understand how the ocean regulates Earth's climate, absorbs heat, and drives weather patterns, scientists need to see the entire three-dimensional structure of the water, from the surface down to the deep sea. For decades, researchers have relied on a patchwork of floating instruments that drift through the water, taking measurements at specific points, but these data points are scattered and sparse compared to the continuous view provided by satellites. This gap has made it difficult to build a complete picture of the ocean's hidden layers or to train the new generation of artificial intelligence tools that could help predict future changes.

A team of researchers has now bridged this gap by creating a massive, unified dataset that pairs what satellites see on the surface with what instruments measure deep below. Called OceanDepths, this new resource brings together nearly twenty-five years of observations, from the year 2000 to 2024, into a single, organized format that computers can easily read. The dataset links high-resolution satellite maps of sea surface temperature, salinity, and height with millions of vertical profiles of temperature and salinity taken by autonomous floats. By aligning these two very different types of data onto the same grid, the researchers have created a training ground for artificial intelligence to learn how the surface conditions of the ocean relate to its hidden depths. This work does not just fill a data hole; it provides a standardized, global testbed that allows scientists to move beyond rough approximations and begin modeling the complex, three-dimensional behavior of the ocean with unprecedented precision.

The challenge of understanding the ocean has long been defined by a mismatch in observation. Satellites provide a continuous, global view of the surface, capturing daily changes in temperature and height across the entire globe. In contrast, the measurements taken beneath the waves come from a fleet of about 4,000 autonomous floats, known as the Argo program, which drift with the currents and periodically dive to measure the water column. While these floats provide the most systematic sampling of the upper ocean, their coverage is extremely sparse relative to the vastness of the sea. In many places, there is only one measurement for every few hundred kilometers, leaving huge gaps in our knowledge. Historically, scientists have tried to fill these gaps using computer models or by reconstructing the data from limited observations, but these approaches often rely on assumptions that may not hold true everywhere. The result has been a lack of a single, reliable source of paired data that could be used to teach artificial intelligence how to infer the hidden structure of the ocean from what is visible on the surface.

To solve this, the researchers behind OceanDepths assembled a dataset that pairs satellite observations with in-situ measurements on a global scale. They collected satellite data for sea surface temperature, sea surface salinity, and sea surface height, which indicates the dynamic shape of the ocean surface. These surface measurements were matched with millions of temperature and salinity profiles taken by instruments in the water. The team carefully aligned these disparate data sources onto a common grid, dividing the entire globe into small squares and ensuring that every surface measurement had a corresponding depth profile nearby. The dataset covers the period from 2000 to 2024, offering a weekly view of the ocean's state. It includes over 9.5 million paired profiles, interpolated to fifty standard depth levels, reaching down to nearly 6,000 meters in some areas. This level of detail allows researchers to study the ocean at a resolution fine enough to see the swirling eddies and currents that dominate ocean dynamics, rather than just the broad, slow-moving trends.

A crucial feature of this new dataset is that it provides the raw, paired observations rather than a pre-reconstructed model. Previous datasets often consisted of smoothed, gap-filled products generated by computer models, which can hide errors or introduce biases. OceanDepths, by contrast, offers the actual measurements alongside a dense reanalysis product, which serves as a reference but is not the primary target. This distinction is vital for training artificial intelligence, as it allows the models to learn directly from the real, messy data of the ocean rather than from a perfect, simulated version of it. The dataset also includes a configurable system that breaks the globe into manageable chunks, or patches, making it easy for researchers to load and process the data for different types of analysis. This approach addresses a major limitation in previous work, where data was often scattered across different formats and resolutions, making it difficult to compare results or reproduce studies.

The researchers tested the utility of this new resource by using it to train simple artificial intelligence models to reconstruct the subsurface state of the ocean. The goal was to see if a model could look at the surface conditions and the sparse measurements from the floats and accurately predict the temperature and salinity at every depth in between. The results showed that while simple statistical methods based on historical averages remained strong competitors, the artificial intelligence models were able to capture more of the complex spatial patterns, particularly for salinity. The study found that models which considered the surrounding horizontal context—looking at the water around a specific point rather than just the vertical column above it—performed significantly better. This suggests that the ocean's hidden structure is deeply connected to the patterns visible on the surface and the broader regional context, a relationship that the new dataset makes possible to explore in detail.

The creation of OceanDepths represents a significant step forward for both oceanography and the field of artificial intelligence. By providing a standardized, global, and high-resolution dataset of paired observations, the researchers have removed a major barrier to entry for scientists looking to apply machine learning to ocean problems. The dataset's extreme sparsity, with observations covering less than one percent of the ocean at any given depth, presents a unique challenge that pushes the boundaries of current AI methods. It forces the development of new techniques capable of learning from incomplete and irregular data, much like how a human might infer the shape of a hidden object by feeling only a few points on its surface. The authors emphasize that this dataset is not a final solution but a foundational resource, a shared baseline that will allow the community to compare different methods and track progress over time.

The implications of this work extend beyond just better maps of the ocean. A deeper understanding of the ocean's three-dimensional structure is essential for predicting extreme weather events, tracking the movement of heat and carbon, and improving climate projections. The ability to reconstruct the hidden state of the ocean from surface observations could lead to more accurate forecasts of marine heatwaves and changes in ocean circulation. Furthermore, the dataset opens the door for observation-based forecasting methods that do not rely solely on complex physical models, potentially offering a new way to predict the future state of the ocean. The researchers have made all the data, code, and tools publicly available, inviting the global scientific community to build upon this foundation and explore the deep, hidden layers of our planet's most vital system.