CLOINet: An Integral Neural Network for 3D Reconstruction of the Oceanic State


CLOINet marks a milestone in the understanding of ocean behaviour by integrating data from various observing platforms. It has also shown a significant improvement in performance, reducing the margin of error by up to 40%, which is crucial for the sustainable management of marine resources, climate change mitigation and the conservation of marine ecosystems.


A team of IMEDEA researchers, in collaboration with SOCIB and IMT Atlantique, has developed CLOINet, an innovative tool specifically designed to address the challenge of integrating sparse observations made in the ocean (in situ) into a complete three-dimensional model of the ocean state.

What makes CLOINet unique is its ability to harness information from remotely sensed ocean imagery, such as satellite imagery, to complement and enrich the information provided by in situ observations.

A multidisciplinary team of experts in oceanography and machine learning led this project, combining their expertise in both areas to develop an innovative solution. The combination of these two disciplines has significantly advanced the understanding and modelling of the state of the ocean, which is crucial for today's environmental and climate challenges.

This study arose due to the large amount of satellite imagery and in situ data available, and the need to develop advanced artificial intelligence techniques to combine this data effectively.


Unlocking Ocean Mysteries: CLOINet's Breakthrough in Precision Reconstruction


During the training process the CLOINet neural network was conducted using an end-to-end approach within a self-supervised framework. This means that the neural network was trained to perform a specific task (in this case, the three-dimensional reconstruction of ocean salinity) while automatically labelling and grouping the provided data into clusters based on their similarities.

To generate these datasets, the researchers used Observing System Simulation Experiments (OSSEs). These experiments involve simulating observations that mimic those that would be collected under real-world conditions. In this case, they relied on the NATL60 simulation, which is a numerical simulation of the North Atlantic Ocean derived from the Nucleus for European Modelling of the Ocean (NEMO) model. These simulations provided synthetic data that were used to train and test the performance of CLOINet.

The main objective of the study was to compare the three-dimensional reconstruction capability of ocean salinity between CLOINet and an always neural network-based version of the classical Optimal Interpolation, called OINet. Optimal Interpolation is a statistical technique commonly used to estimate unknown values between known observations. OINet represents a version of this technique adapted to automatically estimate its parameters and take advantage of the computational power of graphics cards (GPUs).

In addition to comparing the performance between CLOINet and OINet, the study also assessed the capability of both methodologies in practical applications involving real observational data. This allowed researchers to understand how these tools performed in real-world situations, which is fundamental to assessing their usefulness and applicability in practical ocean monitoring and forecasting applications.

The research covered a variety of scenarios, including both randomly and regularly spaced in situ salinity observations, paired with different remote sensing inputs such as Sea Surface Temperature (SST), Sea Surface Height (SSH), or a combination of both. CLOINet was found to be able to resolve scales 1.5 times smaller compared to OINet in dense regular sampling, and showed improved performance in terms of RMSE and correlation in random sampling contexts, especially when the available observations were limited.

Results showed that CLOINet was able to reconstruct smaller details in the salinity distribution compared to OINet, especially in cases where few in-situ observations were available.

Despite not being specifically trained to handle errors or noise in the data, CLOINet showed promising results with real data. It was able to effectively combine noisy chlorophyll and in-situ AUV (Autonomous Underwater Vehicle) temperature information, and successfully reconstructed previously unseen ocean temperature patterns, suggesting its potential in real-world operational applications.

CLOINet's modular design not only improves understanding of its internal processes, but also positions it for future improvements. The team plans to explore the extension of the model to incorporate spatio-temporal dynamics, as well as its application in multi-platform adaptive sampling ocean campaigns. In addition, they hope to take advantage of the upcoming high-resolution SSH observations from the SWOT mission to further refine and apply CLOINet.





CLOINet: ocean state reconstructions through remote-sensing, in-situ sparse observations and deep learning

Eugenio Cutolo, Ananda Pascual,  Simon Ruiz, Nikolaos D. Zarokanellos,  Ronan Fablet










CLOINet: ocean state reconstructions through remote-sensing, in-situ sparse observations and deep learning

Eugenio Cutolo, Ananda Pascual,  Simon Ruiz, Nikolaos D. Zarokanellos,  Ronan Fablet