Learning of Sea Surface Height Interpolation from Multi-variate Simulated Satellite Observations

Overview

• Satellite-based remote sensing has revolutionized our understanding of the ocean
• Space-borne altimetry measures Sea Surface Height (SSH), which is used to estimate surface geostrophic currents
• SSH observations have gaps due to sensor technology, so complete maps are produced using linear Optimal Interpolation (OI)
• Sea Surface Temperature (SST) has higher data coverage and is linked to geostrophic currents through advection
• The authors propose a new Observing System Simulation Experiment (OSSE) to emulate 20 years of SSH and SST satellite observations
• They train an Attention-Based Encoder-Decoder (abed) deep learning network on this data, comparing settings with and without ground truth during training
• The authors evaluate the SSH interpolations in terms of eddy detection and introduce a new learning strategy to transfer knowledge from simulation to real satellite observations

Plain English Explanation

Satellites are revolutionizing our understanding of the oceans. They use altimetry (measuring the height of the sea surface) to estimate the speed of surface currents. However, the technology has some gaps in the data it collects, so scientists use a mathematical technique called Optimal Interpolation to fill in the missing information and create complete maps.

Another type of satellite data is Sea Surface Temperature (SST), which is linked to the speed of the currents. The authors of this paper have created a simulated dataset that mimics 20 years of real satellite observations of SSH and SST. They then trained a deep learning model, called an Attention-Based Encoder-Decoder (abed), to reconstruct the SSH data using this simulated dataset.

The authors compared two different ways of training the abed model - one where it had access to the "ground truth" (the real, complete data) during training, and one where it did not. They also looked at whether including the SST data helped the model perform better at reconstructing the SSH.

In addition, the authors introduced a new way to take what the model learned from the simulated data and apply it to the real satellite observations. They first did some supervised pre-training on the simulated data, and then did additional unsupervised fine-tuning on the real satellite data. This helped the model achieve a 24% reduction in error compared to the standard Optimal Interpolation approach when applied to the real SSH observations.

Technical Explanation

The authors created a multi-variate Observing System Simulation Experiment (OSSE) to emulate 20 years of satellite observations of Sea Surface Height (SSH) and Sea Surface Temperature (SST). They then trained an Attention-Based Encoder-Decoder (abed) deep learning network on this simulated data, comparing two settings: one where the model had access to ground truth during training, and one where it did not.

The abed model was trained to reconstruct the SSH data using either supervised or unsupervised loss functions, with or without the additional SST information. The authors evaluated the SSH interpolations in terms of eddy detection, which is an important oceanographic metric.

To bridge the gap between the simulated and real-world data, the authors introduced a new learning strategy. They first did supervised pre-training on the OSSE data, and then did unsupervised fine-tuning on the actual satellite observations. This combined approach, along with the use of SST data, led to a 24% reduction in root mean squared error compared to the standard Data Unification and Altimeter Combination System (duacs) when applied to real SSH observations from the Ocean Data Challenge 2021.

Critical Analysis

The paper presents a novel approach to SSH reconstruction using deep learning and multi-variate satellite data. The authors' use of a simulated OSSE dataset to train and validate their model is a reasonable approach, as it allows them to control the experimental conditions and have access to ground truth data that is not available in the real-world observations.

However, one potential limitation is the realism of the OSSE dataset. While the authors claim it emulates 20 years of satellite observations, the extent to which it accurately captures the full complexity and variability of the real ocean system is not clear. Additionally, the authors do not provide details on how the OSSE was generated or validated, which makes it difficult to assess the suitability of this dataset for training deep learning models.

Furthermore, the authors' new learning strategy of supervised pre-training followed by unsupervised fine-tuning is an interesting approach, but its effectiveness may depend heavily on the quality and representativeness of the simulated data. It would be valuable to see the authors evaluate this strategy on a wider range of real-world satellite datasets to better understand its broader applicability and robustness.

Despite these potential limitations, the authors' work demonstrates the potential of deep learning and multi-variate data fusion to improve the reconstruction of SSH from satellite observations. Their findings contribute to the ongoing efforts to better understand ocean dynamics and climate change through remote sensing.

Conclusion

This paper proposes a novel deep learning-based approach for reconstructing Sea Surface Height (SSH) from satellite observations, leveraging the complementary information in Sea Surface Temperature (SST) data. By creating a simulated Observing System Simulation Experiment (OSSE) and training an Attention-Based Encoder-Decoder (abed) model, the authors were able to achieve a 24% reduction in root mean squared error compared to standard Optimal Interpolation methods when applied to real-world satellite observations.

The authors' work demonstrates the potential of deep learning and multi-variate data fusion to improve our understanding of the ocean state and dynamics, which is crucial for monitoring and responding to climate change. The new learning strategy of supervised pre-training and unsupervised fine-tuning also presents an interesting approach for bridging the gap between simulated and real-world data, which could have broader applications in Earth system modeling and data assimilation.

Overall, this research represents an important step forward in the use of advanced machine learning techniques for satellite image time series analysis and uncertainty-aware estimation of geophysical parameters, with potential implications for a wide range of environmental monitoring and climate research applications.