AI-Based Subsurface Thermohaline Structure Retrieval from Remote Sensing Observations

Hua Su, Wenfang Lu, An Wang, Tianyi ZhangOpenalex Unknown 2023 · DOI · OpenAlex

Abstract

Observing the ocean’s interior is becoming extremely important since recent evidence suggests widespread warming in the subsurface and deeper ocean as a response to the Earth’s Energy Imbalance (EEI) in recent decades. However, the ocean’s interior observations are sparse and insufficient, severely constraining the studies of ocean interior dynamics and variabilities. Detecting and predicting subsurface and deeper ocean thermohaline structure from satellite remote sensing measurements is quite essential for understanding ocean interior 3D environment and processes effectively. This study proposes several novel approaches based on artificial intelligence to accurately retrieve and predict subsurface thermohaline structure (upper 2000 m) from multiple satellite observations combined with Argo float data. We manage to construct AI-based deep ocean remote sensing techniques with high spatiotemporal applicability to subtly detect and describe global subsurface thermohaline structure, supporting the studies of ocean internal processes and anomalies under global warming.

Introduction

The ocean absorbs over 90% of the Earth’s excess heat, yet the vast majority of this warming occurs beneath the surface, hidden from direct satellite view. While remote sensing provides high-resolution data on the ocean's "skin"—including surface temperature, sea level height, and salinity—our understanding of the interior remains constrained by the sparse distribution of physical sensors. This research demonstrates how machine learning and deep learning can act as "X-ray vision" for the ocean, reconstructing 3D subsurface temperature and salinity structures using only 2D satellite surface observations.

The Challenge of Subsurface Observation

Currently, the scientific community relies heavily on the Argo float network—a fleet of thousands of robotic probes. Despite their utility, the ocean is too vast for these floats to provide continuous global coverage, leaving significant spatiotemporal gaps. Traditional numerical models, while useful, are computationally expensive and often struggle to resolve the complex, non-linear mesoscale dynamics inherent in the ocean's interior. This paper addresses the "retrieval" problem: leveraging high-resolution satellite surface data to accurately predict subsurface variables like temperature and salinity anomalies.

Methodology and Data Integration

The researchers evaluated two primary AI architectures to bridge the surface-to-interior gap:

  1. Ensemble Learning (XGBoost and Random Forests): These models were employed to derive subsurface temperature anomalies (STA) and salinity anomalies (SSA). They are particularly effective at identifying non-linear relationships between surface variables at specific geographic points.
  2. Deep Learning (Bi-LSTM and CNN):
    • Bi-Long Short-Term Memory (Bi-LSTM): Used to capture temporal "memory" by processing time-series data both forward and backward.
  3. Convolutional Neural Networks (CNN): Used to extract spatial features and patterns from satellite imagery to reconstruct the 3D structure of the water column.
Global distribution of subsurface temperature anomalies (STA) at depth, illustrating the complex spatial patterns the AI models must learn to reconstruct.

The models utilized four primary satellite inputs: Sea Surface Height (SSH) from AVISO, Sea Surface Temperature (SST) from OISST, Sea Surface Salinity (SSS) from SMOS, and Sea Surface Wind (SSW) from CCMP. Argo gridded data served as the ground truth for training and validation, focusing on the upper 2,000 meters.

Key Insights and Results

The core premise of this work is that subsurface dynamic processes—such as internal waves or heat sequestration—leave subtle signatures on the surface. These signatures, such as minute bulges in sea level or localized salinity shifts, are detectable by AI even when traditional physical equations struggle to resolve them. By focusing on anomalies (deviations from long-term averages) rather than raw values, the models were forced to learn the stochastic signals related to actual climate variability.

Comparison between observed (a) and AI-predicted (b) global salinity anomalies. The high degree of correlation demonstrates the model's ability to accurately retrieve subsurface structures from surface data.

The AI models showed remarkable consistency with in-situ Argo observations:

  • XGBoost Performance: Achieved an $R^2$ of 0.989 for temperature anomalies and 0.981 for salinity at 600m depth, with a Root Mean Square Error (RMSE) as low as 0.026 °C.
  • CNN Performance: For direct structure retrieval at 200m, the CNN reached an $R^2$ of 0.972 for temperature.
  • Vertical Profiling: The Bi-LSTM model reconstructed meridional vertical profiles where over 99% of predicted points fell within a ±1 °C error margin compared to actual float data.

Conclusion and Future Directions

AI-based approaches demonstrate significant potential for subsurface environment data reconstruction. While currently "tethered" to the quality of Argo training data, these models offer a promising path for investigating ocean interior changes and their role in global climate change. Future work will focus on extending these models over longer time-series to capture decadal shifts and improving architectures to better resolve small-scale mesoscale features like eddies and fronts.

Summary updated

Abstract

Observing the ocean’s interior is becoming extremely important since recent evidence suggests widespread warming in the subsurface and deeper ocean as a response to the Earth’s Energy Imbalance (EEI) in recent decades. However, the ocean’s interior observations are sparse and insufficient, severely constraining the studies of ocean interior dynamics and variabilities. Detecting and predicting subsurface and deeper ocean thermohaline structure from satellite remote sensing measurements is quite essential for understanding ocean interior 3D environment and processes effectively. This chapter proposes several novel approaches based on artificial intelligence to accurately retrieve and predict subsurface thermohaline structure (upper 2000 m) from multiple satellite observations combined with Argo float data. We manage to construct AI-based deep ocean remote sensing technique with high spatiotemporal applicability to subtly detect and describe global subsurface thermohaline structure, so as to support the studies of ocean internal processes and anomalies under global warming. The AI-based approaches demonstrate great potential for subsurface environment data reconstruction and should be a promising technique for investigating ocean interior change and variability as well as its role in global climate change from satellite remote sensing measurements.
BibTeX
@inproceedings{su2023ai,
  author = {Hua Su and Wenfang Lu and An Wang and Tianyi Zhang},
  title = {{AI-Based Subsurface Thermohaline Structure Retrieval from Remote Sensing Observations}},
  booktitle = {Openalex Unknown},
  year = {2023},
  doi = {10.1007/978-981-19-6375-9\_5},
}
Figure