Two-Stage Transfer Learning Improves Three-Dimensional Coastal Ocean Temperature Inversion

Researchers have developed a two-stage residual fine-tuning deep learning framework designed to reconstruct three-dimensional ocean temperature fields over the Yangtze River Estuary and its adjacent shelf waters. Combining multi-variable sea surface observations obtained from satellites with deep learning architectures offers an economical approach for mapping subsurface marine temperatures. Nevertheless, conventional data-driven inversion approaches regularly encounter pronounced systematic biases in dynamic nearshore zones such as estuaries and inner continental shelves.

The study highlights two principal causes behind the inaccuracies found in existing inversion models. First, neural networks pre-trained exclusively on hydrodynamic model simulations inherit systematic discrepancies from biased numerical outputs and fail to accurately mirror real-world coastal marine conditions. Second, training models directly on actual field observations often triggers severe overfitting because in-situ ocean measurements in coastal and shelf environments remain relatively sparse.

To overcome these persistent limitations in shelf seas, the research team designed an integrated framework combining residual convolutional neural networks with the Convolutional Block Attention Module, known as CBAM. The framework utilizes multi-source satellite surface parameters alongside relevant geographic and temporal features as primary inputs.

Ablation experiments conducted within the study demonstrated that incorporating the CBAM attention mechanism alone provided only limited overall performance gains. However, this architectural enhancement provided superior feature initialization for the subsequent stage of transfer learning. Once the baseline network captured the broader physical structures from the numerical simulations, all backbone parameters were completely frozen to preserve the underlying representations.

In the second phase of the architecture, the researchers introduced a lightweight residual correction branch. This secondary branch was fine-tuned using a limited collection of high-quality in-situ vertical temperature profiles, allowing the model to specifically correct deviations originating from the baseline numerical simulations without risking overfitting across the entire network.

Evaluation through spatial leave-one-station-out cross-validation revealed heterogeneous correction capabilities across different mooring monitoring sites. Mooring stations that exhibited large baseline mean absolute errors and root-mean-square errors experienced pronounced error reductions following the fine-tuning process. Concurrently, monitoring stations that already had relatively small baseline prediction errors still achieved moderate performance gains, with no observable prediction degradation identified at any evaluated site.

When compared against a standard constant-bias correction technique, the proposed transfer fine-tuning strategy demonstrated consistent error reductions across all vertical layers and observation stations. The design of the lightweight correction branch ensures that the system requires only sparse in-situ measurements to perform robust temperature inversions in data-constrained coastal shelf settings.

The findings indicate that this two-stage transfer learning strategy provides an effective technical approach for satellite-driven three-dimensional ocean temperature reconstruction. By bridging numerical hydrodynamic simulations and sparse observational records, the method serves as a practical reference for ongoing inversion and observation studies across estuarine environments and marginal sea coastal zones.

Subject of Research: Three-dimensional ocean temperature inversion using transfer learning

Article Title: Two-stage transfer learning for temperature inversion over the Yangtze River Estuary and adjacent shelf

Article References: Deng, H., Zang, N., Jiang, J., Xu, J., Shen, F., Jiang, W., & Zhang, X. (2026). Two-stage transfer learning for temperature inversion over the Yangtze River Estuary and adjacent shelf. Frontiers in Marine Science, 13, Article 1976762. https://doi.org/10.3389/fmars.2026.1976762

Image Credits: AI Generated

DOI: 10.3389/fmars.2026.1976762

Keywords: deep learning, transfer learning, ocean temperature inversion, three-dimensional temperature reconstruction, Yangtze River Estuary, coastal shelf waters, SCHISM model, residual convolutional neural networks, Convolutional Block Attention Module, residual fine-tuning, in-situ temperature profiles, satellite oceanography, hydrodynamic modeling, spatial cross-validation, marine observation, estuarine dynamics, sea surface observations