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

Introduction Satellite‑based multi‑variable sea surface observations combined with deep learning enable low‑cost reconstruction of three‑dimensional ocean temperature fields. However, existing data‑driven inversion models exhibit prominent systematic biases in estuary and inner shelf areas. Two major causes account for these errors: models pre‑trained on biased hydrodynamic outputs fail to represent real‑world coastal marine conditions, whereas direct training using sparse in‑situ measurements easily leads to severe overfitting. Methods To mitigate these issues over shelf seas, this study develops a two‑stage residual fine‑tuning deep learning framework integrating residual convolutional neural networks and Convolutional Block Attention Module (CBAM) for temperature inversion. Using multi‑source satellite surface parameters together with geographic and temporal features as inputs, the CBAM‑augmented baseline model is first pre‑trained on SCHISM hydrodynamic simulation outputs to learn the spatial and vertical temperature patterns of shelf waters. Subsequently, with all backbone parameters frozen, a lightweight residual correction branch is fine‑tuned using limited high‑quality in‑situ temperature profiles to correct model‑originated deviations. Results Ablation experiments demonstrate that incorporating CBAM alone yields only limited overall performance gains. Spatial leave‑one‑station‑out cross‑validation reveals heterogeneous correction performance across mooring sites. Sites with large baseline MAE and RMSE experience pronounced error reduction, while stations with relatively small baseline prediction errors still achieve moderate performance gains, and no obvious prediction degradation is observed at any site. Compared with a simple constant‑bias correction scheme, the proposed transfer fine‑tuning strategy achieves consistent error reduction across all stations and vertical layers. Discussion Owing to the lightweight correction design, the framework only demands sparse in‑situ observations and achieves robust inversion in data‑limited coastal shelf environments. This work presents an effective deep‑learning strategy for satellite‑driven three‑dimensional temperature reconstruction and provides a technical reference for inversion studies in estuarine and marginal sea coastal zones.

Authors

Institutions

Publication Details

Journal
Frontiers in Marine Science
Published
2026-11-13
DOI
https://doi.org/10.3389/fmars.2026.1976762
Primary Topic
Oceanographic and Atmospheric Processes
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

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

Fangzhou Shen, weijie Jiang, Xiangguang Zhang, Jiawei Jiang et al.
4 links
Frontiers in Marine Science
Oceanographic and Atmospheric Processes
article

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

Fangzhou Shen, weijie Jiang, Xiangguang Zhang, Jiawei Jiang, Nan Zang, Jian Xu, Hua Deng
article en

Abstract

Introduction Satellite‑based multi‑variable sea surface observations combined with deep learning enable low‑cost reconstruction of three‑dimensional ocean temperature fields. However, existing data‑driven inversion models exhibit prominent systematic biases in estuary and inner shelf areas. Two major causes account for these errors: models pre‑trained on biased hydrodynamic outputs fail to represent real‑world coastal marine conditions, whereas direct training using sparse in‑situ measurements easily leads to severe overfitting. Methods To mitigate these issues over shelf seas, this study develops a two‑stage residual fine‑tuning deep learning framework integrating residual convolutional neural networks and Convolutional Block Attention Module (CBAM) for temperature inversion. Using multi‑source satellite surface parameters together with geographic and temporal features as inputs, the CBAM‑augmented baseline model is first pre‑trained on SCHISM hydrodynamic simulation outputs to learn the spatial and vertical temperature patterns of shelf waters. Subsequently, with all backbone parameters frozen, a lightweight residual correction branch is fine‑tuned using limited high‑quality in‑situ temperature profiles to correct model‑originated deviations. Results Ablation experiments demonstrate that incorporating CBAM alone yields only limited overall performance gains. Spatial leave‑one‑station‑out cross‑validation reveals heterogeneous correction performance across mooring sites. Sites with large baseline MAE and RMSE experience pronounced error reduction, while stations with relatively small baseline prediction errors still achieve moderate performance gains, and no obvious prediction degradation is observed at any site. Compared with a simple constant‑bias correction scheme, the proposed transfer fine‑tuning strategy achieves consistent error reduction across all stations and vertical layers. Discussion Owing to the lightweight correction design, the framework only demands sparse in‑situ observations and achieves robust inversion in data‑limited coastal shelf environments. This work presents an effective deep‑learning strategy for satellite‑driven three‑dimensional temperature reconstruction and provides a technical reference for inversion studies in estuarine and marginal sea coastal zones.

Frontiers in Marine ScienceVol. 13
Shandong University (CN), Chinese Academy of Sciences (CN), Institute of Oceanology (CN), Qingdao National Laboratory for Marine Science and Technology (CN), Powerchina Huadong Engineering Corporation (China) (CN), Institute of Oceanology (BG), University of Chinese Academy of Sciences (CN), Laoshan Laboratory (CN)
Climate action, Life below water
Oceanographic and Atmospheric Processes
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon