A deep neural network framework for estimating coastal salinity from SMAP brightness temperature data
IntroductionSea surface salinity (SSS) is a critical parameter for understanding ocean circulation, marine ecosystem processes, and climate change. Despite advancements in satellite-based radiometry such as NASA’s Soil Moisture Active Passive (SMAP), significant challenges persist in coastal SSS ret...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Frontiers Media S.A.
2025-06-01
|
| Series: | Frontiers in Marine Science |
| Subjects: | |
| Online Access: | https://www.frontiersin.org/articles/10.3389/fmars.2025.1596325/full |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | IntroductionSea surface salinity (SSS) is a critical parameter for understanding ocean circulation, marine ecosystem processes, and climate change. Despite advancements in satellite-based radiometry such as NASA’s Soil Moisture Active Passive (SMAP), significant challenges persist in coastal SSS retrieval due to radio frequency interference (RFI), land-sea contamination, and complex interactions of nearshore dynamic processes.MethodThis study proposes a deep neural network (DNN) framework that integrates SMAP L-band brightness temperature data with ancillary oceanographic and geographic parameters such as sea surface temperature, the shortest distance to the coastline (dis) to enhance SSS estimation accuracy in the Yellow and East China Seas. The framework leverages machine learning interpretability tools (Shapley Additive Explanations, SHAP) to optimize input feature selection and employs a grid search strategy for hyperparameter tuning.Results and discussionSystematic validation against independent in-situ measurements demonstrates that the baseline DNN model constructed for the entire region and time period outperforms conventional algorithms including K-Nearest Neighbors, Random Forest, and XGBoost and the standard SMAP SSS product, achieving a reduction of 36.0%, 33.4%, 40.1%, and 23.2%, respectively in root mean square error (RMSE). Compared with SMAP SSS products, the baseline DNN demonstrates a reduction of 33.8% and 7.3% in RMSE in nearshore (dis ≤ 50 km) and offshore regions (50 km<dis ≤ 200 km), respectively. The specific models constructed for nearshore and offshore areas, as well as for the four seasons, further improves salinity retrieval accuracy, especially in nearshore regions, highlighting the effectiveness of regional and seasonal optimization strategies in complex coastal environments. The DNN framework significantly mitigates coastal salinity biases caused by RFI and land contamination, providing a robust tool for applications such as coastal hydrological monitoring and marine resource management. |
|---|---|
| ISSN: | 2296-7745 |