Researchers created a machine-learning framework to identify salinity anomalies in the Yangtze Estuary by linking them to meteorological conditions 45 to 90 days early.
Published on Friday, the study uses daily salinity data from 2021 to 2024 to track freshwater availability.
The team tested two machine-learning models: XGBoost and Tabular Prior-data Fitted Network (TabPFN). They used ten meteorological factors for the evaluation. XGBoost showed higher recall and F1-scores. The TabPFN model achieved higher precision but detected only about half of the anomaly days.
Researchers used SHAP to interpret the predictions. This method identified precipitation, near-surface temperature, and wind-related factors as the primary drivers of these anomalies. These findings provide a basis for the retrospective identification of salinity anomalies. This approach was rarely used in previous estuarine research.
Salinity anomalies occur most often between September and the following March. They happen least from May to July. The study says these anomalies associate with warmer conditions and reduced precipitation over the Yangtze River basin in the months before the event. At shorter lead times, the anomalies correlate with near-surface and lower-tropospheric northerly and easterly wind anomalies. They also link to reduced 850 hPa specific humidity.
This research helps explain the Yangtze River Estuary. It is a dynamic coastal system driven by large river discharge, season-reversing monsoon winds, and strong tidal stirring. A separate study appeared on November 13, 2026. It focused on reconstructing three-dimensional ocean temperature fields in the same coastal system.
Predicting seasonal precipitation in the Yangtze River basin also uses advanced neural network approaches. Research published on November 5, 2026, said traditional meteorological methods often struggle with the nonlinear, multi-scale nature of precipitation data. Common neural networks include Back Propagation (BP) networks.
The study uses a 46-year record from 1979 to 2025. This captures interannual variability while maintaining a reasonable predictor-to-sample-size ratio for machine-learning models. This record length exceeds the data span used in the salinity anomaly research.
Related reading
- Honda CBR Riders Unite in Batam for City Roll and Community Gathering
- GTA 6 Wipes Cash, Bank, Stashes & Vehicle at Chapter 2: How to Protect Money
- EU Researchers Design Wide-Field Radio Telescope to Monitor Dynamic Sky (world-today-news.com)
- Milky Way’s Bar Drives Unusual Star Formation, Researchers Find (newsy-today.com)