Researchers have developed a physics-guided noise-enhanced adaptive fusion dual-stream network (PGNAF-DSNet) that integrates observations from the Cyclone Global Navigation Satellite System (CYGNSS) and Sentinel-1 synthetic aperture radar (SAR) to estimate river water levels over the Jiangsu plain in China. The model achieved a root mean square error of 0.2388 meters and a correlation coefficient of 0.9889 on a strictly chronological test set, reducing prediction errors by approximately 50% in regions significantly affected by human activities, according to a study published in Satellite Navigation on October 12, 2026.
Accurate river water level estimation is crucial for flood prevention, disaster mitigation and water resource management, yet strong non-stationarity and nonlinearity in complex plain river-network regions pose significant challenges to conventional approaches. Water level data traditionally come from hydrological stations, but sparse spatial coverage, high operational costs and limited real-time capability hinder continuous monitoring in these regions, where water level variations are jointly influenced by precipitation, sluice regulation, human activities and agricultural irrigation. Satellite remote sensing offers an alternative, but conventional altimeters have sparse cross-track spacing, SAR has limited temporal resolution, and CYGNSS suffers from coarse spatial resolution and observational noise.
The study, conducted by researchers from Nanjing University of Posts and Telecommunications, Henan Polytechnic University, Anhui University and the Nanjing Institute of Geography and Limnology of the Chinese Academy of Sciences, introduces several key innovations. First, researchers constructed physics-guided fused features: PFSWI (Polarization Fusion Surface Water Index), which multiplies Sentinel-1 backscatter differences by CYGNSS reflectivity, and Width_cv (river width coefficient of variation), embedding physical priors into the model. Second, they introduced a selective Gaussian noise enhancement strategy based on wavelet-denoised residual analysis, optimized using the Tree-structured Parzen Estimator algorithm (σ = 0.04157), applied only to channels susceptible to observational errors such as VV, VH, river width and precipitation. Third, the dual-stream architecture employs parallel Transformer and long short-term memory encoders to capture global and local temporal patterns, respectively, before fusion through an Adaptive Feature Fusion module that dynamically weights the two branches at the channel level.
On a strictly chronological dataset (training: January 2021–November 2022; testing: May–October 2023), PGNAF-DSNet outperformed Convolutional Neural Network, Transformer, LSTM and Transformer-LSTM baselines, achieving a mean absolute error of 0.1624 meters and maintaining a correlation coefficient of at least 0.95 across all forecast horizons up to T+6. Ablation experiments confirmed each component's contribution, and SHAP analysis indicated that Width_cv and PFSWI both exhibit high feature importance. The authors said the study's core contribution lies in moving beyond simple data-level fusion, capturing both global trends and local dynamics of water level changes in complex river networks. They added that the selective noise enhancement strategy acts as a regularizer, improving robustness against observational uncertainty and human-induced disturbances.
PGNAF-DSNet offers a practical framework for water level monitoring in plain river-network regions, with potential applications in flood early warning, water resource management and river-network regulation. The model reduced prediction errors by approximately 50% in areas significantly affected by human activities such as sluice regulation and agricultural irrigation, and maintained stable performance at T+6 forecasts, demonstrating promise for operational forecasting. The framework could also be extended to incorporate additional satellite observations, such as Surface Water and Ocean Topography water surface elevations, and adapted to other river systems through transfer learning. The full study is available via DOI 10.1186/s43020-026-00219-7.
