Accurate river water level estimation is critical for flood prevention, disaster mitigation and water resource management, but complex plain river-network regions present severe challenges because water levels are influenced by precipitation, sluice regulation, human activities and agricultural irrigation, producing strong nonlinearity, non-stationarity and spatial heterogeneity. Traditional hydrological stations offer sparse spatial coverage, high operational costs and limited real-time capability, while satellite remote sensing alternatives each have drawbacks: conventional altimeters have sparse cross-track spacing, synthetic aperture radar (SAR) has limited temporal resolution, and the Cyclone Global Navigation Satellite System (CYGNSS) suffers from coarse spatial resolution and observational noise.
To address these gaps, 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 developed a physics-guided noise-enhanced adaptive fusion dual-stream network (PGNAF-DSNet) that integrates Sentinel-1 SAR and CYGNSS observations over the Jiangsu plain in China. The study was published in Satellite Navigation on October 12, 2026.
The model introduces three key innovations. First, it constructs physics-guided fused features: the Polarization Fusion Surface Water Index (PFSWI), which multiplies Sentinel-1 backscatter differences by CYGNSS reflectivity, and the river width coefficient of variation (Width_cv), embedding physical priors into the model. Second, a selective Gaussian noise enhancement strategy based on wavelet-denoised residual analysis, optimized using the Tree-structured Parzen Estimator algorithm (σ = 0.04157), is applied only to channels susceptible to observational errors such as VV, VH, river width and precipitation. Third, a 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 CNN, Transformer, LSTM and Transformer-LSTM baselines, achieving a root mean square error of 0.2388 meters, a mean absolute error of 0.1624 meters and a correlation coefficient (R) of 0.9889. Notably, the model maintained R ≥ 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. Previous approaches mainly used CYGNSS to fill temporal gaps in SAR observations without fully exploiting cross-modal feature interactions, they said. By constructing physics-guided fused features and adopting a dual-stream architecture, the model captures both global trends and local dynamics of water level changes in complex river networks. The selective noise enhancement strategy acts as a regularizer, improving robustness against observational uncertainty and human-induced disturbances.
PGNAF-DSNet 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.
