A physics-guided noise-enhanced adaptive fusion dual-stream network for water level estimation using CYGNSS and Sentinel-1 observations

Accurate river water level estimation is crucial for flood prevention, disaster mitigation, water resource management, and river-network regulation. However, water level variations in complex river-network regions are characterized by strong non-stationarity and nonlinearity, posing significant challenges to accurate estimation. To address this issue, this study proposes a Physics-Guided Noise-Enhanced Adaptive Fusion Dual-Stream Network (PGNAF-DSNet) for water level estimation using the Cyclone Global Navigation Satellite System (CYGNSS) and Synthetic Aperture Radar (SAR) observations from Sentinel‑1 over the complex river network of the Jiangsu plain. The proposed model incorporates physics-guided fusion feature, selective Gaussian noise enhancement, dual-stream temporal encoding, and an adaptive feature fusion mechanism. A strict chronological splitting strategy is adopted. The experimental results show that PGNAF-DSNet outperforms machine learning models, including CNN, Transformer, LSTM, and Transformer-LSTM, achieving Root Mean Square Error ( S RMSE ) of 0.238 8 m, Mean Absolute Error ( S MAE ) of 0.162 4 m, and a Correlation Coefficient ( R ) of 0.988 9 on the test set. Notably, in regions significantly affected by human activities, the prediction error was reduced by approximately 50%, demonstrating strong generalization and anti-interference capabilities. Furthermore, the model maintained high stability in long-term T+6 prediction, with the R consistently exceeding 0.95. These results demonstrate the effectiveness of PGNAF-DSNet for water level prediction in complex river-network regions and provide a reliable technical foundation for dynamic water level monitoring and refined water management in plain river networks.