Physics-Guided Neural Networks Improve River Level Forecasting via Satellite Fusion

A new machine learning framework, the Physics-Guided Noise-Enhanced Adaptive Fusion Dual-Stream Network (PGNAF-DSNet), has demonstrated significant improvements in monitoring water levels across complex river networks. Published on October 12, 2026, in the journal Satellite Navigation, the research addresses the inherent limitations of sparse hydrological station coverage and the observational noise typical of satellite remote sensing.

The research team, comprising investigators from the 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, designed the model to synthesize disparate data streams. By integrating Sentinel-1 Synthetic Aperture Radar (SAR) and Cyclone Global Navigation Satellite System (CYGNSS) observations, the system achieves a more granular understanding of water dynamics in the Jiangsu plain of China.

The architecture relies on physics-guided feature construction, which embeds domain-specific priors into the model inputs. Researchers developed the Polarization Fusion Surface Water Index (PFSWI) by combining Sentinel-1 backscatter metrics with CYGNSS reflectivity data. This approach is supplemented by the river width coefficient of variation (Width_cv), which provides the model with essential physical context regarding river morphology and flow characteristics.

To manage the high levels of observational uncertainty, the authors implemented a selective Gaussian noise enhancement strategy. This module, optimized via the Tree-structured Parzen Estimator (TPE) algorithm with a sigma value of 0.04157, applies targeted denoising to specific channels, including precipitation data and SAR polarization bands. This regularizer effectively mitigates the impact of human-induced disturbances, such as sluice regulation and agricultural irrigation, which often confound standard predictive models.

The core of the network utilizes a dual-stream architecture to process temporal data. A parallel structure consisting of Transformer and long short-term memory (LSTM) encoders allows the system to capture both global trends and local, short-term fluctuations in water levels. An Adaptive Feature Fusion (AFF) module dynamically weights these branches at the channel level to ensure that the most reliable temporal features dominate the final prediction, creating a robust representation of the hydrological state.

Performance benchmarks on a strictly chronological dataset spanning from 2021 to 2023 indicate that PGNAF-DSNet outperforms conventional baselines, including standalone CNN, Transformer, and LSTM models. The system achieved a root mean square error of 0.2388 meters and a correlation coefficient of 0.9889. Notably, the model maintained a correlation coefficient above 0.95 across all forecast horizons up to T+6, highlighting its potential for operational deployment in real-world scenarios.

The significance of this work lies in its departure from simple data-level fusion, which often fails to capture the complex, non-linear interactions between satellite-derived metrics. By moving to feature-level integration, the researchers have created a system that better accounts for the spatial heterogeneity of plain river-network regions. This methodology ensures that the model remains sensitive to the non-stationarity of water levels caused by both environmental and anthropogenic factors.

The SHAP (SHapley Additive exPlanations) analysis performed by the team confirms that the physics-guided features are critical to the model’s predictive accuracy. By quantifying the contribution of each input variable, the researchers demonstrated that Width_cv and PFSWI exert a high degree of influence on the final output. This transparency in feature importance provides a necessary level of interpretability for high-stakes environmental monitoring tasks.

This methodology offers a scalable framework for flood early warning systems and broader water resource management. Because the model demonstrated a 50% reduction in prediction errors in human-impacted regions, it provides a robust alternative to traditional ground-based sensors. The researchers suggest that the framework is adaptable to other river systems through transfer learning and could be further enhanced by incorporating additional satellite inputs like Surface Water and Ocean Topography (SWOT) data.

Future development will likely focus on expanding the model’s geographical scope and integrating real-time data ingestion pipelines. As the team continues to refine the dual-stream architecture, the potential for wider adoption in disaster mitigation and infrastructure management remains a primary objective. The successful application of physics-guided priors suggests a promising path for applying deep learning to high-stakes environmental monitoring tasks.