BWI’s Virtual Station approach integrates remote sensing, hydrological modeling, and machine learning to provide real-time, scalable hydrological forecasts for global water management.
Accurate hydrological forecasting plays a vital role in managing water resources, preparing for natural disasters, and ensuring environmental sustainability. Traditional monitoring methods rely on physical hydrometric stations, which have limitations in terms of spatial coverage, accessibility, and most of all are incapable of providing any forecast. To address these challenges, Blue Water Intelligence (BWI) has developed an innovative Virtual Station approach, yet complementary to physical ground stations. By integrating Earth observation data, hydrological modeling, and machine learning, this system provides real-time, scalable hydrological forecasts. This blog post explains BWI’s methodology and its applications in river basin management in general, and in river flow forecasting in particular.
BWI’s Virtual Stations use remote sensing and in-situ data to monitor and predict surface water conditions. These stations provide valuable insights, both quantitative—such as runoff, discharge, flood levels, drought conditions, and water volume—and qualitative, including water temperature, vegetation health, and pollution levels. By comparing virtual station outputs with data from physical stations or satellite altimetry, whenever possible, BWI does its best to ensure a high level of accuracy and reliability. Below, we focus on describing our February 2025 river discharge forecasting methodology.
BWI’s river flow forecasting system draws on multiple data sources, combining ground sensors, satellite data, meteorological forecasts, and geospatial datasets. Ground sensors, like those in hydrometry networks such as Hydroportail, provide real-time observations. Satellite-based sensors contribute valuable information on topography, soil moisture, and vegetation cover. Meteorological forecasts come from ICON-EU, which offers a five-day lead time with a seven-kilometer resolution and hourly updates, and ECMWF, which extends the forecast up to ten days with updates every three to six hours at a 0.4-degree resolution. Additionally, the system integrates geospatial datasets, including Digital Elevation Models (DEMs) at a 25-meter resolution from ASTER and SRTM, river network data from EU-Hydro v2.1, which includes 160,000 reaches in France, and databases of dams, lakes, and reservoirs. For global river network data, BWI uses HydroSheds, which provides high-resolution hydrographic datasets. For example, HydroSheds includes over 2.9 million river reaches in India, supporting detailed hydrological modeling across diverse river basins.
BWI employs a semi-distributed hydrological model that is designed to be applied globally. The first step in the process is defining subcatchments for each river reach using the D8 algorithm on a 25-meter Digital Elevation Model (EU-DEM) that includes embedded river networks. To drive the model, meteorological inputs come from ICON-EU for short-term forecasts (up to five days) and ECMWF for longer lead times (up to ten days).
For rainfall-runoff simulation, BWI uses a drainage area ratio model to estimate discharge distribution. A three-linear storage model simulates the movement of surface flow, interflow, and subsurface flow. Overland flow propagation is refined using a DEM-based approach to ensure accuracy in runoff predictions. The hydrological model also incorporates an improved Muskingum flow propagation method to simulate how water moves through the river network, while real-time discharge observations further refine forecasts.
BWI enhances its hydrological predictions with machine learning. A Recurrent Neural Network (RNN) dynamically estimates model parameters to improve forecast accuracy. The system updates predictions in near real-time, integrating observed discharge data with meteorological inputs to provide the most up-to-date forecasts. Forecast lead times extend up to five days using ICON-EU meteorological forcing and up to ten days using ECMWF meteorological forcing. Model performance is evaluated using the normalized Nash-Sutcliffe Efficiency (nNSE) metric, which measures how well discharge predictions match real-world observations. Virtual station outputs are continuously benchmarked against physical station data to ensure reliability.
BWI’s forecasting system is already in use in multiple river basins, supporting flood early warning systems, hydrological planning, and water resource management. The methodology has been successfully implemented in France with SYMSAGEL for the Lys basin, in Senegal with SAED, and in Nepal with the Nepal Electricity Authority. Plans for expansion include further deployments in India, Bangladesh, and additional African regions.
BWI continues to advance its hydrological forecasting capabilities with ongoing improvements and technological integrations. AROME-HD meteorological forcing has already been incorporated, providing a 51-hour lead time with a 1.3-kilometer resolution and updates four times daily. Future efforts will focus on further enhancing model precision through improved snowmelt runoff modeling, dam operation simulations, and baseflow estimation techniques. Additionally, the integration of high-resolution soil moisture and land-use data will refine forecast accuracy. BWI is also exploring new collaborations to expand the geographical reach of its services, particularly in regions vulnerable to extreme hydrological events, where advanced forecasting can significantly mitigate risk and support sustainable water management.
BWI’s Virtual Station approach is transforming hydrological forecasting by combining remote sensing, semi-distributed modeling, and machine learning. This scalable system enhances water resource management and disaster preparedness across diverse hydrological environments. As new improvements continue to refine forecast accuracy, BWI is well-positioned to expand its operational capabilities worldwide.
Keywords: Hydrological Forecasting, Virtual Stations, Machine Learning, Rainfall-Runoff Modeling, Flow Propagation, Remote Sensing