Obeng et al. (2026) From forecast to alert: an open-data AI flood early warning system with reservoir monitoring for the White Volta Basin

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2026
• Date: 2026-11-05
• Authors: Joseph Obeng, Hongyong Yuan, Tao Chen, Lida Huang
• DOI: 10.5194/nhess-26-5667-2026

Research Groups

School of Safety Science, Tsinghua University, Beijing, China

Short Summary

This study developed an open-data, AI-driven flood early warning system for the White Volta Basin, integrating reservoir monitoring to address unmonitored dam releases from Burkina Faso's Bagre Reservoir. The system provides highly accurate 1-5 day flood forecasts, significantly outperforming existing global models, and is designed for seamless integration into Ghana's current warning infrastructure.

Objective

• To develop and validate a complete end-to-end, open-data, AI-driven flood early warning system for the White Volta Basin, specifically integrating Bagre Reservoir monitoring to account for unmonitored dam releases, and ensuring compatibility with Ghana's existing institutional infrastructure.

Study Configuration

• Spatial Scale: White Volta Basin (catchment area 92,950 km² above Nawuni gauge), Nawuni gauging station (9.7° N, 1.08° W), Bagre Dam (Burkina Faso), and the lower White Volta floodplain (Central Gonja and Savannah districts, Ghana). CHIRPS rainfall data at 0.05° resolution.
• Temporal Scale: Forecast lead times of 1, 3, and 5 days. GRDC discharge data from January 1975 to February 2007 (core 1987-2006). CHIRPS rainfall and ERA5-Land reanalysis from 1985 to 2023. JRC Bagre reservoir storage proxy from March 1984 to January 2024 (monthly). Sentinel-1 SAR imagery for 2016, 2019, and 2020 flood seasons. Training period: January 1987 to December 1999; Validation period: January 2000 to December 2003; Test period: January 2004 to February 2007.

Methodology and Data

• Models used: Random Forest regressor, XGBoost (Extreme Gradient Boosting), Long Short-Term Memory (LSTM) network (two-layer), and an ensemble (simple unweighted average of the three models). Log-Pearson Type III (LP3) distribution for alert threshold derivation. GloFAS v4 (LISFLOOD model) for benchmark comparison.
• Data sources:
  • Streamflow: Daily mean discharge at Nawuni gauge from the Global Runoff Data Centre (GRDC station 1531450).
  • Rainfall: Daily gridded precipitation from the Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS) version 2.0.
  • Atmospheric reanalysis: Daily 2 m air temperature, potential evapotranspiration, and volumetric soil water content (0-7 cm) from ERA5-Land.
  • Reservoir storage: Monthly water surface extent of the Bagre Reservoir from the JRC Global Surface Water dataset version 1.4, linearly interpolated to daily.
  • Inundation mapping: Sentinel-1 C-band Synthetic Aperture Radar (SAR) imagery (Interferometric Wide Swath mode, VV polarisation).
  • Benchmark: GloFAS v4 reanalysis discharge from the ECMWF LISFLOOD reanalysis archive.

Main Results

• Ensemble Performance (Independent Test Period, January 2004 – February 2007):
  • Kling–Gupta Efficiency (KGE) scores: 0.984 (1 d lead), 0.974 (3 d lead), 0.957 (5 d lead).
  • Root Mean Square Error (RMSE) at 1 d lead: 33.0 m³/s.
  • For extreme flows (observed discharge > 872 m³/s), KGE was 0.808 (1 d lead) and 0.248 (5 d lead).
• GloFAS v4 Benchmark (Nawuni, same test period):
  • KGE: -0.526, Nash–Sutcliffe Efficiency (NSE): -0.567, RMSE: 335.3 m³/s. This indicates performance worse than a climatological mean predictor, primarily due to the absence of Bagre Dam release operations in LISFLOOD.
• Alert Thresholds (Calibrated to 30 wet-season annual maxima, 1975–2006):
  • Green: < 1114 m³/s (< 1.5-year return period)
  • Yellow: 1114–1362 m³/s (1.5–2-year return period)
  • Orange: 1362–1899 m³/s (2–5-year return period)
  • Red: > 1899 m³/s (> 5-year return period)
• Alert Skill (21-fold Leave-One-Year-Out cross-validation, 1987–2007):
  • Red-tier (1 d lead): Probability of Detection (POD) = 0.902 (95% CI: 0.844–0.951), False Alarm Ratio (FAR) = 0.134 (95% CI: 0.080–0.190).
  • Red-tier (5 d lead): POD = 0.770 (95% CI: 0.697–0.844), FAR = 0.266 (95% CI: 0.195–0.331).
  • Mean absolute alert onset timing error: 1.0 day (1 d lead) and 2.0 days (5 d lead).
• Sentinel-1 Inundation Validation: Red-tier exceedance periods were associated with detectable floodplain inundation extents ranging from 50 to 149 km² in the lower White Volta floodplain.
• Bagre Reservoir Storage Proxy: The cloud-corrected JRC-derived reservoir pool area was validated against independent satellite altimetry, showing a Pearson correlation coefficient of 0.94. The reservoir routinely fills to capacity in most wet seasons, functioning as an explicit model input and seasonal situational awareness rather than a standalone flood trigger.

Contributions

• Presents the first end-to-end flood early warning system design for the White Volta Basin, integrating multi-source data, AI-based discharge forecasting, return-period-calibrated alert generation, satellite inundation validation, and a concrete operational architecture compatible with Ghana's existing institutional infrastructure.
• Incorporates Bagre reservoir upstream storage as an explicit model input and independent alert trigger, derived from freely available JRC Global Surface Water imagery, addressing a critical variable previously absent from formal modeling in the basin.
• Provides the first direct comparison of a locally calibrated ensemble against GloFAS v4 reanalysis discharge at Nawuni, demonstrating its significant underperformance (KGE = -0.526) due to the lack of Bagre Dam release rules in LISFLOOD.
• Specifies an operational architecture that integrates directly into Ghana's existing myDEWETRA-VOLTALARM platform and NADMO dissemination chain, requiring no new institutional infrastructure.

Funding

• National Key Research and Development Program of China (grant no. 2025YFE0212800)
• National Natural Science Foundation of China (grant no. 72521001)
• Natural Science Foundation of Beijing (grant nos. L255011, 8242014)

Citation

@article{Obeng2026From,
  author = {Obeng, Joseph and Yuan, Hongyong and Chen, Tao and Huang, Lida},
  title = {From forecast to alert: an open-data AI flood early warning system with reservoir monitoring for the White Volta Basin},
  journal = {Natural hazards and earth system sciences},
  year = {2026},
  doi = {10.5194/nhess-26-5667-2026},
  url = {https://doi.org/10.5194/nhess-26-5667-2026}
}