AI System Forecasts White Volta Floods Using Reservoir Data

Researchers from Tsinghua University have developed an artificial intelligence-driven flood early warning system for the White Volta Basin in northern Ghana, designed to address a critical gap in current global forecasting models. The study, published in Natural Hazards and Earth System Sciences, presents an end-to-end framework that integrates open data sources, machine learning algorithms, and satellite monitoring of upstream reservoirs to provide more accurate and timely flood alerts for communities in the region.

The White Volta River, which drains approximately 106,000 square kilometers across Burkina Faso and northern Ghana, is highly susceptible to flooding driven by both local rainfall and operational decisions made at the Bagre Dam in Burkina Faso. Existing global warning systems, such as the Global Flood Awareness System (GloFAS), often fail to account for the specific impact of unmonitored dam releases from this upstream reservoir. The new system aims to bridge this gap by explicitly incorporating reservoir state data into its forecasting models, offering a more localized and responsive approach to flood risk management.

The core of the system is an ensemble of machine learning models, including Random Forest, XGBoost, and Long Short-Term Memory (LSTM) networks. These models were trained on a combination of observed gauge discharge data from the Nawuni station, satellite rainfall estimates from CHIRPS, atmospheric reanalysis data from ERA5-Land, and a novel proxy for Bagre reservoir storage derived from the Joint Research Centre (JRC) Global Surface Water dataset. The ensemble approach allows the system to leverage the strengths of different algorithms, improving overall predictive performance and robustness.

In testing on an independent period from January 2004 to February 2007, the ensemble model achieved high accuracy in forecasting river flows. The Kling-Gupta Efficiency (KGE) scores, a standard metric for evaluating hydrological model performance, were 0.984, 0.974, and 0.957 for 1-day, 3-day, and 5-day lead times, respectively. These results indicate that the system can reliably predict discharge levels several days in advance, providing valuable lead time for emergency preparedness and response.

A direct comparison with the GloFAS v4 reanalysis discharge at the Nawuni gauge revealed a significant performance advantage for the locally calibrated ensemble. GloFAS v4 achieved a KGE of -0.526 on the same test period, performing worse than a simple climatological mean predictor. This poor performance is attributed to the fact that the LISFLOOD model used in GloFAS does not represent the release operations of the Bagre Dam, which is a dominant driver of extreme flood events in the basin. The locally calibrated ensemble, by contrast, incorporates the Bagre storage proxy as an explicit input, allowing it to capture the downstream flood signal more accurately.

The system translates continuous discharge forecasts into a four-tier alert system, calibrated to 30-year flood return periods. The tiers are Green (below 1,114 cubic meters per second), Yellow (1,114-1,362 cubic meters per second), Orange (1,362-1,899 cubic meters per second), and Red (above 1,899 cubic meters per second). Cross-validation results show that the system achieves a high probability of detection for Red-tier alerts, with a score of 0.902 at a 1-day lead time and 0.762 at a 5-day lead time. The corresponding false alarm ratios are 0.134 and 0.268, respectively, indicating a reasonable balance between sensitivity and specificity.

To validate the spatial relevance of the discharge-based thresholds, the researchers used Sentinel-1 synthetic aperture radar (SAR) imagery to map flood inundation extents during three validation seasons: 2016, 2019, and 2020. The results showed that periods of Red-tier alert exceedance were associated with floodplain inundation extents ranging from 50 to 149 square kilometers. This spatial grounding confirms that the gauge-level alert tiers correspond to physically meaningful inundation conditions at the community scale, enhancing the system’s practical utility for disaster management.

The Bagre reservoir storage proxy, derived from JRC Global Surface Water imagery, was validated against independent satellite altimetry data, achieving a Pearson correlation coefficient of 0.94. While the reservoir fills to capacity in most wet seasons, making it a less discriminating signal for predicting specific flood years, it serves as an important indicator of seasonal situational awareness. The system uses this signal as an explicit model input and as a trigger for seasonal alerts, providing advance warning that the reservoir is capable of spilling, even if it does not predict the exact timing of a flood event.

The operational architecture of the system is designed to integrate seamlessly with Ghana’s existing institutional infrastructure, specifically the myDEWETRA-VOLTALARM platform operated by the Ghana Meteorological Agency. The system requires no new institutional infrastructure, relying instead on existing data sources and dissemination channels. Alerts are formatted for direct entry into the platform and disseminated through social media, SMS, and community radio, reaching the Community Disaster Volunteer network that forms the last link in the warning chain.

Despite its promising results, the system has several limitations. The independent test period did not include any Orange- or Red-tier discharge events, meaning that the alert skill for these higher tiers is based on cross-validated evidence rather than strictly held-out extreme events. Additionally, the Bagre reservoir signal relies on monthly JRC updates, which introduce a lag of several weeks relative to real-time conditions. The researchers suggest that near-real-time Sentinel-1 monitoring of the reservoir surface area could reduce this latency and improve the system’s responsiveness to changing reservoir conditions.

The study highlights the potential of locally calibrated AI models to improve flood forecasting in data-sparse regions where global models may underperform. By integrating open data sources and satellite monitoring, the system offers a cost-effective and scalable solution for enhancing flood early warning capabilities in West Africa. The findings underscore the importance of considering local hydrological dynamics and operational factors in the design of effective early warning systems.

Future work will focus on extending the training data to include post-2007 flood events and developing near-real-time monitoring of the Bagre reservoir. The researchers also plan to explore the integration of the discharge forecast with a distributed inundation model to provide community-scale depth and extent warnings. This study provides a valuable framework for developing AI-driven early warning systems in other regions facing similar challenges with transboundary water management and data scarcity.

The study area, the White Volta Basin, experiences a single wet season from June to October, driven by the West African Monsoon. Hydrological conditions in the basin are strongly modulated by the Bagre Dam, located approximately 180 km upstream in Burkina Faso. The reservoir, completed in the early 1990s, has a storage capacity of roughly 1.7 billion cubic meters. Because the dam was constructed primarily for irrigation and hydroelectric generation rather than flood control, its release decisions are driven by storage management objectives. Consequently, release events are not subject to formal transboundary notification requirements, creating a significant gap in the current early warning infrastructure for downstream communities in Ghana.

To address this gap, the researchers developed a storage proxy for the Bagre reservoir using the JRC Global Surface Water dataset. The raw satellite-derived water extent initially showed a physically inverted seasonal cycle, with higher values in the dry season due to cloud suppression of optical water detection. To correct this, the team normalized monthly values by the count of valid observations, which restored the expected wet-season filling pattern. The corrected pool area was then validated against independent satellite altimetry data from the DAHITI virtual station, achieving a Pearson correlation coefficient of 0.94 over 135 monthly observations. This validation confirms that the proxy accurately tracks reservoir level, providing a reliable measure of upstream storage state for the forecasting model.

The alert system’s thresholds were derived using a Log-Pearson Type III frequency analysis applied to 30 wet-season annual maxima at the Nawuni gauge. This statistical approach accounts for the mild negative skewness of the discharge distribution, which reflects the physical constraints of the dam-regulated system. Bootstrap confidence intervals indicate substantial sampling uncertainty at lower return periods, with the Yellow threshold carrying a 95% confidence interval of 905 to 1,361 cubic meters per second. These intervals suggest that the exact tier boundaries should be treated as statistically informed estimates rather than precise physical thresholds, highlighting the need for local calibration in operational deployment.

Event-based hydrograph-shape validation further assessed the fidelity of the forecast during flood episodes. Across 13 independent Yellow-tier and above episodes, the system reproduced alert onset timing with a mean absolute error of 1.0 to 2.0 days, depending on the lead time. Peak magnitude was captured with a mean absolute error ranging from 3.5% to 7.9%. However, the analysis revealed that the most extreme peaks were systematically under-predicted, with a mean peak bias of up to -5.6% at the 5-day lead time. This systematic under-prediction of the largest peaks suggests that while the system is effective for early warning, it may require additional calibration to accurately forecast the magnitude of the most severe flood events.

The operational architecture is designed to function within the existing institutional framework of the Ghana Meteorological Agency. The system requires a minimum technical footprint consisting of a Python environment, automated data download scripts, and approximately 50 GB of storage. Alerts are disseminated through a cascade from national to regional and district levels, targeting Community Disaster Volunteer networks within four hours of each daily forecast run. Message formats are provided in Gonja, Dagbani, and English, ensuring that warnings are accessible to the local population. This integration allows the system to leverage existing dissemination channels without requiring new capital investment or institutional restructuring.

Despite the system’s design, the researchers acknowledge that the people-centered layer of the warning system remains a design specification rather than a deployed reality. The impact-based warning matrix, which links discharge thresholds to expected physical impacts and protective actions, is presented as a template for field co-design with local authorities. The precise discharge levels at which specific roads are cut or communities are isolated depend on local topography and asset placement, which only field validation can establish. Therefore, the current study provides the technical foundation for the warning system, but the translation of these statistical thresholds into locally meaningful action levels requires further engagement with district offices and community networks.

Subject of Research: From forecast to alert: an open-data AI flood early warning system with reservoir monitoring for the White Volta Basin

Article Title: From forecast to alert: an open-data AI flood early warning system with reservoir monitoring for the White Volta Basin

Article References: Obeng, J. J., Yuan, H., Chen, T., & Huang, L. (2026). From forecast to alert: an open-data AI flood early warning system with reservoir monitoring for the White Volta Basin. Natural Hazards and Earth System Sciences, 26(11), 5667-5689. https://doi.org/10.5194/nhess-26-5667-2026

Image Credits: AI Generated

DOI: 10.5194/nhess-26-5667-2026

Keywords: forecast, alert, open-data, flood, early, warning, system, reservoir, monitoring, White, Volta, Basin