HydroSphereAI Case Study: Sauble River at Allenford — Spring Melt 2025
In late March and in the first days of April 2025, the Grey Sauble Conservation Authority issued an “All Watersheds” Flood Watch in anticipation of significant rainfall and elevated flows. A forecasted weather system was expected to bring up to 50 mm of total precipitation, following weeks of already saturated conditions. For the Sauble River at Allenford (Station 02FA004), this setup resulted in two distinct streamflow peaks within a five-day span— first on March 30, then again on April 3.
The Grey Sauble watershed in southern Ontario is located roughly between Lake Huron and Georgian Bay. Station 2FA004 is found on the Sauble River in the town of Allenford.
As this weather event unfolded, Aquanty’s HydroSphereAI (HSAI) machine learning-based streamflow forecasting platform tracked and predicted both peaks in near real time, providing ensemble forecasts that supported local preparedness efforts. The forecasts from March 28 and 29, issued ahead of the peaks, aligned closely with observed flows and predicted the first peak well, with daily mean values between 65 and 70 m³/s— closely aligned with observed streamflows. While daily averages tend to smooth sharp increases in flow, the forecasts accurately captured the timing and structure of the event. Forecasts remained stable across these days, giving local stakeholders advance notice of rising flows.
The second peak, which occurred on April 3, was more difficult to predict at longer lead times. Early forecasts from March 26 and 27 underpredicted the median outcome, although the actual flow did fall within the upper end of the ensemble range. As new precipitation forecasts became available, HydroSphereAI rapidly adjusted. By April 2, the forecast had sharpened significantly—demonstrating a clear shift from wide uncertainty to focused accuracy. The ensemble spread narrowed meaningfully, and the median value converged toward the observed outcome, offering a high-confidence signal just ahead of the event. This improvement wasn’t just technical— it provided timely, actionable information for decision-makers, showcasing how the system evolves in real time to reflect changing conditions. This evolution highlights a core strength of HydroSphereAI: its probabilistic ensemble forecast design allows users to interpret uncertainty and risk, rather than relying solely on a single deterministic forecast
Looking at both peaks, HydroSphereAI consistently delivered strong performance in predicting the structure and timing of the events. While peak magnitudes— especially for the second event— were underpredicted at longer lead times, the platform successfully indicated the potential for high flows and provided valuable advance notice. As precipitation forecasts from Environment Canada became more aggressive— rising from initial conservative estimates around 20 mm/day to observed totals above 40 mm/day for the first event and over 25 mm/day for the second— HSAI’s forecasts adjusted accordingly. This highlights another important consideration: HydroSphereAI’s accuracy is closely tied to the reliability of the weather forecasts (in this case ECCC’s 16-day GEPS forecast ensemble). As forecasts from the Meteorological Service of Canada are updated, HSAI integrates those changes to improve its outputs. Looking ahead, we are also actively exploring additional sources of meteorological forecasts to further enhance system accuracy, flexibility, and resilience across a range of conditions.
Contrary to initial assumptions, snowmelt played a limited role in the spring 2025 event. While earlier reports noted elevated snowpack in earlier in the year, by March 28 the average snow water equivalent (SWE) had dropped to approximately 8.7 mm in the watershed. The primary driver of both streamflow peaks was precipitation, not melting snow, reinforcing the need for real-time data to guide interpretation and planning.
Importantly, the usefulness of HydroSphereAI’s forecasts extended beyond precise magnitude prediction. The platform consistently indicated the correct timing of flood risk, flagged both events with meaningful lead time, and captured uncertainty through its ensemble structure. For conservation authorities and other users, this allowed for proactive decision-making and clear communication with stakeholders— even when exact peak flow volumes were difficult to pinpoint.
In the Sauble River watershed, HydroSphereAI provided a dynamic, evolving picture of flood potential across a multi-day event window. It helped frame the conversation around risk before peak conditions were reached and offered a detailed retrospective view for validating forecast quality after the event. As hydrologic systems grow more variable under a changing climate, this type of responsive, ensemble-based forecasting will continue to be a critical tool for effective water resource management.