Assessing the riverine flood forecast skill of GloFAS and Google Flood Hub with impact data and discharge observations to support early actions in Mali

Els Kuipers, Valentijn Oldenburg, Phuoc Phung, Edwin Sutanudjaja, Andrea Ficchi, Peter Salamon, Marc van den Homberg

Riverine floods are among the most destructive and frequent natural hazards in Mali. To reduce their impacts, the Mali Red Cross implemented an Early Action Protocol (EAP) to enable anticipatory actions through pre-defined triggers and forecast information. Currently, the protocol relies on upstream water levels from the National Directorate of Hydraulics (DNH) to predict downstream flooding. However, this model does not consider meteorological forcing, limiting lead times to a maximum of four days. Recent advancements in global flood forecasting systems present opportunities to enhance Mali's EAP by leveraging increasingly skilful medium-range weather forecasts as inputs of both physics-based models, as in the Global Flood Awareness System (GloFAS), and AI-based models, as in Google Flood Hub (GFH). We compare GloFAS v4.0 and v3.0, GFH, and Mali’s current trigger model using discharge observations and district-level impact data derived from multiple sources and text-mined news-articles. Model performance was assessed for a range of lead times and discharge thresholds. GloFAS and GFH demonstrate sufficient skill for early action beyond 4-day lead time in frequently flooded regions and have a larger spatial coverage compared to the current trigger model, suggesting early action plans could operate with 7-day lead time and span a larger area. Overall, this study highlights the potential and challenges of flood forecasting for anticipatory action in flood-prone, data-scarce regions. In particular, we (i) assess the usability of two different ground truths (observed discharge and impact data) for forecast validation; (ii) assess the possibility of extending Mali’s current trigger model in lead time and spatial coverage; and (iii) evaluate user-oriented forecast skill across models and contexts.