Google Flood Hub paired AI flood forecasts with pre-disaster cash aid in Bangladesh

Bangladesh's Jamira island farmers face unpredictable flooding driven by climate change, with traditional warnings giving only three days' notice. Google's Flood Hub uses AI to learn patterns from hydrological and meteorological data, issuing five-day forecasts for river basins in over 80 countries — performance comparable to the physics-based GloFAS model.

GiveDirectly partnered with Google to cross-check AI forecasts against river gauge data and deliver unconditional cash transfers to villagers four days before a flood. The nonprofit mapped flood-prone areas using Google's historical data combined with socioeconomic indicators, targeting villages where over half the area is forecast to flood.

An earlier trial in Nigeria's Kogi State in August 2024 provided early cash to 4,463 people before floods, with many investing in farm improvements and seeds. A separate IRC pilot found Nigerian farmers who received early cash bought more inputs and animals than those who received later aid, and diversified their income.

For 2026, GiveDirectly plans a randomized controlled trial across Gaibandha and surrounding regions in northern Bangladesh covering over 100,000 families. Households would receive 10,375 Bangladeshi taka ($85), with four groups testing different intervention timings to measure which approach best protects livelihoods.

GiveDirectly will send pre-flood cash to 100K+ Bangladeshi families via AI forecasts, based on African pilots where recipients invested in livestock and assets.

Google Flood Hub provides five-day flood warnings using AI patterns rather than physics-based models, reaching river basins without streamflow gauges.

GiveDirectly flips the traditional humanitarian model by delivering aid before disaster strikes rather than after, using AI to pinpoint at-risk villages.

African pilot data showed farmers who received early cash invested in livestock, tools, and seeds — measurably improving livelihoods compared to post-disaster aid recipients.

The randomized trial design across 100,000+ families will produce scientific evidence on whether pre-disaster cash outperforms traditional relief, potentially reshaping humanitarian practice globally.

AI prediction combined with pre-positioned funding lets communities protect livelihoods before disaster strikes, outperforming traditional post-disaster relief delivery.

The randomized controlled trial in northern Bangladesh is planned for 2026, partly funded by Google.org. If results confirm pre-disaster cash outperforms post-disaster relief, the model could scale across Flood Hub's 80+ country coverage. GiveDirectly has already demonstrated the approach works in Nigeria, where early cash recipients invested in farm improvements before flood season.