The Strengthening Early Warning Systems for Anticipatory Action (SEWAA) project in Ethiopia applies machine-learning post-processing to numerical weather prediction (NWP) outputs in order to generate improved, localised, probabilistic rainfall forecasts that are linked to pre-agreed triggers for anticipatory humanitarian activation. The project is a multi-partner collaboration led by the World Food Programme (WFP) and the IGAD Climate Prediction and Applications Centre (ICPAC), working in close partnership with the Ethiopia Meteorological Institute (EMI), the European Centre for Medium-Range Weather Forecasts (ECMWF), the University of Oxford, and other technical partners, with grant funding from the Government of Denmark and support from Google.org.
The core AI technique employed is the use of conditional generative adversarial networks (cGANs) to post-process global ensemble and NWP rainfall forecast outputs, primarily from the ECMWF Integrated Forecasting System (IFS). The cGAN model takes coarse-resolution global forecast data as input and generates high-resolution probabilistic rainfall forecasts tailored to specific localities in Eastern Africa, including Ethiopia. This machine-learning post-processing approach addresses a critical gap: while global NWP models provide broad forecasts, they lack the spatial resolution and local calibration needed for actionable early warning at the woreda (district) level. The cGAN architecture learns the statistical relationship between large-scale forecast fields and observed local rainfall patterns from historical meteorological and observational datasets, enabling it to generate ensemble-like probabilistic output that captures forecast uncertainty. The resulting forecasts are designed to be more accurate and more actionable than raw NWP outputs for local decision-makers.
The forecasts generated through this AI post-processing pipeline feed directly into the anticipatory action governance framework established under the United Nations Office for the Coordination of Humanitarian Affairs (OCHA). This framework operates on three pre-agreed pillars: triggers, activities, and financing. When AI-enhanced forecasts indicate that rainfall conditions are likely to cross pre-defined hazard thresholds, this triggers the release of pre-arranged financing and the implementation of pre-agreed anticipatory activities. The trigger mechanism employs a two-step process with committee oversight: initial model-based trigger identification is followed by endorsement by coordination mechanisms before activation proceeds. This governance architecture ensures that AI model outputs do not autonomously drive humanitarian response but rather inform a structured human decision-making process.
In September 2024, this system contributed to an anticipatory action activation in Ethiopia's Somali region, triggered by forecasts of La Nina-induced drought conditions during the October-December 2024 period. The July 2024 meteorological forecast indicated that 49 out of 76 woredas met the threshold to trigger anticipatory actions for severe drought scenarios. With approximately 19 million people in areas at risk of drought nationally, WFP activated responses across 15 woredas in the Somali region. The activation deployed four intervention types: early warning dissemination translating EMI forecasts into actionable advisories, cash transfers of USD 228.27 per household over three months, livestock support including rangeland management and feed vouchers, and water infrastructure rehabilitation. The total activation cost was USD 6.6 million, funded through Germany and Denmark's multi-year funding commitments. The Somali Disaster Risk Management Bureau led the initiative in consultation with regional technical working groups.
The AI post-processing component operates in a cloud-based compute environment to support national meteorological services, although the specific cloud provider and data jurisdiction have not been publicly specified. The system is designed to reduce forecast generation time and costs compared to traditional methods, while building capacity within national agencies such as EMI to generate and interpret ML-enhanced forecasts. ICPAC serves as the regional hub for the forecasting system, with the SEWAA project having initially piloted in Kenya and Ethiopia before expanding to Uganda and Rwanda in March 2025.
The SEWAA project represents an implementation-stage deployment where the AI forecasting model has been integrated into operational anticipatory action workflows with demonstrated real-world activation. The project's mid-term achievements, as documented by WFP in September 2025, include reduced forecast generation time and costs via cloud-based processing, capacity-building for national meteorological agencies, and the evidenced Ethiopia activation in September 2024. Post-activation evaluation and learning loops form part of the governance framework, ensuring continuous improvement of both the AI model performance and the anticipatory action response protocols.
From a social protection perspective, the SEWAA system operates at the intersection of climate forecasting and adaptive social assistance. The anticipatory cash transfers and in-kind support provided through activation constitute emergency social assistance measures triggered by AI-enhanced risk assessment. The system exemplifies how predictive AI can be embedded within humanitarian governance frameworks to enable proactive rather than reactive delivery of social protection in climate-vulnerable contexts, particularly in low-income countries with limited domestic forecasting capacity.