How effective is a hybrid deep learning approach for predicting drought dynamics in Ethiopia?

This policy briefing is based on a study conducted focuses on six drought-prone administrative districts (woredas) located in the arid and semi-arid lowlands of eastern Ethiopia. Most local studies rely on isolated indices, failing to capture the synergistic interactions of drought, a phenomenon driven by interconnected meteorological, agricultural and ecological processes (Mishra and Singh 2010). Consequently, even when machine learning is applied, high-order deep learning architectures remain underutilised and poorly optimised for the unique spatio-temporal complexities of the vulnerable Afar and Somali regions. This creates a critical gap in capturing non-linear interplay between thermal stress and vegetation health. Addressing this, the study, conducted between March and November 2025, developed an integrated deep learning model utilising climate indices. It specifically investigated how these architectures can be optimised to model the complex, non-linear interactions inherent in drought forecasting for moisture-stressed environments, providing an enhanced model for national early warning systems.