In Ethiopia, drought is a common threat to food and water security making millions depend on humanitarian aid year after year. Both the severity and uncertainties of drought are increasing. Its assessment and measurement is also becoming more complex. Specifically, measuring and forecasting the likelihood of drought of small counties is important for humanitarian response and resource planning. But currently, the approach in use is not simple, accurate, and cost effective. In this regard, we compare performance of models developed using stochastic, classical machine learning, and deep learning techniques on satellite imagery datasets of 55 wordas (places) based on historical 474 months’ time series imagery datasets. Performances of our models reveal that some places (wordas) have some kind of easy learnable patterns compared to some other places. For example models for wordas in Oromia, SNNPR, and Somali regions demonstrate higher prediction accuracy than wordas in Afar and Amhara regions. Specifically, R2 of deep learning models vary from the lowest (42% for Tenta Worda of Amhara to 95% for Boloso Worda of SNNPR). We also found that deep learning models outperform classical machine learning as well as stochastic models for all wordas we considered. From the context of humanitarian aid planning and resource allocation perspective, deep learning models can enable decision makers to make better and informed decisions regarding which locations (wordas) are heading to extreme droughts and which are not. The higher the R2 value a model has the higher the quality of the decision.