Deep Multi-task Learning For Early Warnings Of Dust Events Implemented For The Middle East

This study proposes a meteorology-based Deep Neural Network model for dust events forecasting in 12-72 hours lead time. The model’s multi-task framework is designed to predict regional PM fields and Local in situ measured PM. The model is able to detect 76% of the dust events at 67% precision in 24 hours ahead. The Middle East is subjected to dust storms that originate from several sources, with the Sahara desert being the largest. In Israel, North African dust contributes 60-80% of the eolian material, while arid and semi-arid regions to the east of Israel, and local dust sources contribute the rest.