LAUSR

Abstract Because of the lack of ground-based observations in large parts of West Asia, Aerosol Optical Depth (AOD) is mainly monitored by using remote sensing techniques. AOD can also be predicted by short term forecasts with commonly called Deterministic weather prediction models (DMs). The skill of DMs in reproducing remotely sensed observations when averaged over monthly time scales over West Asia is rather limited due to significant uncertainties in inputs and complexity of dust, which is the dominant type of aerosols in the region. Machine Learning Algorithms (MLAs), which require much less computational expenses than DMs, can be used. Using Moderate Resolution Imaging Spectroradiometer (MODIS) Deep Blue (DB) AOD as the representative of response variable, MLAs, especially Multivariate Adaptive Regression Splines (MARS) and Support Vector Machines (SVM), outperformed DMs on monthly time scale. MLAs have yielded lower prediction error (RMSE) and higher correlation with observations than DMs. In addition, findings disclosed that DMs, especially MACC, have failed to simulate observed AOD values over western Iran where the Zagros Mountains prevent advection of fine dust particles to the east of the study area. Prediction errors of MLAs and DMs along with major DB AOD peaks, over Iraq, can be traced back to the rough resolution of variable datasets, omission of some unknown influential predictors representing the life cycle of dust and/or other aerosols, and scarcity of extreme cases. It also remains to be tested in how far the results presented can be generalized to other regions and time scales.