Dust storms cause widespread damage to social health, economy, welfare, and environment of inhabitants in the affected areas. Although early detection of dust fails to focus on the cause of this phenomenon, it can help develop collective and individual action plans. Therefore, some researchers have become interested in detecting dust concentration in recent years. Accordingly, a hybrid approach based on machine learning was developed in this study to detect dust concentrations by using meteorological and MODIS data. The proposed model consists of two main sections, the first case uses the particle swarm optimization algorithm to extract appropriate features, whereas the second uses innovatively a nonlinear ensemble approach to detect dust concentration and horizontal visibility. First, five machine learning methods included GMDH neural network, a multilayer perceptron neural network, multiple linear regression, a random forest algorithm, and a support vector machine technique are used as reference models. Then, each of the five reference model is used as a nonlinear ensemble model. Also, the common approach of voting-based on ensemble model is used to compare with the result of the proposed approach and to ensure the optimality of the method. Ultimately, it is observed that the ensemble approach has significantly increased the precision of the results with respect to references models.
