Zohre Ebrahimi-Khusfi

It is necessary to predict wind erosion events and specify the related effective factors to prioritize management and executive measures to combat desertification caused by wind erosion in arid areas. Therefore, this work aimed to evaluate the applicability of nine machine learning (ML) models (including multivariate adaptive regression splines, least absolute shrinkage and selection operator, k-nearest neighbors, genetic programming, support vector machine, Cubist, artificial neural networks, extreme gradient boosting, random forest) and their average for predicting the seasonal dust storm index (DSI) during 2000–2018 in arid regions of Iran. The results showed that the averaging method outperformed the other individual ML models in predicting DSI changes in all seasons. For instance, the averaging methods improved the prediction accuracies for winter, spring, summer, autumn, and dusty seasons by 22%, 39%, 28%, 32%, and 26%, respectively, compared to the multivariate adaptive regression splines. Furthermore, the most important factors in predicting DSI were detected as follows: wind speed for winter, enhanced vegetation index for spring, maximum wind speed for summer, autumn and dusty seasons. In general, our results indicate that the combining of the individual ML models by averaging method help us to develop a more accurate approach for predicting the temporal changes of the dust events in arid regions. Furthermore, the obtained results in this study can be applicable for prioritizing measures in order to minimize the dangers of wind erosion based on the major driving factors.