Machine-learning algorithms for predicting land susceptibility to dust emissions: The case of the Jazmurian Basin, Iran

In this study, we apply six machine-learning algorithms (XGBoost, Cubist, BMARS, ANFIS, Cforest and Elasticnet) to investigate the susceptibility of the Jazmurian Basin in southeastern Iran to dust emissions. This research is the first attempt to apply several machine-learning techniques (e.g., BMARS, ANFIS, Cforest and Elasticnet) to mapping of dust emissions from land surfaces. Fourteen parameters associated with meteorology, lithology, soil, and human activity were considered as potentially effective dust emission factors implemented in our modelling. Collinearity among the parameters and their weighted importance were examined statistically. To evaluate the accuracy of our predictive models and their performance, we applied the Taylor diagram (involving RMSE and correlation coefficient), the Nash Sutcliffe coefficient (NSC), and mean absolute error (MAE). The prediction accuracy of the six algorithms for identifying susceptibility to dust emissions, as assessed by the Taylor diagram, was as follows: Cforest (NSC = 98% and MAE = 3.2%) > Cubist (NSC = 90% and MAE = 10.6%) > Elasticnet (NSC = 90% and MAE = 10.7) > ANFIS (NSC > 90% and MAE = 11%) > BMARS (NSC = 89% and MAE = 11.2%) > XGBoost (NSC = 89% and 11.3%). Based on the map produced by Cforest (i.e., the best-performing algorithm in our assessment), we identify four dust susceptibility classes, and their respective total areas ranging from low (32%), moderate (8.2%), high (10%), to very high (50%). We identify the dry lakebed of Hamun-e-Jaz Murian as the most productive area for dust emissions.