Dust source susceptibility mapping in Tigris and Euphrates basin using remotely sensed imagery

Abstract

The present study aims to develop an applicable, robust, and generalizable approach building on the integration of machine learning classifiers into dust source susceptibility mapping. Five machine learning classifiers were used for this purpose, including Support Vector Machine, Random Forest, Naïve Bayes, Logistic Regression, and Artificial Neural Network. The study area includes the Tigris-Euphrates river basin, considered one of the world’s largest active dust sources during the last two decades. A total of 411 active dust sources (i.e. hotspots) were identified as ground truth via visual interpretation of sub-daily MODIS-Terra/Aqua time-series imagery from 2000 to 2020 and then used to train and validate the classifiers. The main environmental drivers of dust formation were incorporated based on remote sensing data/products, including Tasseled Cap components, Soil Crust Index, Modified Fournier Index, and the Sentinel-5 Absorbing Aerosol Index. Dempster-Shafer (D-S) theory was used to merge the results from all classifiers to obtain an accurate susceptible dust sources map of the region. Each of the five classifiers and the final integrated D-S model were assessed using the ROC curve, AUC, and Calibration Plot. The mean prediction accuracy of all five classifiers was 87.7% and 86.3% for dust sources susceptible and non-susceptible classes, respectively. The D-S increased the classification accuracy in susceptible and non-susceptible classes by 7.5% and 12.6%, respectively. Finally, the wind erosion thresholds, as a limiting environmental factor in the emission of dust, were applied over the D-S-based susceptible dust sources map to measure the degree of susceptibility in the identified dust sources.