Dust source mapping using satellite imagery and machine learning models

Predicting dust sources area and determining the affecting factors is necessary in order to prioritize management and practice deal with desertification due to wind erosion in arid areas. Therefore, this study aimed to evaluate the application of three machine learning models (including generalized linear model, artificial neural network, random forest) to predict the vulnerability of dust centers during the years 2005 to 2018 in the Central Desert of Iran. For this purpose, the dust source areas were extracted in the study area using MODIS satellite images using four indicators including BTD3132, BTD2931, NDDI and variable D, and finally 135 hotspots were identified and used in modeling. In this study, conditional factors affecting dust were considered for modeling including land use, soil science, geology, distance from waterway, normalized vegetation difference index (NDVI), land slope and climate. The results showed that among the applied algorithms, random forest with 63.5% accuracy was the most accurate model and followed by artificial neural network with 43.4% accuracy and generalized linear model with 43.2% accuracy. In addition, among factors, land use and soil were identified as the most effective factors on dust source area. The results of this study can provide valuable information for regional managers and policy makers and help them to make useful decisions in management.