Document Type : Original Research
Authors
1
Ph.D Student in Desert Management and Control,Gorgan University of Agricultural Sciences and Natural Recourses, Gorgan, Iran.
2
2. Associate Professor, Department of Arid zone Management, Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, Iran.
3
Associate Professor, Department of Arid zone Management, Gorgan University of Agricultural Sciences and Natural Resources, Gorgan, Iran
4
PhD in Remote Sensing and Geographic Information Systems, National Cartographic Center (NCC), Tehran, Iran.
5
Professor, Department of Remote Sensing and GIS, Faculty of Geography, University of Tehran, Tehran, Iran
Abstract
Dust storms, characterized by their capacity to transport aeolian sediments across extensive distances from their origins, represent significant threats to human societies and are linked with considerable detrimental impacts on public health, ecological systems, and the economic stability of communities. In light of the swift dispersal and extensive distribution patterns of dust particulates, coupled with their transportational dynamics via wind currents, numerous occurrences remain elusive to detection and monitoring, thereby necessitating the elucidation and spatial delineation of their origin regions. The Kara-Bogaz Gol Basin, alongside the Karakum Desert, has emerged as one of the principal contributors to dust emissions affecting Golestan Province in recent years. This study critically assesses the efficacy of sophisticated machine learning algorithms in pinpointing dust emission origins within the Kara-Bogaz Gol Basin. In the present investigation, a robust analytical framework founded on the amalgamation of remote sensing data and machine learning methodologies was employed. Environmental datasets, encompassing nine distinct parameters—namely, dusty days, soil moisture, soil texture, precipitation, wind velocity, Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM), air temperature, and land cover—were systematically extracted and processed utilizing the Google Earth Engine platform across the temporal scope of 2003-2023. A total of 340 dust emission sources were discerned through visual interpretation of MODIS satellite imagery, serving as training data for the machine learning algorithms. The findings revealed the following accuracies in identifying areas with high dust emission potential: Random Forest 91.8%, Artificial Neural Network 70.9%, XGBoost 89.9%, Gradient Boosting 87.9%, Bagged CART 89.9%, and LightGBM 91.8%. Notably, Random Forest and LightGBM exhibited superior performance in the identification of dust sources. An examination of explainability techniques indicated that three variables—vegetation index (27% contribution), soil moisture (23%), and DEM (19%)—exerted the most substantial influence on the prediction of dust-prone areas. The primary contribution of this research resides in the formulation of a hybrid machine learning framework adept at identifying regions susceptible to dust emissions and quantifying the role of environmental parameters in the genesis of these storms.
Keywords
Subjects
The Journal of Spatial Planning and Geomatics