Integrating environmental predictors and deep learning approach for spatiotemporal dust storm risk mapping and impacts modeling on land, health and food security

Abstract

Dust storms originating from the desiccated bed of Lake Urmia pose a serious threat to public health, agriculture, and ecosystem sustainability. This study assessed dust storm susceptibility using four hybrid models integrating deep learning and machine learning approaches, namely CNN–RF, CNN–LR, SVM–RF, and SVM–LR, developed from 16 environmental predictors and verified dust occurrence and non-occurrence samples. The model performance was evaluated using standard validation metrics, which revealed that the CNN-based hybrid models consistently outperformed the traditional machine-learning combinations. The resulting susceptibility maps indicate extensive high-risk zones across the basin, with substantial exposure of the population, agricultural lands, and orchards. Feature analyses highlight pH soil, salinity, Normalized Difference Vegetation Index (NDVI), wind speed and sun hours as dominant drivers. The high-resolution maps delineate priority areas for mitigation and land-use planning and offer a transferable framework for dust storm risk assessment in arid and semi-arid lake basins experiencing severe desiccation.