Identification of dust sources in a dust hot-spot area in Iran using multi-spectral Sentinel 2 data and deep learning artificial intelligence machine

Abstract

The drying of wetlands in Iran due to climate change and indiscriminate human activities has increased dust production. Dust storms have become a major problem in arid and semi-arid regions and cause adverse social, economic, and environmental effects. The Jazmurian wetland in Kerman Province is one such area. To identify dust sources in the Jazmurian basin, high resolution Sentinel 2 data were used. From these, sediment supply was mapped. Three artificially intelligent algorithms—artificial neural network (ANN), support vector machine (SVM), and deep-learning neural network (DLNN)—were used to model dust-production potential in the study area. The results show that portions of the Jazmurian basin that have dried up in recent years have a very high potential for dust production. Evaluation of the models’ performances using area-under-curve (AUC) statistics revealed that the DLNN model is more efficient (AUC = 0.97) than either the ANN (AUC = 0.91) or SVM (AUC = 0.92). All three models reveal that NDVI, elevation, annual rainfall, and windspeed are the four most important factors influencing dust-production potential in the study area. This remote sensing-artificial intelligence framework should be tested for mapping dust-production potential in other regions as this study demonstrates highly accurate, high-resolution results. This study yielded fundamental information to identify locations in need of desertification management and mitigation of dust production in the Jazmurian basin.