Dust emission forecasting using time series spatial data and machine learning algorithms

The objective of our research is to use remote sensing, reanalyzed time series data, and a machine learning algorithm to forecast dust emissions from sources in the Tigris-Euphrates basin. This basin is one of the most significant global dust source areas and includes more than 10,000 dust hotspots of dust emission sources. To our knowledge, this is the first attempt to predict the emission of dust hotspots one year in advance. The methodology was developed based on an analysis of the spatiotemporal behavior of environmental parameters and the use of machine learning algorithm, including support vector machine, random forest, multinomial Naïve Bayes, and k -nearest neighbors. Additionally, the ensemble model Dempster–Shafer theory was used to merge the results from all machine learning algorithm to obtain dust source emission forecasting. Time series data were used to identify the primary dust drivers, including vegetation cover, precipitation, soil moisture, wind speed, temperature, soil texture, soil thickness, elevation, and slope. The location and temporal behavior (January 2000 to December 2020) of 10,422 dust events, obtained from a previous study conducted by the authors, along with data from January 2021 to December 2021, were utilized as the basis for the analysis. The training data set encompasses the period from January 2000 to December 2020, with January 2021 to December 2021 serving as the forecasting test data. The accuracy of the algorithms for prediction dust source emissions forecasting was evaluated using the AUC-ROC curve, with the random forest model achieving the highest performance of about 69%. Using the Dempster–Shafer model, we combined four machine learning models, and the accuracy reached about 73%, which improved the forecasting accuracy of dust emissions by 4.7%. • This paper studies dust source emission forecasting in the Tigris-Euphrates basin. • First attempt to forecast hotspot emissions one year ahead using ML and environmental data. • Our ML algorithms achieved a 73.48% accuracy, evaluated by the AUC-ROC curve.