Predicting wind erosion rate using portable wind tunnel combined with machine learning algorithms in calcareous soils, southern Iran

Wind erosion is a critical factor in land degradation worldwide, particularly in arid and semi-arid regions of southern Iran, which have been severely exposed to wind erosion in the recent years due to climate change and land use changes. The main objective of the present study was to predict the wind erosion rate (WER) using easily measurable soil properties combined with some data mining approaches. For this purpose, the WER was measured at 100 locations with different land uses and soil types in the Fars Province, southern Iran using a portable wind tunnel. The WER was predicted by multiple linear regression (MLR), support vector regression (SVR) and decision tree (DT) algorithms using easily measurable soil properties. Results revealed that land use and soil type had significant effect on the WER. The highest mean WER was observed in Entisols with the lowest organic matter (OM), the lowest penetration resistance (PR) and the lowest aggregate mean weight diameter (MWD). Bare lands with the lowest OM and MWD showed the highest WER compared to other land uses. R² and RMSE of the non-linear regression models developed based on the type of the relationship between the WER and easily measurable soil properties improved by 15% and 12%, respectively, compared to the linear regression model. In both train and test datasets, the SVR and DT models coupled to a genetic algorithm (GA) used for selecting the effective easily measurable soil properties had higher performance than the SVR and DT models using all easily measurable soil properties for predicting WER. With respect to statistical indices, the SVR model with R² = 0.91 and RMSE = 0.68 g m^-2 s^-1 outperformed the MLR and DT for predicting the WER. We concluded that combining the SVR with GA could be an applicable and promising method for predicting WER.