Salt dome related soil salinity in southern Iran: Prediction and mapping with averaging machine learning models

In order to manage soil salinity effectively, it is necessary to understand the origin and the spatial distribution of salinity. There are about 120 salt dome outcrops in southern Iran and little is known about their contribution as the potential sources of salts and the spatial pattern of salts around them. Six machine learning algorithms were applied to model topsoil electrical conductivity (EC) and sodium adsorption ratio (SAR) in the Darab Plain (surrounded by six salt domes), Fars Province. Decision trees (DT), k-nearest neighbours (kNN), support vector machines (SVM), Cubist, random forests (RF) and extreme gradient boosting (XGBoost) were used as primary models and the Granger-Ramanathan (GR) method was used to combine the predictions of these models. The results showed that remotely sensed data contributed more to predict EC and SAR than terrain-based data. In terms of root mean square errors (RMSE), Cubist followed by the RF model, tended to give the best estimates of EC, whereas for SAR, RF performed best and was followed closely by SVM and Cubist. Compared to the primary models, the GR method on average resulted in a decrease of 6.1% and 3.9% in RMSE and an increase of 10% and 10.9% in R-2 for EC and SAR, respectively. The spatial pattern of SAR and EC suggested that the contribution of salt domes in soil salinization varied significantly according to their hydraulic behaviour in relation to adjacent aquifers and their activity. In general, the model averaging approach showed the potential to improve the estimates of EC and SAR.

Automated summary

The study utilized machine learning algorithms to model soil salinity, showing that remote sensing data contributed more to predicting electrical conductivity and sodium adsorption ratio, and that model averaging approach has the potential to improve prediction accuracy.