Digital mapping of soil moisture retention properties using solely satellite-based data and data mining techniques

Soil moisture retention is an important environmental factor that controls water availability in agro-ecosystems. Comprehensive information on spatial distribution and patterns of soil properties controlling moisture retention such as organic carbon (OC), clay content, and saturation percentage (SP) are crucial for effective land management and sustainable development. This study seeks to employ two data mining algorithms named Multivariate Adaptive Regression Splines (MARS) and Gene Expression Programming (GEP) as predictive models in digital soil mapping (DSM) and quantify the associated uncertainty at a grid resolution of 30 m using satellite-based covariates. For each model, the features selected based on their interior algorithm during the training of the models. The performance and accuracy of MARS and GEP were evaluated through nine statistical/quantitative and graphical criteria, including mean error (ME), mean absolute error (MAE), Root mean squared error (RMSE) and coefficient of determination (R-2), relative RMSE (RMSE%), Taylor diagram, scatter, curve fitting, and point density plots. For each model, the prediction maps of soil properties and their associated uncertainty maps were generated. The results revealed that MARS outperformed GEP in providing predictions with superior performance and accuracy. Moreover, MARS performed better in showing the spatial distributions and patterns of all the studied soil properties. In addition, the MARS model produced less prediction uncertainty and can predict soil moisture retention properties more accurately. This study highlights the key role of data mining/numerical modeling algorithms as predictive models in DSM toward the most accurate predictions. Moreover, the study expressed the capabilities of remote sensing derived data in predicting complex soil properties. This study opened a new research line for accurate DSM.