Digital Modeling of Three-Dimensional Soil Salinity Variation Using Machine Learning Algorithms in Arid and Semi-Arid lands of Qazvin Plain

Soil salinity, as one of the most important indicators of soil quality, has crucial roles in land use planning and land management in arid and semi-arid regions. The aim of this study was to model soil salinity at five standard depth (0-5, 5-15, 15-30, 30-60, and 60-100 cm) of global digital soil mapping project in 60,000 hectares of Qazvin plain with spatial resolution of 15m. Field studies included a sampling of 278 soil profiles and then the EC was measured in the laboratory. The recursive feature elimination (RFE) method was employed to select environmental covariates including parameters extracted from Landsat 8 image (OLI/TIRS) data, topography, and climatic parameters. Four machine learning algorithms as random forest (RF), cubist (CB), decision tree regression (DTr), and k-nearest neighbors (k-NN) were applied for predicting and mapping soil salinity. According to RFE, 10 covariates were chosen for each standardized depth. The results of modeling showed that the CB model at the depth of 0-5 and 15-30 cm with R2 values of 0.92 and 0.85 and RMSE 4.77 and 7.90 dS/m and the RF model at depths of 5-15, 30-60, and 60-100 cm with R2 values of 0.93, 0.94, 0.96 and RMSE 6.65, 5.10 and 3.20 dS/m, respectively, had the highest accuracy compared to two other models i.e., DTr and k-NN. Furthermore, the covariates extracted from RS data had more impact on topsoil salinity prediction while the climate and topographic attributes influence subsurface soil salinity. Generally, The RF and CB models along with appropriate environmental covariates were able to present salinity variation of study standard depths.