Predicting the spatio-temporal distribution of absorbable heavy metals in soil is needed to identify the potential contaminant sources and develop appropriate management plans to control these hazardous pollutants. Therefore, our aim was to develop a model to predict soil adsorbable heavy metals in arid regions of Iran from 1986 to 2016. Soil adsorbable heavy metals were measured in 201 samples from locations selected using the Latin hypercube sampling method in 2016. A random forest (RF) model was used to determine the relationship between a suite of geospatial predictors derived from remote sensing and digital elevation model data with georeferenced measurements of soil absorbable heavy metals. The trained RF model from 2016 was used to reconstruct the spatial distribution of soil absorbable heavy metals at three historical timesteps (1986, 1999, and 2010). Results indicated that the RF model was effective at predicting the distribution of heavy metals with coefficients of determination of 0.53, 0.59, 0.41, 0.45, and 0.60 for Fe, Mn, Ni, Pb, and Zn, respectively. The predicted maps showed high spatio-temporal variability; for example, there were substantial increases in Pb (the 1.5–2 mg/kg⁻¹ class) where its distribution increased by ~25% from 1988 to 2016—similar trends were observed for the other heavy metals. This study provides insights into the spatio-temporal trends and the potential causes of soil heavy metal contamination to facilitate appropriate planning and management strategies to prevent, control, and reduce the impact of heavy metal contamination in soils.

Soil organic carbon (SOC) is an essential property of soil, and understanding its spatial patterns is critical to understanding vegetation management, soil degradation, and environmental issues. This study applies a framework using remote sensing data and digital soil mapping techniques to examine the spatiotemporal dynamics of SOC for the Yazd-Ardakan Plain, Iran, from 1986 to 2016. Here, a conditioned Latin hypercube sampling method was used to select 201 sampling sites. A set of 37 environmental predictors were obtained from Landsat imagery taken in 1986, 1999, 2010 and 2016. Here, SOC was modeled for 2016 using the Random Forest (RF), support vector regression (SVR), and artificial neural networks (ANN) machine-learners by correlating environmental predictors with soil data. The results showed that RF yielded the highest accuracy (R² = 0.53), compared to the other two learners. By performing a variable importance analysis of the RF model, normalized difference vegetation index, modified vegetation index, and ground-adjusted vegetation index were determined to be the most important environmental predictors. By applying the model calibrated from 2016 data to 1986, 1999 and 2010, the results showed a substantial decrease in SOC; these decreases in SOC were mainly attributed to land use changes and agricultural activities.