2026/07/01 by M. Dyani Frye, Gordon L. Rees
paper · doi:10.1002/saj2.70300
Abstract Leveraging spectroscopic technology to indirectly estimate important soil characteristics can help overcome the challenge of acquiring accurate, high‐density soils data needed for sustainable land management. This study was conducted to determine whether elemental data from portable x‐ray fluorescence (pXRF) could be used to predict chemical and physical properties of soils from diverse land uses and spanning a large spatial extent within the state of California. Random forest models were constructed to determine the accuracy with which soil properties, including pH, soil texture, cation exchange capacity (CEC), percent total nitrogen (TN %), and percent soil organic carbon (SOC %), could be predicted from pXRF‐derived elemental concentrations. Variable importance plots (VIPs) were also created to determine the most important elemental predictors for each property. Correlation matrices for the overall dataset and two of the geographically related subsets were used to examine the strength and direction of relationships between soil properties and elemental concentrations. Overall, developed models gave fair estimates, and subset‐specific models for two subsets showed further improvement across the majority of evaluation metrics. Model performance was strongest for CEC (cmolc kg −1 ) (root mean square error [RMSE] = 6.79, R 2 = 0.79) and TN % (RMSE = 0.06, R 2 = 0.78) with lower prediction accuracy found for SOC % (RMSE = 1.14, R 2 = 0.74), sand % (RMSE = 10.8, R 2 = 0.66), clay % (RMSE = 6.06, R 2 = 0.63), and pH (RMSE = 0.49, R 2 = 0.49). VIPs and correlation analysis were used to investigate relationships between chemical covariates and target properties. While the acceptable level of error from modeling estimates will depend upon the project goals, random forest models to predict certain properties such as CEC, TN %, and soil texture class show promise over broad geographic ranges using a heterogeneous set of soil data.