A random forest model for predicting soil properties using Landsat 9 bare soil images

Abstract Digital soil mapping (DSM) provides a low‐cost approach for characterizing the spatial variation in soil properties, which contributes to inconsistent productivity. This study utilized random forest (RF) models to facilitate the DSM of apparent soil electrical conductivity (EC a ), cation exchange capacity (CEC), and soil organic matter (SOM) in agricultural fields across the Lower Mississippi Alluvial Valley based on a bare soil composite of Landsat 9 multispectral imagery and digital elevation models. Model data were collected during nongrowing seasons from 2019 through 2024. Field data included EC a of the upper 0.5 m of soil from agricultural fields ( n = 347) and soil‐test‐estimated values for CEC and SOM collected on a ∼0.4‐ha (1‐acre) grid from the upper 0‐ to 15‐cm depth ( n = 14,349). The RF model utilized a stratified K ‐fold cross‐validation with fivefolds. Stratification by farm was used to ensure each fold in the cross‐validation process contained a representative distribution of data drawn from all farm locations. Data were divided into an 80/20 split for training and testing purposes. Models had moderate accuracy ( R 2 = 0.45, 0.74, and 0.72 for EC a , CEC, and SOM) with moderate predictability (ratio of performance to deviation = 1.35, 1.95, and 1.91 for EC a , CEC, and SOM). The contrasting performance between the CEC and SOM models with the EC a model is likely due to the dynamic nature of soil properties, which is more pronounced in EC a . Accordingly, models could have benefitted from covariates such as soil moisture and climatic factors or higher spectral resolution imagery, such as hyperspectral.