A hybrid approach for predictive soil property mapping using conventional soil survey data

Abstract

AbstractSoil property maps are important for land management and earth systems modeling. A new hybrid point‐disaggregation predictive soil property mapping strategy improved mapping in the Colorado River basin by increasing sample size approximately sixfold and can be applied to other areas with similar data, including the conterminous United States. Random forests related environmental raster layers representing soil‐forming factors with samples to predict pH, texture fractions, rock, electrical conductivity (ec), gypsum, CaCO3, sodium adsorption ratio (sar), available water capacity (awc), bulk density (dbovendry), erodibility (kwfact), and organic matter (om) at seven depths as well as depth to restrictive layer (resdept) and surface rock size and cover. Cross‐validation R2 values averaged .53 (range, .20–.76). Mean absolute errors ranged from 3 to 98% of training data averages (mean, 41%). Models of pH, om, and ec had the best accuracy (R2 > .6). Most texture fractions, CaCO3, and SAR models had R2 values from .5 to .6. Models of kwfact, dbovendry, resdept, rock models, gypsum, and awc had R2 values from .4 to .5; near‐surface models tended to perform better. Very‐fine sands and 200‐cm estimates for other models generally performed poorly (R2 = .2–.4), and sample size for the 200‐cm models was too low for reliable model building. Uncertainty estimates were also developed by creating relative prediction intervals, which allow end users to evaluate uncertainty easily. Average error increased in areas with higher relative prediction intervals (higher uncertainty), which also had low sampling densities, suggesting that additional sampling in these areas may improve prediction accuracy. Greater uncertainty was observed in areas with highly stratified shale parent materials and physiographic settings uncommon relative to the broader study area.