Ensemble Approach to Provide Uncertainty Estimates of Soil Bulk Density

Large scale environmental impact studies typically involve the use of simulation models and require a variety of inputs, some of which may need to be estimated in absence of adequate measured data. An important input, soil bulk density (Db) affects conditions for soil aeration, solute transport, and storage as well as the outcome of soil C stock calculations. Correct representation of Db in simulation studies is essential since any bias or uncertainty will propagate through a variety of processes and time steps. We used the U.S.-wide NRCS National Soil Survey Characterization (NSSC) database of point measurements and the ‘k-nearest neighbor’ (k-NN) pattern recognition algorithm combined with random resampling without replacement to estimate Db and its uncertainty. Soil particle-size distribution and organic C content were utilized as inputs and soil taxonomy classification, sample depth, and soil horizon notation were tested as optional grouping and limiting factors to the calibration (reference) data. We obtained an overall root-mean-squared error (RMSE) of 0.17 g cm−3 and mean error (ME) of 0.01 g cm−3 Grouping samples by taxonomic classification proved to be advantageous, while limiting samples by depth helped avoid depth-specific bias in the estimations. Grouping samples by horizon notation did not yield significant improvement due to the great variability of bulk density within horizons of the same notation. Varying the proportion of data used in the resampled data subsets can be used to establish greater or lesser degree of confidence in mean estimates without affecting those mean estimates. Simulation based environmental impact and risk assessment studies can be direct beneficiaries of data with characterized uncertainty.