Non-parametric imputation of properties for soil profiles with sparse observations

Abstract

Soil profile data are a collection of soil property values associated with a series of non-overlapping depth intervals. The CSIRO National Soil (NatSoil) Archive database contains full soil profile data recorded for over 56,000 such depth intervals at approximately 9500 sites around Australia. Another database developed by CSIRO uses spectroscopic estimates of soil attributes for soil sampled as part of Geoscience Australia's Geochemical Survey of Australia. That survey is made up of soils collected at two depth intervals (0–0.1m and 0.6–0.8m) from 2244 different sites in Australia. The key question of interest is how complete soil profile data can be used to impute or “fill in” missing soil depth interval data to increase the utility of the two-depth spectroscopic data for a range of environmental analyses. We demonstrate our approach through using the complete NatSoil database to impute missing two-depth interval data to create complete profiles of total phosphorus, total nitrogen and total potassium. A parametric modeling approach to imputation was initially considered but having data at just two depth intervals led us to non-parametric approaches. We simulated from a large quantity of complete profile data with similar first and second order properties as the original data and drew random samples from the simulated data to predict (impute) the two-depth data with quantified certainty for each intervening Global Soil Map depth interval and profile. The complete imputed profiles can be used in future modeling and mapping. We believe our imputation procedure can be extended to other scenarios in soil science where joint imputation of multiple soil properties can be used to fill the gaps arising from incomplete soil profiles.