Mapping available soil water capacity in New South Wales, Australia using sparse data - An inverse Bayesian approach

Data sparsity is a major limitation in modelling soil hydraulic properties across large spatial extents, such as the continent of Australia, and can lead to erroneous models being developed. Inverse modelling can be used to make inferences about underlying environmental processes, even when the data is subject to error. Bayesian inverse modelling is one method to quantify uncertainties from multiple sources, including data, model structure, and model parameters. Although, Bayesian inverse modelling approaches have been proven useful in many applications, they are still not commonplace in digital soil mapping (DSM). In this work, we present a Bayesian inverse model for predicting available water capacity (AWC) of soils using sparse data. Specifically, we estimate upper and lower drainage limits (DUL and LL15) using a Bayesian geostatistical model. DUL and LL15 are then differenced to estimate the AWC. This study was carried out using measurements of upper and lower drained limits (DUL & LL15) across NSW, Australia. Two separate Bayesian inverse models were calibrated and validated for the DUL and LL15 soil hydraulic parameters. The calibrated models were then used to spatially predict AWC of soils across a cropping region in New South Wales (NSW), Australia on a 90-m grid. We also present the uncertainty of the parameters and output maps. Our results gave similar estimates of AWC values when compared to those predicted using the Cubist machine learning model. Although the predictions are similar, the Bayesian method allows for the direct quantification of the prediction uncertainties, while the Cubist method cannot.