A two-step modelling approach to map the occurrence and quantity of soil inorganic carbon

Soil inorganic carbon (SIC) accounts for approximately 30–40% of global soil carbon stocks and up to 90% of total carbon stocks in arid and semi-arid regions. The quantity of SIC can vary considerably over short distances, and SIC primarily accumulates at depth as carbonates, which makes it a challenge to model and map across space. Furthermore, carbonates are not found in all soils, meaning that many datasets of SIC are zero-inflated and highly skewed compared to the distributions of other soil properties. This study uses data from a soil survey performed in 2015 to model and map subsoil (0.3–0.5 m) inorganic carbon content in a semi-arid, irrigated cotton-growing region in the lower Lachlan River valley in south-west New South Wales, Australia. A two-step mixture model is used to overcome the zero-inflated and highly skewed nature of the SIC dataset. Such an approach involves using a random forest model to initially predict the presence or absence of SIC in the study area, then using a separate model to predict SIC content using values from the dataset that are above-zero only. The two maps produced from this process are then combined to create a SIC content map. This two-step method had more accurate predictions when compared to a simple one-step modelling approach. Overall, the two-step mixture model proved useful for mapping SIC content, and shows promise for modelling other environmental properties that have a similar zero-inflated and skewed distribution.