Better understanding the spatial distribution of soil organic carbon (SOC) stocks is important for the management and enhancement of soils for production and environmental outcomes. We have applied digital soil mapping (DSM) techniques to combine soil-site datasets from legacy and recent sources, environmental covariates and expert pedological knowledge to predict and map SOC stocks in the top 0.3m, and their uncertainty, across South Australia’s agricultural zone. In achieving this, we aimed to maximise the use of locally sourced datasets not previously considered in national soil C assessments. Practical considerations for operationalising DSM are also discussed in the context of working with problematic legacy datasets, handling large numbers of potentially correlated covariates, and meeting end-user needs for readily interpretable results and accurate maps. Spatial modelling was undertaken using open-source R statistical software over a study area of ~160000km2. Legacy-site SOC stock estimates were derived with inputs from an expert-derived bulk-density pedotransfer function to overcome critical gaps in the data. Site estimates of SOC were evaluated over a consistent depth range and then used in spatial predictions through an environmental-correlation regression-kriging DSM approach. This used the contemporary Least Absolute Shrinkage and Selection Operator penalised-regression method, which catered for a large number (63 numeric, four categorical, four legacy-soil mapping themes) of potentially correlated covariates. For efficient use of the available data, this was performed within a k-fold cross-validation (k=10) modelling framework. Through this, we generated multiple predictions and variance information at every node of our prediction grid, which was used to evaluate and map the expected value (mean) of SOC stocks and their uncertainty. For the South Australian agricultural zone, expected value SOC stocks in the top 0.3m summed to 0.589 Gt with a 90% prediction interval of 0.266–1.086 Gt.