Since the first documented soil survey in Tanzania by Milne (J Ecol 35:192-265, 1936), a number of other soil inventory exercises at different scales have been made. The main challenge has been the fragmented nature of the often outdated detailed soil maps and small-scale less-informative country-wide soil maps. Recent advances in information and computational technology have created vast potential to collect, map, harness, communicate and update soil information. These advances present favorable conditions to support the already popular shift from qualitative (conventional) to quantitative (digital) soil mapping (DSM). In this study, two decision tree machine learning algorithms, J48 and Random Forest (RF), were applied to digitally predict k-means numerically classified soil clusters to update a soil map produced in 1959. Predictors were derived from 1 arc SRTM digital elevation data and a 5 m RapidEye satellite image. J48 and RF predicted the soil units of the legacy maps with greater detail. However, RF showed superiority for predicting clusters J48 could not predict and for showing higher pixel contiguity. No significant difference (P = 0.05) was observed between the soil properties of the predicted soil clusters and the actual field validation points. Young soils (Entisols and Inceptisols) were found to occupy about 56 % of the study site’s 30,000 ha followed by Alfisols, Mollisols and Vertisols at 31, 9 and 4 %, respectively. This study demonstrates the usefulness of DSM techniques to update conventionally prepared legacy maps to offer soil information at improved detail to agricultural land use planners and decision makers of Tanzania to make evidence-based decisions for climate-resilient agriculture and other land uses. © Springer International Publishing AG 2016.