Abstract While most legacy soil maps are available at coarse spatial detail with composite mapping units, high resolution and detailed soil maps are desired for various land resource applications. In time and resource constrained circumstances, the application of disaggregation methods and modelling approaches that capitalise on existing, less detailed soil maps is an important alternative method for a more rapid generation of soil maps at finer resolutions. A legacy soil map of 1:1,000,000 scale for Nepal was disaggregated using “Disaggregation and Harmonisation of Soil Map units through Resampled Classification Trees” (DSMART) tool with the C5.0 classification tree algorithm and an area proportional virtual sampling technique. Environmental covariates sourced from remote sensing, digital elevation model, climatic databases, and national databases were used for predictive mapping of soils. The predicted map was found to show more detailed soil information in comparison to the original soil map. Accuracy assessment with independent datasets showed that the overall accuracy of prediction was 40.4% (51.2% on 3x3 window) while considering the level of Reference Soil Groups only, and 22.1% (32.6% on 3x3 window) for the soil groups with 1st principal qualifier. Geology was the most important covariate, followed by the minimum temperature of the coldest month, elevation, valley depth and land cover. Amidst the scarcity of spatially explicit detailed soil information, this disaggregated soil map can be a useful resource as a more detailed version of the legacy soil map of Nepal for individuals concerned with research, planning and management of land resources. Environmental covariates used in this study may be useful when disaggregating soil maps in similar environmental settings.