Fine scale maps of soil properties enable efficient land management and inform earth system models. Recent efforts to create soil property maps from field observations tend to use similar tree-based machine learning interpolation approaches, but often deal with depth of predictions, validation, and uncertainty differently. One of the main differences in approaches is whether to model individual depths of interest separately as ‘2D’ models, or to create models that incorporate depth as a predictor variable creating a ‘3D’ model that can make pre-dictions for all depths. It is unclear how choice of 2D or 3D approach influences model accuracy and uncertainty due to lack of direct comparison and inconsistent presentation of results in past studies. This study compares 2D and 3D methods for mapping soil electrical conductivity (salinity), pH, sum of fine and very fine sands, and organic carbon at 30 m resolution for the upper 432,000 km 2 of the Colorado River Watershed of the United States of America. A new, simple, model-agnostic relative prediction interval (RPI) approach to report un-certainty is presented that scales prediction interval width to the 95% interquantile width of the original training sample distribution. The RPI approach enables direct comparison of uncertainty between properties and depths and is easily interpretable by end users. Results indicate that 3D mapping of soil properties with strong variation with depth can result in substantial areas with much higher uncertainty that coincide with unrealistic predictions relative to 2D models, even though 3D models had slightly better global cross-validation scores. Maps and global model summaries of RPI proved helpful in identifying these issues with 3D models. These results suggest that the use of RPI or similar approaches to evaluate models can identify accuracy problems not evident in global va-lidation diagnostics.