In field-, watershed-, and regional-scale projects, soil water retention is often estimated from soil textural classes shown in soil maps. The textural classes are relatively broad, often only the dominating textural class is shown, cartographers routinely use error-prone field judgments of soil texture, and soil texture is known to vary along slopes and to depend on the land surface shape. We, therefore, hypothesized that including topographic information in water retention estimation would increase accuracy. To test this hypothesis, we extracted data on 216 soil pedons for soils of moderate and large extent from the Natural Resources Conservation Service (NRCS) soil characterization database. Textural classes, genetic horizon numbers, slopes, position on the slope classes, and land surface shape classes were the field descriptors that we used to estimate water retention at −33 and −1500 kPa potentials for each horizon in each pedon. Because our input variables were both categorical and continuous, regression trees were used for subdividing the samples into the smallest number of the most homogeneous groups, which we tentatively called topotextural groups (TTG). The jackknife cross-validation was used to prune the regression trees to prevent overparameterization. Ten or fewer TTGs were defined for both the −33 and the −1500 kPa retention. The TTGs were different for the two matric potential levels. Using topographic variables and soil horizon seemed to be the most accurate way to make up for errors made in field determination of texture. For the A horizon, the topotextural grouping resulted in estimates that were more accurate than those using laboratory textures only. Although most of the topographic variables in this work are categorical, those variables seemed to be useful for improving estimates of water retention.