Extracting pseudo points by mining soil-landscape relationships embedded in historical soil maps is an important method for updating these maps. However, existing pseudo-point selection strategies suffer from limited representativeness and spatial coverage. To address this, this study developed a soil-typical landscape-position identification algorithm, grounded in pedogenetic environmental characteristics, to extract representative typical pseudo-points. A total of 860 samples were used for training and 172 independent validation samples for validation (60 field-verification samples and 112 randomly selected typical pseudo-points), with the polygon-centroid and random-point methods serving as baselines. Elevation, short-wave infrared, parent material, soil organic matter content, topographic wetness index, and normalized difference vegetation index were selected as key environmental covariates to construct the typical pseudo-point extraction scheme; predictive mapping under different pseudo-point selection methods was performed using a multi-layer random forest. Results showed that, relative to the polygon-centroid and random-point methods, the typical pseudo-point approach markedly improved historical soil map updating accuracy (overall accuracy increased by 18% and 8%, respectively; Kappa by 0.20 and 0.10; and CI decreased by 0.12 and 0.04, respectively). By optimizing pseudo-point selection, this study enhanced the representativeness and quality of pseudo points, offering a methodological reference for updating historical soil maps.