In this study, I investigated the use of digital soil mapping (DSM) techniques for predicting and mapping soil properties. I reviewed two major challenges in DSM: limited observations in many parts of the world and the complexity of soil systems. To address the first challenge, I explored the extrapolation of soil properties from areas with observations to those without, based on similarities in soil-forming factors. For the second challenge, I combined multivariate geostatistics with machine learning (ML) algorithms to model the spatial interdependence of soil variables. The study included three case studies: comparing ML models for small-scale soil mapping, extrapolating ML models to different areas, and mapping salt-affected soils in Hungary using ensemble ML and multivariate geostatistics. The results showed the effectiveness of RF as the best ML model in the first case study, limitations in extrapolation due to soil-landscape complexity in the second case study, and the efficiency of ensemble ML with high-resolution mapping in the third case study. The study highlights the importance of considering scorpan factors and limitations of ML algorithms when applying DSM techniques for soil property predictions.