Due to time, budget, and/or technical constraints, geotechnical site investigation data from a specific site are often limited and sparse, leading to a long-lasting challenge in characterization of spatially varying geotechnical properties. During preliminary stages of site characterization, geotechnical data from neighboring sites or sites with similar geological conditions are often collected and used as valuable prior knowledge in geotechnical engineering practice. Nevertheless, existing methods for spatial variability characterization often rely solely on site-specific data and cannot effectively incorporate prior knowledge or existing databases. To address this issue, this study proposes a novel machine learning method that systematically combines sparsely measured data at a specific site with existing data from neighboring sites or sites with similar geological settings for characterization of property spatial variability in a data-driven manner. The proposed method starts with the construction of a dictionary that draws the dominant spatially varying patterns from a property measured at sites with similar geology under a dictionary learning framework. Leveraging the developed dictionary, the spatial variability of a property is interpreted from sparse site-specific measurements using Bayesian learning. The effectiveness of the proposed method is demonstrated using real data, and improved performance over existing methods is observed.