Non-raster scanning has been proved to be an effective approach to dramatically improve the efficiency of the scanning probe microscopy without sophisticated scanners and controllers. However, the metrological performance of the method largely depends on the posterior data processing techniques which are responsible for recovering measured surfaces from non-gridded and subsampled data. This paper casts the surface reconstruction as a regression problem and proposes a Gaussian process regression model to study the spatial correlation of the non-gridded data and to give credibility to the recovered surfaces in terms of covariance matrix. The statistical nature of the Gaussian process regression offers great flexibility in reconstructing various topographies with high accuracy by designing task-specific covariance functions. Comparisons with the well-known Delaunay triangulation-based interpolation method verify the effectiveness of the proposed method in enhancing the metrological performance of non-raster scanning microscopy.