Extended Data Fig. 6: Data coverage and void-filling with Gaussian process (GP) regression.

(a) Black areas denote regions with 3D photogrammetric (SfM) constraints from the 1936/1938 aerial images (Fig. 1) and white regions denote void areas infilled with the GP regression (Methods). The SfM-generated point clouds have void areas because of occlusion and poor feature matching in low-contrast areas. There is no reconstruction for the eastern portion of Austfonna (Fig. 2), since no photographs of that region were acquired during the 1936/1938 expeditions⁷ (Extended Data Fig. 3a). (b-e) An illustration of the void filling procedure, applied to Oscar II Land in western Svalbard. To fill the holes in the 1936 DEM, we first compute the ∆h map, differencing the ∼2010 reference DEM to the 1936 reconstruction (b). Next, we train a GP regression to estimate the ∆h values in the void areas. The GP regression is trained using x, y, and z (the 2010 elevation) as predictor variables to infer ∆h as the response variable, and thereby incorporates both the spatial information in (a) and the elevation-dependence of ∆h illustrated in (c). (d) The error of the GP-regression-infilled values is estimated on random subsets of data points (60%) held-out from model training. Finally, subtracting the infilled ∆h map in (e) from the 2010 reference DEM yields the 1936 surface (Extended Data Fig. 3d).