Comparing the effectiveness of image inpainting techniques over standard interpolation procedures for high-resolution data analysis

The increasing availability of high-resolution Digital Terrain Models (DTMs) in recent years represents a great opportunity for Geosciences being this input information a fundamental prerequisite for an accurate representation of the surface and of the processes acting on it. Nowadays, a number of techniques in order to produce high-resolution DTMs are available. Depending on the scale of analysis, the most used survey techniques are: i) Radar, ii) LiDAR and iii) Photogrammetry. The raw survey-derived products are often in need of a careful post-processing in order to filter out unwanted features. This, for example, could be the case of vegetation removal for surface processes modeling. After the filtering procedures, usually either the point density under filtered areas drops down or missing data areas are created. Resulting interpolated surfaces, in particular over these areas, could vary significantly, mainly depending on the selected interpolation algorithm, sometimes leading to the creation of artifacts that are not able to mimic the original surface trend and texture. In this work, we devised an experiment in order to compare the results of an image inpainting technique over missing data holes against commonly used interpolators (IDW, Spline, Kriging, Natural Neighbor) quantifying the accuracy of the approaches. The image inpainting technique has demonstrated in all the cases a significantly better performance in reconstructing the original surface. An assessment of the extent up to which such an approach could be regarded as robust has also been carried out. Selected applications of surface propagation models to the reconstructed surfaces have been reported so as to highlight how surface variability and uncertainty in surface reconstruction can influence (positively/negatively) model results. Applications of such an approach could pave the way for accurate surface interpolation and/or surface reconstruction in case of the need to remove selected features to characterize several modeling scenarios or to fill the gaps to reproduce the original surface in a consistent way.