Structure-from-Motion (SfM) photogrammetry is one of the most common approaches used to elaborate high-resolution Digital Elevation Models (DEMs) nowadays. Factors that influence the final error associated to the derived DEM are: camera-toground distance, camera-sensor system parameters, image network geometry, matching performance, terrain type, lighting conditions and referencing methods. Here, a strategy focused on minimizing the occlusion produced by topography and determine optimal camera locations for image acquisition is presented. This methodology is based on using a viewshed analysis implemented in a Geographical Information System (GIS) to identify the best images for the SfM workflow of a specific survey-site. The suitability of the workflow presented against conventional acquisition strategies was tested using three different datasets (one terrestrial and two aerial) and analyzing differences between SfMderived DEM produced using: 1) a dataset acquired following conventional overlap requirements (i.e. one image every 5-10o around the target for terrestrial close-range oblique SfM and 7060% frontal and side overlap for aerial surveys), 2) a dataset overloaded with images (i.e. one image every 3-4o around the target and >95-95% frontal and side overlap for aerial surveys), and 3) images selected using the viewshed analysis. The resulting DEMs were tested against Terrestrial Laser Scanner-derived (TLS) DEMs. SfM results showed denser point clouds for the datasets elaborated using the viewshed analysis. Differences were particularly important for the terrestrial case indicating a stronger line-of-sight effect on the ground. Point cloud density absolute differences and no-data zones in the datasets produced using the conventional strategies resulted in larger Mean Absolute Errors (MAE) in the DEMs. DEMs produced using the viewshed criteria showed lower MAEs than the conventional dataset and similar to the dataset overloaded of images. Additionally, the processing time of the datasets that used viewshed criteria was much shorter than the datasets overloaded of images.