Using visibility analysis to improve point density and processing time of SfM‐MVS techniques for 3D reconstruction of landforms

Abstract Image network geometry, including the number and orientation of images, impacts the error, coverage, and processing time of 3D terrain mapping performed using structure‐from‐motion and multiview‐stereo (SfM‐MVS). Few studies have quantified trade‐offs in error and processing time or ways to optimize image acquisition in diverse topographic conditions. Here, we determine suitable camera locations for image acquisition by minimizing the occlusion produced by topography. Viewshed analysis is used to select the suitable images, which requires a preliminary digital elevation model (DEM), potential camera locations, and sensor parameters. One aerial and two ground‐based image collections were used to analyse differences between SfM‐MVS models produced using: (1) all available images (ALL); (2) images selected using conventional methods (CON); and (3) images selected using the viewshed analysis (VIEW). The resulting models were compared with benchmark point clouds acquired by a terrestrial laser scanner (TLS) and TLS‐derived DEMs. The VIEW datasets produced denser point clouds (28–32% more points) and DEMs with up to 66% reduction in error compared with CON datasets due to reduction of gaps in the DEM. VIEW datasets reduced processing time by 37–76% compared with ALL, with no reduction in coverage or increase in error. DEMs produced with ALL and VIEW datasets had similar slope and roughness, while slight differences that may be locally important were observed for the CON dataset. The new method helps optimize SfM‐MVS image collection strategies that significantly reduce the number of images required with minimal loss in coverage or accuracy over complex surfaces. © 2020 John Wiley & Sons, Ltd.