摘要： Measuring river geometry and its evolution through time has always been a cornerstone of fluvial geomorphology. While experimental and numerical modeling of fluvial dynamics has been central in understanding long-term dynamics and testing ideas, they remain simplified versions of complex natural systems and cannot necessarily include all relevant processes. Field measurements are thus central to our understanding of elementary processes such as sediment entrainment and deposition, bank erosion, bedrock incision as well as the macroscopic dynamics of river reaches such as channel bed accretion/erosion, bedforms mobility, and river meandering. It is therefore not surprising that fluvial geomorphologists have quickly embraced the use of terrestrial laser scanner (TLS) to study rivers (e.g., Heritage and Hetherington, 2007; Hodge et al., 2009a). TLS allows 3D digitization of fluvial environment in a dense (sub-cm), accurate (mm precision), and nearly exhaustive way (Fig. 1). The very large range of spatial scales covered is particularly impressive, from individual pebbles to km long river reaches (e.g., Brasington et al., 2012). Sub-cm accuracy also offers the possibility of detecting very subtle changes (Lague et al., 2013), a key attribute to measure slow processes such as bedrock abrasion (Beer et al., 2017). Given the recent emphasis on the role of riparian processes on fluvial processes, the ability to digitize vegetation in 3D in relation to channel morphology offers a unique perspective in biogeomorphology. However, many of the promises of TLS have not really been fulfilled, and the scientific potential of the TLS dataset remains often untapped. This is largely due to the challenging aspects surrounding the processing of TLS data which, to a large extent, also apply to structure from motion (SfM) surveys (Passalacqua et al., 2015). Three challenges, akin to typical Big Data issues can be identified as follows: 1. Data Complexity: TLS data are 3D data and nearly exhaustive. This makes for very rich data but also extremely complex to process as the relevant information (e.g., ground, grains, riverbanks, vegetation) must be detected prior to scientific analysis (Fig. 1). TLS data is also natively non-regularly sampled, with strong spatial variations in point density and requires processing methods that are more complex than for 2D raster-based data such as satellite imagery. 2. Data Volume: the latest generation of TLS instruments generates billions of points in a day. Manual processing cannot realistically be applied, and automatic processing methods are paramount. This requires good programing skills as well as a culture of machine learning and computer vision approaches that are not necessarily part of the training of geomorphologists and requires bridging the gap with computer sciences. 3. Data Incompleteness: despite the very large field of view of TLS sensors, the resulting 3D data do not sample the entire surface (Fig. 1). The ground-based viewpoint imparts missing data behind obstacles (grains of any size and vegetation) and the laser is generally fully absorbed by water resulting in the lack of bathymetric data, a strong limitation in river environments. Processing methods must account for this lack of information.