Date of this Version
Earth Surf. Process. Landforms 42, 969–986 (2017), DOI: 10.1002/esp.4066.
The availability of high-resolution, multi-temporal, remotely sensed topographic data is revolutionizing geomorphic analysis. Three-dimensional topographic point measurements acquired from structure-from-motion (SfM) photogrammetry have been shown to be highly accurate and cost-effective compared to laser-based alternatives in some environments. Use of consumer-grade digital cameras to generate terrain models and derivatives is becoming prevalent within the geomorphic community despite the details of these instruments being largely overlooked in current SfM literature.
A practical discussion of camera system selection, configuration, and image acquisition is presented. The hypothesis that optimizing source imagery can increase digital terrain model (DTM) accuracy is tested by evaluating accuracies of four SfM datasets conducted over multiple years of a gravel bed river floodplain using independent ground check points with the purpose of comparing morphological sediment budgets computed from SfM- and LiDAR-derived DTMs. Case study results are compared to existing SfM validation studies in an attempt to deconstruct the principle components of an SfM error budget.
Greater information capacity of source imagery was found to increase pixel matching quality, which produced eight times greater point density and six times greater accuracy. When propagated through volumetric change analysis, individual DTM accuracy (6–37 cm) was sufficient to detect moderate geomorphic change (order 100 000m3) on an unvegetated fluvial surface; change detection determined from repeat LiDAR and SfM surveys differed by about 10%. Simple camera selection criteria increased accuracy by 64%; configuration settings or image post-processing techniques increased point density by 5–25% and decreased processing time by 10–30%.
Regression analysis of 67 reviewed datasets revealed that the best explanatory variable to predict accuracy of SfM data is photographic scale. Despite the prevalent use of object distance ratios to describe scale, nominal ground sample distance is shown to be a superior metric, explaining 68% of the variability in mean absolute vertical error.