Influence of the Number and Spatial Distribution of Ground Control Points in the Accuracy of UAV-SfM DEMs: An Approach Based on Generalized Additive Models

The accuracy of digital elevation models (DEMs) obtained with unmanned aerial vehicle (UAV)-SfM photogrammetry depends on the quality, number, and distribution of ground control points (GCPs). In this work, generalized additive models (GAMs) are used to analyze the relationship between both, the root mean square error (RMSEz) and the mean absolute error (MAEz) in the
$Z$
coordinate, in a group of checkpoints and in a set of covariates related to the number and spatial distribution of the GCPs. Beyond the exploratory data analysis frequently used to analyze the effect of GCPs on UAV-generated DEM accuracy, our approach also allows the determination of the shape of the association between the response (RMSEz and MAEz) and the predictor variables describing GCP number and distribution when they are all combined in a single model. Among the different predictor variables studied, the number of GCPs has, by far, the greatest influence on vertical accuracy. Other variables such as mean distance between control points (CPs) or distance of checkpoints from their nearest CP are statistically significant but contribute much less to explaining RMSEz and MAEz. The analysis was performed by constructing 4600 DEMs from different GCP combinations, divided into four sampling methods: random, stratified, spatially balanced, and contour (edge) distribution. As expected, the random method produced the poorest results, while the stratified and contour distribution methods produced the smallest range and dispersion of errors.