Processing TLS heterogeneous data by applying robust Msplit estimation

Terrestrial laser scanning provides a point cloud containing hundreds or thousands of points. One should suppose that some points are mismeasured; hence, the observation set is not homogeneous, requiring the application of modern statistical methods, such as M-split estimation. That novel method is designed for processing observation sets that are unrecognized mixtures of realizations of at least two random variables (however, a priori, there is no information on subset division). The basic M-split estimates are not robust against outlying observations. The paper proposes new variants of M-split estimation designed as robust against outliers. The example applications prove that the new variants are twice as accurate as of the existing M-split estimates or four times more accurate than the conventional robust assessments. What is more, new variants can provide acceptable results even if the share of outliers exceeds 50%, which is impossible for traditional robust methods.

Automated summary

Terrestrial laser scanning provides a point cloud with hundreds or thousands of points, some of which may be mismeasured. Modern statistical methods, such as M-split estimation, are required to process the non-homogeneous observation set. This paper proposes new robust variants of M-split estimation that are twice as accurate as existing methods and can tolerate up to 50% outliers, exceeding the capabilities of traditional robust methods.