FEATURE RELEVANCE ANALYSIS FOR 3D POINT CLOUD CLASSIFICATION USING DEEP LEARNING

We present a method to extract surface change information from 4D time series of topographic point clouds recorded with a terrestrial laser scanner. The method uses sensor information to spatially and temporally smooth the data, reducing uncertainties. The Kalman filter used for the temporal smoothing also allows us to interpolate over data gaps or extrapolate into the future. Clustering areas where change histories are similar allows us to identify processes that may have the same causes.

The rock glacier in Äußeres Hochebenkar (Austria) moved faster in 2021–2022 than it has in about 70 years of monitoring. It is currently destabilizing. Using a combination of different data types and methods, we show that there have been two cycles of destabilization at Hochebenkar and provide a detailed analysis of velocity and surface changes. Because our time series are very long and show repeated destabilization, this helps us better understand the processes of rock glacier destabilization.