3DGeo Research Group GEODYNAMO4D
Tracing geographic dynamics on 4D Point Clouds
- Duration: 2019 - 2022
- Team: Bernhard Höfle, Lukas Winiwarter
- Contact: Prof. Bernhard Höfle
- Funding: Heidelberg University
Research news can be found in our GIScience News Blog.
Motivation
The Earth’s surface is subject to changes at a range of temporal and spatial scales. Using remote sensing methods, such as laser scanning and dense image matching, we can create three-dimensional point clouds of this surface. By analysing point clouds acquired at different points in time, we want to develop our understanding of the various geomorphic processes that contribute to this change. While methods to quantify change between two 3D point clouds have been presented, they often do not take advantage of knowledge about the data itself, about the method of its acquisition, and the monitored object. Use of this additional information is anticipated to enable an improved quantification of change itself, but also of the estimation of significance, effectively lowering the level at which change can be detected.
Objective
The GEODYNAMO4D research project develops methods for the analysis of time series of point cloud data, referred to as 4D point clouds. The goal therein is robust quantification of geometric change, tracing of the change direction, and the identification of areas where we cannot reliably determine change. Identifying areas with similar spatiotemporal change patterns will allow the separation of superimposed processes working in geometrically similar, yet distinguishable ways; for example, in separating creep of a rock glacier from rolling/sliding events of individual boulders.
The task will be approached on three different levels:
- Incorporating knowledge on the sensing process, i.e. measurement errors, using error propagation.
- Using the full time series to extract patterns that are otherwise attributed to noise by applying signal processing methods and machine learning.
- The observed objects and their changes are not arbitrary, but rather follow certain laws, that can be understood as system constraints. These constraints can e.g. be enforced in a neural network, increasing training speed and reducing network complexity.
Datasets
The developed methods should be very general and applicable to many tasks concerned with time series of topographic point cloud data. This includes all processes that have a temporal and a spatial component and specific characteristics in their change over time, for example geomorphological processes and vegetation dynamics (cf. Project SYSSIFOSS).
Virtual laser scanning (HELIOS++) will provide additional simulated measurements that can be used as datasets with exactly-known ground truth. Combined with simulation of geomorphological processes, this allows easy testing of the developed algorithms over multiple time scales and supplies practically unlimited training data for machine learning applications.
Related Projects
- HELIOS++: Heidelberg LiDAR Operations Simulator.
- Auto3DScapes: Autonomous 3D Earth observation.
- SYSSIFOSS: Synthetic structural remote sensing data for improved forest inventory models.
- Geomorph4D: Characterising multi-process geomorphic change.
- AHK-4D: High-resolution and high-frequency monitoring of the rock glacier Äußeres Hochebenkar (AHK) in Austria.