Automatic Classification of Surface Activity Types from Geographic 4D Monitoring Combining Virtual Laser Scanning, Change Analysis and Machine Learning

We advance the characterization of landscape dynamics through analysis of point cloud time series by integrating virtual laser scanning, machine learning and innovative open source methods for 4D change analysis. We present a novel approach for automatic identification of different surface activity types in real-world 4D geospatial data using a machine learning model trained exclusively on simulated data.

Our method focuses on classifying surface activity types based on spatiotemporal features. We generate training data using virtual laser scanning of a dynamic coastal scene with artificially induced surface changes. Scenes with surface change are generated using geographic knowledge and the concept of 4D objects-by-change (4D-OBCs) [1, 2], which represent spatiotemporal subsets of the scene that exhibit change with similar properties. A realistic 3D scene modelling is essential for accurately replicating the dynamic nature of coastal landscapes, where morphological changes are driven by both natural processes and anthropogenic activities.

The Earth's landscapes exhibit complex dynamics, spanning large spatiotemporal scales, from high-mountain glaciers to sandy coastlines. The challenge lies in effectively detecting and classifying diverse surface activities with varying magnitudes, spatial extents, velocities, and return frequencies. Effective characterization of these dynamics is crucial for understanding the underlying environmental processes and their interplay with human activities. Supervised machine learning classification of surface activities from point cloud time series is challenging due to the limited availability of comprehensive and diverse real-world datasets for training and validation. Our approach combines virtual laser scanning with machine learning-based classification, enabling the generation of comprehensive training datasets covering the full spectrum of expected change patterns [3].

In our approach, the simulation of LiDAR point clouds is performed in the open-source framework HELIOS++ [4, 5]. HELIOS++ allows the flexible simulation of custom LiDAR campaigns with diverse acquisition modes and settings together with automatic annotations of artificially induced surface changes. We train a supervised machine learning model to classify synthetic 4D-OBCs into typical surface activity types of a sandy beach (e.g. dune erosion/accretion, sediment transport, etc.). Moreover, we investigate descriptors for 4D-OBCs, assessing their suitability for representing general types of surface activity (transferable between use cases) and types specific to particular surface processes.

We evaluate our model for 4D-OBC classification in terms of its capacity to discriminate surface activity types in a real-world dataset of a sandy beach in the Netherlands [6]. 4D-OBCs are extracted, classified into our target classes and validated with manually labelled reference data based on expert evaluation.

Our study showcases the efficacy of coupling virtual laser scanning, innovative open-source 4D change analysis methods, and machine learning for classifying natural surface changes [7]. Our findings not only contribute to advancing the understanding of landscape dynamics but also provide a promising approach to mitigating environmental challenges.

REFERENCES

[1] Anders et al. (2022): DOI: https://doi.org/10.5194/egusphere-egu22-4225

[2] py4dgeo: https://github.com/3dgeo-heidelberg/py4dgeo 

[3] Zahs et al. (2022): DOI: https://doi.org/10.1016/j.jag.2023.103406

[4] HELIOS++: https://github.com/3dgeo-heidelberg/helios

[5] Winiwarter et al. (2022): DOI: https://doi.org/10.1016/j.rse.2021.112772 

[6] Vos et al. (2022): DOI: https://doi.org/10.1038/s41597-022-01291-9

[7] CharAct4D: www.uni-heidelberg.de/charact4d