Fully automatic spatiotemporal segmentation of 3D LiDAR time series for the extraction of natural surface changes
Graphical abstract
Introduction
Time series of topographic 3D data pose great possibilities to geographic analyses, but also challenges to the detection and delineation of surface change information from these 4D geospatial data. Surface change analysis using topographic point clouds has long since gained considerable importance in the observation of Earth surface processes and in advancing geoscientific research in general (Eitel et al., 2016, Qin et al., 2016). Ongoing repetitions of topographic surveys have led to the cumulation of data in the temporal dimension that enable change detection for many use cases of Earth surface observation, among them studies of changes to landslides (Oppikofer et al., 2009, Pfeiffer et al., 2018), rockfalls (Abellán et al., 2010, Rosser et al., 2007), rock glaciers (Bodin et al., 2018, Zahs et al., 2019), snow cover (Grünewald et al., 2010, Fey et al., 2019), and the coast (Fabbri et al., 2017, Miles et al., 2019).
Most recently, the acquisition strategy of permanent terrestrial laser scanning (TLS) generates time series of 3D point clouds at (sub-)hourly intervals over periods of weeks to months (e.g., Kromer et al., 2017, O'Dea et al., 2019, Stumvoll et al., 2020, Vos et al., 2017, Williams et al., 2018). Such high-frequency time series of point clouds can alternatively be acquired by photogrammetric techniques (e.g., Eltner et al., 2017, Kromer et al., 2019), albeit obtained datasets have different properties from laser scanning point clouds and may require other specific processing strategies. The unprecedented temporal density of these 4D datasets provides many more epochs for change analysis to compare evermore combinations of pairwise states of the topography, which are ideally adapted to the rates of target changes, respectively. Even more importantly, the information these data potentially contain on temporal properties of change processes hold opportunity for new insights on spatiotemporal characteristics of topographic activity and, consequently, to extend our fundamental knowledge about the investigated geographic phenomena (Eitel et al., 2016, Eltner et al., 2017).
To leverage the temporal dimension of 3D time series for change detection and delineation, a method of spatiotemporal segmentation was developed that makes use of the full history of surface change to extract periods and spatial extents of surface changes (Anders et al., 2020b). The time series-based approach is designed to advance established approaches of pairwise change detection. Pairwise change analysis typically serves to identify areas of accumulation or erosion over the selected analysis period and to quantify change rates. Based on the bitemporal change information, patterns and underlying drivers of change are interpreted (e.g., Anders et al., 2020a, Eltner et al., 2017, Fey et al., 2019, Zahs et al., 2019). Standard methods for pairwise change detection are the differencing of Digital Elevation Models (DEMs, James et al., 2012) or point cloud distance computation (Girardeau-Montaut et al., 2005, Lague et al., 2013). Alternatively, change can be assessed in object-based approaches, where observed objects are first identified based on morphometric features or even previously derived bitemporal surface change, and subsequently changes in object properties are analysed, such as their location and size (e.g., Mayr et al., 2018).
In the following, we reveal drawbacks of pairwise change detection methods for the analysis of 3D time series. These drawbacks arise from the general circumstance of observing natural, Earth shaping processes that it is not a priori known when and where changes occur within a scene, and what the spatial and temporal properties of these changes are. The required selection of epochs in pairwise change analysis entails that the periods for detecting change are pre-defined. Temporary surface changes, which only persist for a limited amount of time within the observed scene, may hence be missed in the analysis (Anders et al., 2019) as their timing and/or existence are not known to the analyst and their disappearance is not expressed in later topographic information (Fig. 1A). Performing pairwise change analysis for all combinations of epochs to solve this drawback would be somewhat impractical, and has not been done so far to our knowledge. Pairwise change analysis, therefore, is not adequate to analyse 3D time series for changes that occur with highly varying temporal characteristics, that is timing, change rate, duration of change processes, and persistence of change forms. Surface changes further occur at varying spatial scales regarding their extent, shape and magnitude and can therefore not be extracted generically with one-for-all settings. Where morphologic boundaries of objects or forms moreover are not distinct, it is difficult to spatially delineate them in individual scenes. Binary surface change information (change/no change) has been used to identify and delineate so-called change objects (Liu et al., 2010). However, these spatially contiguous areas of surface change do not necessarily stem from the same change-inducing process (Fig. 1B). Separating them into individual objects without a priori knowledge or information on external influences is improved when integrating the history of surface change that is contained in the 3D time series in the analysis (Anders et al., 2020b).
Time series clustering (Kuschnerus et al., 2020) has been proposed for extracting change information from large 4D geospatial data. The method is useful to extract areas that are homogenous in their change dynamics and thereby finding dominant change patterns in the observed scene. It is not possible, though, to identify individual, spatially and temporally limited change occurrences as the full time series of the observation period are used as input.
Object extraction by integrating the history of surface change can be performed with the concept of 4D objects-by-change (4D-OBCs; Anders et al., 2020b). This method identifies areas in the scene where the surface changes similarly over time within sub-periods in the time series at neighbouring locations. Sub-periods are automatically detected in the temporal domain of a location and are subsequently used as seeds for spatial region growing with time series similarity as homogeneity criterion. The seed locations at which to perform the temporal change detection have been selected manually so far. However, changes occurring are spatially variable within a scene and their timing and location is in general not known to the analyst. Automatic extraction of changes from entire datasets will therefore require fully automatic seed selection from all seed candidates. These can be obtained as sub-periods via temporal change detection at all locations in the scene. Many of these detected sub-periods will be both spatially neighbouring and temporally overlapping, and thereby likely belong to the same change form. One could perform the segmentation for all detected sub-periods at all locations and subsequently aggregate segments into unique 4D-OBCs in a post-processing step. This option is hardly viable, given the large data volumes of 3D time series and considering the extreme redundancy of computations if each location within a change form is used as seed to obtain the same, single object.
The drawbacks outlined above become particularly apparent in settings with continuous surface morphology and dynamic changes of the surface due to material transport induced by varying external drivers. Therefore, the use case of this paper is TLS-based monitoring of a sandy beach, using an hourly time series spanning five months. Sandy beaches are highly active in their morphodynamics through multiple processes acting on the surface, as these coastal landscapes are subject to continual sediment transport by wind, waves, as well as anthropogenic modifications. Their surface is hence shaped by a variety of (temporary) forms of accumulation, erosion, and transported material (Walker et al., 2017). Therefore, the target changes to be extracted from our data are temporary accumulation and erosion forms which typically exist over periods of days to few weeks.
The objective of this paper is to develop a fully automatic workflow to extract surface changes as temporary accumulation or erosion forms in their varying spatial and temporal extents from a long and dense 3D time series dataset. To achieve this, we implement methods of automatic seed selection and locally adaptive thresholding for spatiotemporal segmentation. We consider the following methodological aspects:
- •
Integrating the history of surface change in temporal change detection will avoid missing temporary surface changes in the analysis which may not persist throughout epochs that are selected for fixed-period analyses.
- •
Sorting and selecting seeds for region growing by an appropriate metric and considering previous segments throughout continued segmentation allows avoiding redundant calculations but also prevents skipping relevant change occurrences.
- •
A locally adaptive approach of threshold selection is more suitable than pre-defined thresholds, albeit strict or loose, in order to avoid general over- or underestimation of spatial extents due to dependence on magnitude, duration, and change rate of the respective change form.
The developed automatic spatiotemporal segmentation approach is designed to improve both the spatial separation of change forms, e.g. two co-occurring 4D-OBCs that would otherwise be extracted as one accumulation or erosion object, and the temporal separation of change forms, e.g., two consecutive 4D-OBCs that could be aggregated or missed with alternative methods (cf. Fig. 1). Our approach thereby provides an important contribution to the toolset for change analysis in large 4D geospatial data.
Section snippets
Data and methods
We use a time series of hourly TLS data with 2,942 epochs acquired at a sandy beach to perform a fully automatic extraction of 4D-OBCs. The results are validated by a group of expert analysts, who assess the detection and extraction performance of our method for sample locations. In addition, we compare extracted 4D-OBCs with changes derived from the baseline method of pairwise change analysis. The main steps of our approach and the investigation of results are visualised in Fig. 2, with
Results
In the spatial–temporal extent of the dataset, a total of 306,728 surface changes are detected as sub-periods at 192,901 locations in the scene, which derive from over 15 billion LiDAR points in the full 3D time series. The detected sub-periods are sorted by decreasing neighbourhood similarity and therein decreasing change volume, and provide the seed candidates for the segmentation (Section 2.2). A total of 7,893 segments are generated by the full segmentation until the end of the seed
Discussion
The ability to identify and separate individual change forms spatially and temporally with time series-based change analysis is an important improvement regarding drawbacks of standard pairwise change analysis approaches (cf. Section 1). The change analysis becomes independent of selecting analysis periods and also of selecting a reference epoch that most suitably represents the initial state of the surface, relative to which accumulation and erosion are determined. Particularly in natural
Conclusion
In this paper, we present a fully automatic approach to change analysis from 3D time series data. The method detects changes in the time series at locations in a scene and makes use of spatiotemporal segmentation to delineate change forms. This enables the extraction of 4D objects-by-change (4D-OBCs), i.e. temporary surface changes induced by material transport on continuous surface morphology that are difficult to detect in space and time using single topographic snapshots or fixed-period
Data statement
The Python scripts to perform the spatiotemporal segmentation are published openly together with material and results of the validation in the data repository of Heidelberg University (https://doi.org/10.11588/data/4HJHAA). The 3D time series data used in this article are available upon reasonable request to S.E. Vos ([email protected]).
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
We are greatly thankful to six colleagues who took the time to perform the evaluation of results for validation of the method. This work was supported in part by the Heidelberg Graduate School of Mathematical and Computational Methods for the Sciences (HGS MathComp), founded by DFG grant GSC 220 in the German Universities Excellence Initiative. The acquisition of the TLS time series data was financed by the ERC Advanced Grant Neashore Monitoring and Modeling (grant number 291206) and by the
References (40)
- et al.
Detection and spatial prediction of rockfalls by means of terrestrial laser scanner monitoring
Geomorphology
(2010) - et al.
4D objects-by-change: Spatiotemporal segmentation of geomorphic surface change from LiDAR time series
ISPRS J. Photogramm. Remote Sens.
(2020) - et al.
Mapping burned areas from Landsat TM/ETM+ data with a two-phase algorithm: Balancing omission and commission errors
Remote Sens. Environ.
(2011) - et al.
Beyond 3-D: The new spectrum of lidar applications for earth and ecological sciences
Remote Sens. Environ.
(2016) - et al.
Geomorphological analysis and classification of foredune ridges based on Terrestrial Laser Scanning (TLS) technology
Geomorphology
(2017) - et al.
Cloud detection algorithm comparison and validation for operational Landsat data products
Remote Sens. Environ.
(2017) - et al.
Mapping forest change using stacked generalization: An ensemble approach
Remote Sens. Environ.
(2018) - et al.
3D micro-mapping: Towards assessing the quality of crowdsourcing to support 3D point cloud analysis
ISPRS J. Photogramm. Remote Sens.
(2018) - et al.
Geomorphic change detection using historic maps and DEM differencing: The temporal dimension of geospatial analysis
Geomorphology
(2012) - et al.
Accurate 3D comparison of complex topography with terrestrial laser scanner: Application to the Rangitikei canyon (N-Z)
ISPRS J. Photogramm. Remote Sens.
(2013)