 |
| Titel |
Modelling spatiotemporal change using multidimensional arrays Meng |
| VerfasserIn |
Meng Lu, Marius Appel, Edzer Pebesma |
| Konferenz |
EGU General Assembly 2017
|
| Medientyp |
Artikel
|
| Sprache |
en
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 19 (2017) |
| Datensatznummer |
250152733
|
| Publikation (Nr.) |
 EGU/EGU2017-17610.pdf |
|
|
|
|
|
| Zusammenfassung |
| The large variety of remote sensors, model simulations, and in-situ records provide great
opportunities to model environmental change. The massive amount of high-dimensional data
calls for methods to integrate data from various sources and to analyse spatiotemporal and
thematic information jointly. An array is a collection of elements ordered and indexed in
arbitrary dimensions, which naturally represent spatiotemporal phenomena that are identified
by their geographic locations and recording time. In addition, array regridding (e.g.,
resampling, down-/up-scaling), dimension reduction, and spatiotemporal statistical
algorithms are readily applicable to arrays. However, the role of arrays in big geoscientific
data analysis has not been systematically studied: How can arrays discretise continuous
spatiotemporal phenomena? How can arrays facilitate the extraction of multidimensional
information? How can arrays provide a clean, scalable and reproducible change modelling
process that is communicable between mathematicians, computer scientist, Earth system
scientist and stakeholders?
This study emphasises on detecting spatiotemporal change using satellite image time
series. Current change detection methods using satellite image time series commonly analyse
data in separate steps: 1) forming a vegetation index, 2) conducting time series analysis on
each pixel, and 3) post-processing and mapping time series analysis results, which does not
consider spatiotemporal correlations and ignores much of the spectral information.
Multidimensional information can be better extracted by jointly considering spatial, spectral,
and temporal information. To approach this goal, we use principal component analysis to
extract multispectral information and spatial autoregressive models to account for
spatial correlation in residual based time series structural change modelling. We also
discuss the potential of multivariate non-parametric time series structural change
methods, hierarchical modelling, and extreme event detection methods to model
spatiotemporal change. We show how array operations can facilitate expressing these
methods, and how the open-source array data management and analytics software
SciDB and R can be used to scale the process and make it easily reproducible. |
|
|
|
|
|
|