 |
| Titel |
Reproducible Earth observation analytics: challenges, ideas, and a study case on containerized land use change detection |
| VerfasserIn |
Marius Appel, Daniel Nüst, Edzer Pebesma |
| Konferenz |
EGU General Assembly 2017
|
| Medientyp |
Artikel
|
| Sprache |
en
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 19 (2017) |
| Datensatznummer |
250144669
|
| Publikation (Nr.) |
 EGU/EGU2017-8525.pdf |
|
|
|
|
|
| Zusammenfassung |
| Geoscientific analyses of Earth observation data typically involve a long path from data
acquisition to scientific results and conclusions. Before starting the actual processing, scenes
must be downloaded from the providers’ platforms and the computing infrastructure needs to
be prepared. The computing environment often requires specialized software, which in turn
might have lots of dependencies. The software is often highly customized and provided
without commercial support, which leads to rather ad-hoc systems and irreproducible results.
To let other scientists reproduce the analyses, the full workspace including data,
code, the computing environment, and documentation must be bundled and shared.
Technologies such as virtualization or containerization allow for the creation of
identical computing environments with relatively little effort. Challenges, however,
arise when the volume of the data is too large, when computations are done in a
cluster environment, or when complex software components such as databases are
used.
We discuss these challenges for the example of scalable Land use change detection on
Landsat imagery. We present a reproducible implementation that runs R and the
scalable data management and analytical system SciDB within a Docker container.
Thanks to an explicit container recipe (the Dockerfile), this enables the all-in-one
reproduction including the installation of software components, the ingestion of the data,
and the execution of the analysis in a well-defined environment. We furthermore
discuss possibilities how the implementation could be transferred to multi-container
environments in order to support reproducibility on large cluster environments. |
|
|
|
|
|
|