 |
| Titel |
Some safe and sensible shortcuts for efficiently upscaled updates of existing elevation models. |
| VerfasserIn |
Thomas Knudsen, Allan Aasbjerg Nielsen |
| Konferenz |
EGU General Assembly 2013
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 15 (2013) |
| Datensatznummer |
250083918
|
|
|
|
|
|
| Zusammenfassung |
| The Danish national elevation model, DK-DEM, was introduced in 2009 and is based on
LiDAR data collected in the time frame 2005–2007. Hence, DK-DEM is aging, and it is time
to consider how to integrate new data with the current model in a way that improves the
representation of new landscape features, while still preserving the overall (very high) quality
of the model.
In LiDAR terms, 2005 is equivalent to some time between the palaeolithic and the
neolithic. So evidently, when (and if) an update project is launched, we may expect some
notable improvements due to the technical and scientific developments from the last half
decade.
To estimate the magnitude of these potential improvements, and to devise efficient and
effective ways of integrating the new and old data, we currently carry out a number of case
studies based on comparisons between the current terrain model (with a ground sample
distance, GSD, of 1.6 m), and a number of new high resolution point clouds (10-70
points/m2).
Not knowing anything about the terms of a potential update project, we consider multiple
scenarios ranging from business as usual: A new model with the same GSD, but improved
precision, to aggressive upscaling: A new model with 4 times better GSD, i.e. a 16-fold
increase in the amount of data.
Especially in the latter case speeding up the gridding process is important.
Luckily recent results from one of our case studies reveal that for very high resolution
data in smooth terrain (which is the common case in Denmark), using local mean (LM) as
grid value estimator is only negligibly worse than using the theoretically “best” estimator, i.e.
ordinary kriging (OK) with rigorous modelling of the semivariogram. The bias in a leave one
out cross validation differs on the micrometer level, while the RMSE differs on the 0.1 mm
level.
This is fortunate, since a LM estimator can be implemented in plain stream mode, letting
the points from the unstructured point cloud (i.e. no TIN generation) stream through the
processor, individually contributing to the nearest grid posts in a memory mapped grid
file.
Algorithmically this is very efficient, but it would be even more efficient if we did not
have to handle so much data.
Another of our recent case studies focuses on this. The basic idea is to ignore data that
does not tell us anything new. We do this by looking at anomalies between the current
height model and the new point cloud, then computing a correction grid for the
current model. Points with insignificant anomalies are simply removed from the point
cloud, and the correction grid is computed using the remaining point anomalies
only.
Hence, we only compute updates in areas of significant change, speeding up the process,
and giving us new insight of the precision of the current model which in turn results in
improved metadata for both the current and the new model.
Currently we focus on simple approaches for creating a smooth update process for
integration of heterogeneous data sets. On the other hand, as years go by and multiple
generations of data become available, more advanced approaches will probably become
necessary (e.g. a multi campaign bundle adjustment, improving the oldest data using
cross-over adjustment with newer campaigns).
But to prepare for such approaches, it is important already now to organize and evaluate
the ancillary (GPS, INS) and engineering level data for the current data sets. This is essential
if future generations of DEM users should be able to benefit from future conceptions of
“some safe and sensible shortcuts for efficiently upscaled updates of existing elevation
models”. |
|
|
|
|
|
|