1 Spatial Modelling of the Terrestrial Environment - Georeferencial

120 Spatial Modelling of the Terrestrial Environment
Table 6.1 Check data, photogrammetrically acquired points for elevations that were not inundated
and the associated error statistics
Number of Number of Check points Mean Standard
check points photogrammetrically as a percentage error deviation
DEM available acquired data points of dry points (m) of error (m)
February 3700 2 451 166 0.15 0.542 2.113
1999
March 241 2 232 187 0.01 0.252 0.883
1999
February 1661 2 288 138 0.02 0.064 0.926
2000
bed cannot be seen either because it is vegetated, or because it is inundated with water
which, in the case of the Waimakariri, is commonly too turbid to allow through water
photogrammetry (e.g. Westaway et al., 2000). However, even after erroneous data points
are removed (Figure 6.2(b)), it is clear that many errors remain. To quantify the error in
this surface, it is necessary to consider data points that have been stereo-matched and not
those where matching has failed and which have been interpolated. Interpolated data points
may be affected by incorrectly matched points. This will violate the use of equation (2) as
it is possible that adjacent data points are not independent when surface derivatives such
as slope are calculated.
Whilst the percentage of data points that could be checked is exceptionally low due to
the relatively small size of the check dataset (Table 6.1), the magnitude and frequency of
DEM errors are sufficient to give exceptionally poor data quality parameters (Table 6.1)
given the total topographic relief of about 2 m. There is systematic error in the surface,
associated with MEs significantly greater than zero, and poor surface point precision with
very high-standard deviations of error. However, there are clearly areas of the riverbed
which appear to be of a very high quality in visual terms and even with the errors present,
there is considerable spatial structure present in the DEM. This aside, three issues emerge.
First, many of the errors appear to be spatially concentrated along the main river channels.
Similarly, there appears to be some sort of banding present in the default DEM, which may be
due to some form of inaccuracy in the functional model (e.g. specification of the collinearity
equations). Second, if these DEMs were used to calculate a mean bed level, there would
be a systematic error of between 0.064 m (February 2000) and 0.542 m (February 1999).
These are significantly greater than those that would arise from measurement using the
conventional survey methods applied in the Waimakariri (levelling), except where bias
is introduced due to the relatively large spacing between levelled cross-sections. In this
study, these errors are too great to allow meaningful estimation of mean bed levels for river
management purposes. Third, the number of data points used to represent the surface makes
manual checking of each data point using stereo-vision unfeasible; a trained operator may
check up to 500 points per hour, although this checking rate cannot always be sustained
(Lane, 1998). This requires approximately 4900 hours to check each DEM. Similarly, the
number of check points available for assessing each data point is a very low percentage
of the number of data points generated photogrammetrically. Field data collection had to
involve check data stratification to areas close to and along the channel edge, where it