1 Spatial Modelling of the Terrestrial Environment - Georeferencial

128 Spatial Modelling of the Terrestrial Environment
Table 6.2 The response of data quality parameters to the exclusion of data points in their
derivation. Note that the changes in ME and SDE are statistically significant at the 95% level
when compared with the raw data case
Excluding points Excluding points Excluding points
Raw within 1 m of a within 2 m of a within 5 m of a
data wetted channel wetted channel wetted channel
ME (mean error), m 0.197 0.014 0.051 0.063
SDE (standard 0.313 0.195 0.193 0.258
deviation of error), m
Maximum error, m 0.828 0.231 0.231 0.231
Minimum error, m 0.265 −0.265 −0.265 −0.265
eliminated without recourse to error removal once the data has been generated. However,
this is an idealistic view. In this section we explore the causes of the errors identified in
section 6.4 and demonstrate that even if data re-collection is not possible, explaining the
cause of DEM error goes some way to justifying the correction methods adopted.
6.5.1 Causes of Localized Error
The methods described in Section 6.4.1 allowed the identification of erroneous pits and
spikes in the dataset. The first step in attempting to explain these errors involved visualizing
and quantifying their spatial distribution. Plate 4 shows a contour plot of error superimposed
upon a wet–dry classification of an orthorectified image. This immediately shows that the
errors are associated with proximity to wet areas, with some concentration of error along the
water edges. To assess the extent to which this pattern could be supported by quantitative
analysis, edge detection was applied to an orthorectified image of the area shown in Plate 4.
Data points in the raw photogrammetric DEM were then removed from the error analysis
according to their distance from the water’s edge.
Table 6.2 confirms the observation in Plate 4 that errors were significantly greater
along the channel margin. Exclusion of points within a1mradius results in the largest
reduction in ME and a major reduction in the SDE. With 2 m exclusion, similar patterns are
found. With5mexclusion, data quality starts to worsen, albeit marginally as compared with
the raw data case. It is also interesting to note that the minimum error is insensitive to point
exclusion as compared to maximum error. In this case, positive error is where check point
elevation is higher than photogrammetric point elevation, suggesting that photogrammetrically
acquired data that are close to the channel margin appear to be too low compared to
their correct elevation. As we are not considering interpolated data points here, only those
that have been matched successfully using the photogrammetric method, there are two hypotheses
for this effect: (1) structure of river bed relief; and (2) the stereo-matching process.
Large discrepancies between check data and photogrammetric data might be expected in the
vicinity of breaks of slope, where the river banks are steep. This relates to the resolution of
data collection; witha1mdata collection interval, elevations are averaged over a 1 m 2 area
due to the template based stereo-matching. However, the check data refer to point samples
that are not areally weighted in any way. This means that larger errors would be expected