PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2014
119
What is the most appropriate methodology
to calculate swath relative accuracy?
Which version of the data is most appro-
priate to perform the assessment: the
boresighted and calibrated flight lines
themselves (pre-classification) or using
the classified tiled data? Does one have
advantages/disadvantages over the other?
Are they essentially the same? How would
the data sampling be approached? Should
an RMSE be calculated for each indepen-
dent swath instead of an AOI as a whole?
Bobby Riley, Project GIS Specialist
URS Corporation, Germantown, Maryland USA
Dr. Abdullah:
In all the published lidar data specifications,
there is no clear strategy given concerning the testing
methodology of the relative accuracy for a lidar dataset. The
closest incident where the relative accuracy is mentioned is
in the USGS Lidar Base Specification v.1. According to these
specifications, the relative accuracy for lidar data is given as
follows:
• Within individual swaths:<= 7 cm RMSE
• Within overlap between adjacent swaths:<=10 cm RMSEz
It also calls on data providers to document the procedure
they applied without limiting them to a certain procedure as
it is stated below:
“Swath-to-swath and within swath accuracies (relative) are
to be documented. A detailed report of this validation process
is a required deliverable.”
Such a void in the current standards and specification
forced data providers to adopt different techniques for the
verification of the relative accuracy to ensure the quality
of their data is meeting certain specifications. Evaluating
the relative accuracy within a swath is the easiest task and
usually not formally performed by most data providers. The
relative accuracy of the data within a swath is a measure
of how much the elevation of a point in a point cloud differs
from the elevation of an adjacent point (of the same class)
when both points are reflected from a surface that has the
same elevation. It resembles low-level noise that affects the
smoothness of the data and therefore can be evaluated by
examining the standard deviation of the elevations from a
sample. Testing the relative accuracy of the data within a
swath is best achieved by evaluating a smooth, homogenous,
flat and level surface such as a large building roof, airport
runway or leveled segment of highway. For this purpose,
industrial buildings, highways, airport runways, airport
tarmacs or any flat reflective surfaces that are free from
any slope should be used for testing. Larger, flat-topped
commercial buildings that have a minimal asphalt water
barrier would be the best candidates. Avoid roads as much
as you can, as even minimally crowned roads can expand the
standard deviation and exhibit erroneous conclusions. As for
the number of samples, I suggest the following strategy:
Samples from building roofs:
Number of samples: at least five
Length: at least 10 × NPS (where, NPS is Nominal Post
Spacing)
Width: at least 5 × NPS
Therefore, for a lidar dataset with NPS of 1.4 m, you will test
a minimum of five samples each around 100 m
2
.
“In all the published lidar
data specifications, there
is no clear strategy given
concerning the testing
methodology of the
relative accuracy for a
lidar dataset.”
