International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
I. INTRODUCTION
It is widely acknowledged that classification of remotely sensed
imagery has variable and often poor quality. The cause and
nature of these errors has been the subject of extensive research
in order to improve the accuracy of remotely sensed products.
Error in this context can be defined as some discrepancy
between the situations depicted on the generated image (map)
and reality (Arbia et al., 1998). Performing spatial data analysis
operations on data of unknown accuracy will result in a product
with low reliability and restricted use in the decision-making
process, while errors deriving from one source can propagate
through the database via derived products (Lunetta et al., 1991).
The quality of data is a function both of the inherent properties
of those data and the use to which they are to be put. Hence,
knowledge of error levels is necessary if data quality is to be
estimated.
There are two different components of accuracy in the context
of remote sensing: positional and thematic accuracy (Janssen
and Van der Wel, 1994). Positional accuracy determines how
closely the position of discrete objects shown on a rectified
image (map) or in a spatial database agree with the true position
on the ground, while thematic accuracy refers to the non-
positional characteristic of a spatial data entity, the so-called
attributes (which are derived from radiometric information).
Quality control of cartographic products is usually
accomplished by computing the discrepancy between cach
member of a set of well defined points present in one
cartographic document with the corresponding points observed
in the field, using a technique that guarantees sufficient
accuracy for the analysis. In spatial databases generated from
remotely sensed data, it is equally necessary to have knowledge
of the discrepancies (errors). However, in some imagery where
the number of control points is not sufficient, or where their
spatial distribution is suboptimal, the use of generic features
(such as roads, edges, polygons, ctc) to provide a means of
relating two spatial data sets is an important alternative.
The aim of this paper is to review standard methods for
assessing the quality of cartographic products in the context of
remote sensing. A further aim is to present a methodology to
assess the positional accuracy of spatial databases using generic
features (and their spatial distribution) within the image in the
validation phase, and also present the user with an indication of
the thematc reliability of the remote sensing products.
2. STUDY AREA AND DATA
Two separatet datasets were used in order to assess the
positional and thematc accuracy of remote sensing products.
2.1 Positional Accuracy
The study area is located near the town of Uberaba-MG, in
southeastern Brazil. This area is located at approximately 700m
above sea level, and possesses an undulating topography. The
economic activities of the region are based on dairy farming, as
most of the area is covered by grassland.
Two remotely sensed multi-band images were used in this
study, one acquired by the Landsat Thematic Mapper (at 30m
resolution) and the other by the High Resolution CCD Camera
(HRCCC) carried by the China-Brazil Earth Resources Satellite
(CBERS). This camera has a spatial resolution of 20 m. For the
purposes of this study, a single waveband of each of the two
multi-band images (TM and HRCCC) were used. A 1:25,000
scale map of the study region was used to provide ground
reference data.
980
The two single-band images were geometrically registered to
the UTM reference system (zone -23 S) using the Córrego
Alegre horizontal datum (Brazil). Image to map registration
used 14 and 12 ground control points respectively for the
Landsat TM and CBERS HRCCC images, with nearest
neighbour resampling, since this technique maintains the
original pixel values (Jensen, 1986). In each case, the root-
mean-square (RMS) error associated with registration was less
than 0.5 pixels (i.e., the RMS for Landsat TM was 0.4721 pixel
and the RMS error for the CBERS image was 0.479 pixel).
Atmospheric correction was not performed since comparisons
are not being made directly between images.
2.2 Thematic Accuracy
A SPOT High Resolution Visible (HRV) multispectral (XS)
image (14 June 1994) of a region of flat agricultural land
located near the village of Littleport (E. England) is used in this
study, together with Field Data Printouts for summer 1994,
These printouts are derived from survey data supplied by
individual farms, and provide details of the crop or crops
growing in each field in the study area. On the basis of
examination of the areas covered by each crop, the geographical
scale of the study, and the spectral separability of the crops,
seven crop categories were selected: potatoes, sugar beet,
wheat, fallow, onions, peas and bulbs.
Image processing operations were performed using ERDAS
[Imagine (version 8.0) and the IDRISI GIS. Neural network
application used the SNNS software. Some in-house programs
were written to carry out specific procedures. Registration of
the image to the Ordnance Survey (GB) 1:25,000 map was
performed using 17 ground control points and nearest neighbour
interpolation. The RMS error was 0.462 pixels.
3. TECHNIQUES FOR ESTIMATING POSITIONAL
AND THEMATIC ACCURACY
3.1 Positional Accuracy
A standard method of assessing the positional accuracy of
cartographic products is based on comparison of deviations
between homologous control points that can be accurately
located on both the reference map and the geometrically
corrected image. The deviations at these homologous points are
used to compute statistics that are used to perform specific tests
to evaluate the accuracy of the geometric corrected image.
An alternative approach to assess the positional accuracy of
cartographic products is based on the use of geometric features
(Galo et al., 2001). Examples of geometric features are roads,
edges, and other boundaries. They should be easily located and
represented as a set of sequential coordinates in both documents
(i.e., the image to be corrected and the reference map). Three
feature-based methods were used by Vieira et al. (2002) to
measure the correspondence between features shown on à
reference map and a remotely sensed image. These are the
Generated Point Method, the Areal Method and the Equivalent
Rectangle Method. As these alternatives methods work with the
relative distances between homologous points, there is no need
to apply trend analysis to check the presence of systematic
errors on the directions E and N (E and N are the directions X
and Y respectively on the Universal Transverse Mercator
coordinate system). One of these methods is presented in the
following sections.
Internatio:
Le
3l: V
An impor
maps is th
2001), wh
homologo
(Xm: Yn)
homologo
compute :
evaluate |
corrected
reference
assess the
According
check for
sample me
applied, u
estimating
hypothesis
conclusior
error in the
The Stude
test. The
tables (wh
confidence
for the de
than f lc
east-west
then the ge
errors on
values of
equations:
where o; a
À; and An
Accuracy .
deviations
defined (t
hypothesis
for each o!
employed
geometrica
the directi
involving |
the tabled
The value
following «
Where 6x
function
formulae: (
Error (SE
Instance, i
classifies ¢
quality (sec
