International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
23 E s. E
Figure 6. Change detection using two images within a time
lapse of 10 years. Bright details are subject of change
If we assume that two points taken into consideration are
independent and their specific accuracy is different, the relative
position error can be calculated from the square sum of the two
values for the positional accuracy:
4.2 Classification data error
Another source of error is the thematic classification of data.
A method to empirically assess the classification accuracy is to
select several classes and to compare them with the reference
data. Reference data is usually named “ground truth”. By
comparing the data sets, the percentage of the pixels correctly
classified can be estimated.
From every class representative pixels are selected and
compared to the reference data.
A statistical approach of this problem is to select random pixels
from the thematic map and to compare them to the reference
data. Here, the main impediment is that large classes have the
tendency to be represented by a larger number of points and the
small classes may be not represented at all. The solution to this
problem could be the stratified random sampling, in this case, a
set of strata are predefined and the random sampling is carried
out in each of these collections. A regular grid can be used or a
random selection of pixels in each class, in order to assess the
class accuracy. A confusion matrix will result.
(Van Genderen, 1982) and (Rosenfield,1978) have, along with
others, determined guidelines for the minimum sample size.
The estimated accuracy for a class can be calculated (Freund,
1962):
Xx — nà
= eds uf
g né 9) Wa e
x= number of correct identified pixels, n= total number of pixels
in the sample, 6= the map accuracy, (1-a)= a confidence limit,
imposed by the analyst.
The remote sensing characteristics that affect the change
assessment accuracy are: temporal, spectral and spatial
resolution, look angle.
In order to perform accurate change analysis, the data must be
acquired at approximate the same time of the day and at
significant calendar dates regarding the environmental changes
that are under observation.
If data used to detect changes are from the sensors with the
same IFOV (Instantaneous Field of View), it is easy to register
the two data sets. Geometric rectification algorithms can be
used to register the images to a standard map projection (most
of the available software and maps are in UTM but for the
specific case of Romania, a Stereographic 1970 map projection
is necessary- for this, standard datum and standard geoid is
provided).
(2)
782
4.3 Environmental considerations
To obtain robust change detection, some environmental factors
and variables must be taken into consideration, such as
atmospheric conditions, soil characteristics, vegetation cycles,
hydrologic cycles and others. Most of the environmental
features are extremely dynamic, in most of the cases the
temporal resolution of remote sensed data cannot cover the
dynamic domain of the environmental parameters evolution
(atmospheric conditions, soil moisture, other environmental
related phenomena). The atmospheric effect can be corrected
with specific path radiance atmospheric correction models or an
image-to-image normalization method.
Many factors, related to the specific phonological characteristics
of the vegetation canopy must be taken into consideration.
Attention must be given to differences in the phenological state
of different varieties of the same species and the time the data
sets were acquired . Meteorological aspects and the hydrologic
regime of the area along with the agricultural work schedule are
important aspects when change detection analysis is performed.
Depending of the meteorological conditions, the river network
of the studied area can suffer changes and thus affect the soil
humidity conditions. These aspects are predictible if we have
appropriate geomorphologic analysis is and soil quality is
assessed for the zone in study.
5. Image processing and change detection
In order to obtain environmental changes information, once we
selected the appropriate data and classification scheme, special
radiometric and geometric corrections must be applied,
followed by change detection and classification techniques,
creation of thematic products and finally the error assessment.
Image normalization reduces the pixel brightness variations.
Using simple regression equations between the brightness
values of radiometric normalization targets in the base scene
and the scene to be normalized can perform image
normalization.
Ground targets that spectrally invariant in the two images can be
used to normalize multitemporale data sets to a single reference
scene. The acceptance criteria for radiometric normalization are
(Eckhardt, 1990):
The target must be at the same elevation, must contain as little
vegetation as possible the terrain must be as flat as possible, the
scene features must remain unchanged in both scenes.
This method calculates the additive term (path radiance
correction) from a constant (D) and then obtains the
multiplicative factor:
1
EA Lh 2. Aref
ref
COS Op ref
M =
1 (3)
COS O0norm| —z— Anorm
norm
C - Dre — (Dnorm)- (M)
where: 1/A= Radiance interval of brightness value, C= additive
correction, 057 solar zenith angle, ES= Earth-Sun distance,
Ref= reference scene, Norm= scene to be normalized, D= dark
normalization target of brightness value. This approach ignores
differences in atmospheric attenuation and phase angle between
data sets. The radiation received is dependent of the relative
orientation of the terrain from the Sun. All these methods
require a DEM (Digital Terrain Model).The DEM and the
image was registered and resample to the same spatial
resolution as images. And then the value for each pixel is
Inter
proc
shot
The
proj
char
dire
regi:
conf
conl
grou
to ci
are |
on tl
The
- Ch
This
in a
chan
- Mi
sets
can
class
a cl
Com)
- Im
Diff:
perf(
and 1
Whe
time
trans
(nort
i= lir
The.
boun
5.1.
Area
vectc
from
chan,
wher
band
$2.
In c
proce
colun
rows
This
categ
The «
chan;
of th
stasti
KAP]
differ
The K
