incorrectly displaced into the class "deciduous forest"
(2.09%).
Table 2: Displacement errors(0.5 pixel to east and north)
Changes of coniferous forest
conif. ---> conif. 86.74 %
conif. ---» mixed | 11.17 96
conif. ---» decid. | 2.09 96
Changes of deciduous forest
decid. ---» conif. | 20.18 96
decid. ---> decid. | 69.07 96
decid. ---> mixed | 10.75 %
Changes of mixed forest
mixed ---» conif. | 8.63 %
mixed ---> decid. | 16.99 %
mixed ---> mixed | 74.38 %
By shifting the forest type classification by an assumed
image distortion of 1 pixel to east and 1 pixel to north the
following correspondence values were obtained:
Table 3: Displacement errors(1.0 pixel to east and north)
Changes of coniferous forest
conif ---> conif 83.55 %
conif ---> mixed 13.96 %
conif ---> decid 2.49 %
Changes of deciduous forest
decid ---> conif 25.98 %
decid ---> decid | 60.07 %
decid ---> mixed | 13.95 96
Changes of mixed forest
mixed ---> conif 10.00 %
mixed ---> decid | 22.48 %
mixed ---> mixed | 67.52 %
The test has proven that a pixelwise comparison of
signatures or classification results cannot be
recommended for forest monitoring, since changes in
forest areas occur in most cases in smaller dimensions as
the miss-registration found in the test in the alpine region.
This is particularly true for forest types that are
characterised by heterogeneous spatial distribution such
as alpine mixed stands. This can be demonstrated by the
correspondence values of the class "mixed forest’
wherein values of only 74.38% (0.5 pixel shift) and
67.52% (1.0 pixel shift) could be noticed.
Calculation of a change vector in a moving window
In order to overcome the above discussed geometrical
superposition errors as well as sampling restrictions in the
image data sets, larger reference units than one single
pixel have to be used for signature comparison. As a
solution it is proposed to apply a moving window (kernel)
on the geocoded, topographical normalised and
absolutely or relatively calibrated multi-temporal data
sets. The selection of the relevant bands will be the same
as used for the classification of the forest composition
(see chapter 4.4). Within this moving window the change
vector has to be calculated by the mean value of the
relevant channels.
Thresholding of change vector
To ensure that the signature changes described by the
change vector are due to real changes and not due to
calibration errors only changes which are characterised
by change vectors larger than the estimated calibration
error has to be considered. The threshold respectively the
necessary length of the change vector can be empirically
estimated using the root mean square error of regression
function of the calibrated images t1 and t2. The root mean
square errors calculated from regression functions of the
corresponding TM-bands are:
TM1: 1.3 digital numbers
TM2: 0.7 digital numbers
TM3: 1.1 digital numbers
TM4: 4.3 digital numbers
TM5: 3.0 digital numbers
TM7: 1.4 digital numbers
Thus, only change vectors that are larger than the
attained RMS-errors are assigned to the category "real
changes".
4.6 Interpretation of changes by using the
classification results
Since only the change vector magnitude will be used for
detection of the changed areas, the result of the step
described in chapter 4.5 are areas which show significant
changes in signatures between the acquisition dates. The
interpretation of these changes can now be carried out by
means of the classification results. The advantage of this
approach is that no complicate labelling or referencing of
the change vectors using ground truth data of different
acquisition dates is necessary.
The following change categories were defined:
conifer -» broad-leafed
conifer -> mixed
broad-leafed -> conifer
broad-leafed -» mixed
mixed -» conifer
mixed -» broad-leafed
other wooded land-» non-forest
non-forest -» other wooded land
changes between canopy closure classes
forest area -> non-forest
major species groups >- clearings
0:00 000 00 00:0
4.7 Visual Approval
For certain change categories very high accuracy
requirements have to be fulfilled to deliver useful results
270 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
for e
above
Suppx
classi
into «
such
opere
applic
const
limite
refore
symp
From
monit
an e!
tasks.
remo!
objec
chan
it off
advar
recort
to ot
monit
dealir
Howe
proce
Extrei
impor
monit
theref
accur
are s
image
be af
geocc
Effort
under
For
multis
radior
atmos
proble
methc
restric
remot
compl
appro
differe
proble
range
in on
constr
consic
of mul
order
