International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
half the total pixel amount. As can be seen by using this
technique, the probability for representative data was increased
and the curves could be flattened. Figure 3 shows two cover
types before and after the setting of the limit: the smoothing
effect can be seen well in the sraphs of coniferous forest
plantation (a) and dry heather moorland (b).
07
n5 4 v s Bm
| ie m | h gH
= ; NA y d
as c ^l à | à T Wt m
A d 5 | | a
s 04 4 a qi à d i f y | |
m" | m
= | "P? 5 m 1] B or
| |
02
0,1 4
0 T T T = : + T
0 5 10 15 20 25 30 35 40 45 50
week
| —5— original data a-— reduced data :
(a)
07
05 4
05 + m * (1 > 94.
Ju up j fi f ks" a
e e re
$553 T AP / X }
8 | «1 y \/ | a
A :
0 T T T T T T T - —
20 25 40 45 50
week
o
o
[zi
a
m
a
| —B- original data —-- reduced data | (b)
Figure 3: Polygons of coniferous forest plantation (a) and dry
heather moorland (b) before and after the selection
of weeks.
3.4 Filtering of the dataset
Filtering is a common technique for smoothing NDVI data.
Usually. either a weighted or unweighted moving average filter
or a median filter is used, sometimes both are combined
(Duchemin, 1999; Reed et al., 1994; Malingreau, 1986).
The moving average formula is based on the average value of
the variable over a specific number of preceding periods, e.g. 3
or 5. The moving average of 3, for example, is calculated using
the running NDVI and the values just before and after in the
time series.
After experimenting with a weighted moving average filter
(weighting coefficient of 0.25, 0.5. 0.25), median filtering, and
moving average filters of 3 and 5 points, an interval of 5
composite periods was finally selected for all cover types except
for pasture and arable land which were filtered with a 3 point
filter. It was discovered that a 5 point filter would smooth out
too many peaks and troughs in these cover types, so that the
overall trend was more difficult to establish. Figure 4 gives an
example (wet heather moorland 1996) of how the data were
smoothed, starting with the original data which is noisiest; the
reduced average. i.e. only pixels above a certain limit had been
included in the calculation, shows clearly that most troughs are
flattened. It can be seen that all the filtered curves are much
smoother. As a comparison the moving average of 5 points of
the original unsmoothed data is given. It is by about 0.01 to
0.04 units lower than the moving average of the curve with the
pixel limit.
UE
05 4
=
0.4 4 FA
03 4 je A 5
Ee of Ve X
2 à ao £84 oR FEN
2 Ca KM
02d É gi E A X
Fuld V
; 01 y x
0.1 / |
|
fi 4 i
-0,1 : 7 : 7 : : :
a 10 20 week 30 40 so
—4— original data —s-- reduced data
original data, filtered —#— reduced data, filtered
Figure 4: Filtering of the dataset (wet heather moorland 1996,
polygon 3): Original weekly composites, reduced
data and smoothed data (moving average filter of 5).
\ c e
4. RESULTS
[t was apparent that the original data curves of the NDVI are
very noisy. Sub-pixel clouds, cloud shadow. atmospheric effects
and the small number of polygons used in certain cover types
might all be factors that lead to the noisiness of the curves. The
missing atmospheric correction also means that the NDVI
values are lower as if correction had been undertaken.
The cloudiness of the images also results in a limitation of the
data that is available in each year. It can be observed that only
values for the weeks around 5-50 exist which means that
Scotland is extremely cloud-covered during the winter months.
Through the necessary smoothing of the curves, weeks were
climinated so that effectively NDVI data can only be displayed
for roughly weeks 8-45. A look at climate graphs validates that
the months with the highest precipitation are November (0
February. This, along with the low sunshine duration, implies a
high cloud probability. As photosynthetic activity is low during
this time and the winter months are therefore less interesting in
vegetation studies, the NDVI observations allow for an
interpretation of the most interesting part of the vegetation
cycle. Interpretation of some forest polygons is made difficult,
especially of semi-natural and mainly coniferous forest because
of limited data availability due mainly to cloud cover problems.
4.1 Intra-annual variations of the NDVI of different
vegetation types
The individual polygons of each vegetation type cover different
areas of Scotland. It was investigated whether differences in the
behaviour of the NDVI curve in the polygons of each cover type
were apparent. This is of interest for looking at the advance of
the ‘green wave’, the start of the growing period. Differences
arc to be expected: depending on the vegetation type, in the
Highlands and Islands of Scotland the green-up is likely to start
later than in the Lowlands. In the graphs, those differences are
only existent to a minor extent and could therefore not be
480
Intern
detect
figure
the b
take |
comp
Differ
polyg
the or
the gr
0
HDVI
o
Figur
4.2 I
4.2.1
Howe
unexp
respoi
annua
shows
the ye
consis
pastur
This i:
(0.5-0
respor
land,
season
in the
seen |
1998.
growit
values
class i
in the
the di:
class
outcro
the m
NDVI
to the
soils a
plants
Blanke
in mag
either
OF IS el