Use of digital aerial photography for sub-alpine ...labgis.ibot.cas.cz/wp-content/uploads/pdf... ·...
Transcript of Use of digital aerial photography for sub-alpine ...labgis.ibot.cas.cz/wp-content/uploads/pdf... ·...
Use of digital aerial photography for sub-alpine vegetation mapping: Acase study from the Krkonoše Mts., Czech Republic
Jana MüllerováInstitute of Botany, Academy of Sciences of the Czech Republic, 25243 Pr°uhonice, Czech Republic; (tel:+420271015278; fax: +420267750031; e-mail address: [email protected])
Received 23 July 2002; accepted in revised form 12 March 2004
Key words: Digital image processing, Field survey, Methodology assessment, Multispectral photography, Re-mote sensing, Vegetation analysis
Abstract
Multispectral aerial photographs from 1986, 1989 and 1997 were used for mapping and assessment of changes insub-alpine vegetation of the Krkonoše Mts Plateau �1300 to 1400 m a. s. l.�, Czech Republic. Scanned aerialphotographs were digitally processed by various remote sensing methods, such as histogram equalization, prin-cipal component transformation, Normalised Difference Vegetation Index, and classification �both unsupervised,i.e., Iterative Self-Organizing Data Analysis Technique, and supervised, i.e., Parallelepiped and Maximum Like-lihood algorithm�. The results were geometrically corrected by registration into orthophoto images. Field datafrom 1991 assisted in class labelling and vegetation description. A detailed vegetation analysis was conductedusing 1989 multispectral photographs since the field-verification and comparison to ground-based vegetation mapcould be done with only two-year difference. The unsupervised classification resulted in six classes �three ofthem mixed� with a 60.6% overall accuracy of unmixed classes. Some important vegetation types could not beseparated. The supervised classification resulted in nine classes with the overall accuracy of 81.1%. The mostcommon vegetation types were: mat-grass communities �40.4% of the total area� and tall grass and herb com-munities �23.1%�. Pine stands, the best separated of vegetation types, were evaluated on the multispectral pho-tographs from 1986 and 1997 and the P. mugo stands doubled in area over 10 years. The aerial-photograph basedmaps were compared with an available vegetation map obtained by a field survey. The correspondence betweenboth ranged from 24 � 75%, with the highest values obtained for pine stands and mat-grass communities. Incomparison to the field survey, aerial photography provided less detail and vegetation was characterized mostlyby dominant species. The approach failed to map small patches of heterogeneous or atypical vegetation and todifferentiate vegetation types with dominants of similar appearance. Nevertheless since this method is compara-bly less time- and labour-consuming, it seems suitable for mapping distinct vegetation types, surveying largerareas and identifying changes over time.
Introduction
Traditional field methods of vegetation mapping �ac-cording to the Braun-Blanquet floristic approach,Braun-Blanquet 1965� provide with a possibility todescribe vegetation in a large detail but are time/la-bour intensive and are biased by the subjectivity ofthe interpreter �Congalton 1991�. Remote sensing
�RS� can partly eliminate these problems but the na-ture of resulting information differs �Mosbech andHansen 1994�. Instead of focusing on communitystructure and floristic composition, RS is based ondominant species, their biomass �i.e., chlorophyllcontent and leaf area�, water content, physiognomy�e.g., height of plants, width of leaves and their posi-tion�, canopy structure and its character �e.g., canopy
Plant Ecology 175: 259–272, 2004.© 2004 Kluwer Academic Publishers. Printed in the Netherlands.
259
closure – a ratio of bare ground visible through thecanopy�, and soil conditions �Graetz 1989; Franklinet al. 1994�.
The relatively high spatial resolution of multispec-tral aerial photography makes this tool useful formapping vegetation composed of a mosaic of smallunits. The digital processing employed in the presentstudy has several advantages and disadvantages com-pared to the manual one. Manual vegetation mappingis not only time/labour intensive, hence expensive,but also subjective. Although a higher number ofclasses �some very small and unique� can beachieved, the classification is not reproductive.Labour intensity of manual interpretation usuallylimits the spatial extent of databases and the manualclassification accuracy is not readily assessed �Carmeland Kadmon 1998�. Digital RS classification is morehomogeneous and consistent. It often distinguishesfewer classes, but their number increases as the op-erator gets more experienced, the software/classifica-tion more sophisticated, and the amount of input data�multitemporal� or quality of input data �from multi-spectral to hyperspectral� increases. Computer pro-cessing must deal with changes in illumination acrossthe scene and other data irregularities. Yet it enablespartial correction of geometrical distortions, imageenhancement, filtering and other pre-processing tech-niques, statistical evaluation, visualisation of resultetc. �Pope et al. 1996�.
The present study examined the application ofmultispectral aerial data to the mapping of sub-alpinevegetation that is vulnerable to human-induceddisturbances and extremely valuable from the conser-vation point of view. The usefulness of digital inter-pretation of analogous aerial photography for vegeta-tion cover-type mapping, that would adequatelyaddress the needs for conservation purposes in thearea, is evaluated. The study was aimed at finding amethodology that could be applied to future aerialdata of the area so as to assess and monitor temporalchanges in the area �mainly due to human disturb-ances�.
The study was designed to answer the followingquestions: �1� Can aerial multispectral data be a use-ful tool in classifying sub-alpine vegetation? �2� Towhat extent can be different vegetation types sepa-rated on this kind of imagery? �3� Are the results ofaerial data analysis comparable with those obtainedin the field? �4� To what extent it is possible to detecttemporal changes in vegetation using time series ofaerial data?
Study Area
The Krkonoše Mts �Giant Mts, Riesengebirge� lie inthe northeastern part of the Czech Republic. Thestudy site �Figure 1� is located on the western of thetwo sub-alpine plateaus �1300 to 1400 m a. s. l, 15°33� E; 50° 45� N� belonging to the uppermost areasituated above timberline. It represents a particularlandscape system, often called arcto-alpine tundra�Soukupová et al. 1995�, with affinities to both sub-arctic and high-mountain regions and subjected toextreme climate conditions, such as exposition tostrong northwestern winds. Average temperature is 0to �1 °C, snow cover lasts more than 180 days, an-nual precipitation is 1400 mm, and there is periodicalsoil regelation and deflation. The plateaus are formedby crystalline metamorphic rocks, covered by pooracidic soils of the alpine sod podzol type, patches ofpeat bogs, raised bogs, polygonal and patterned soils.The treeless summits of the Krkonoše Mts harboursome unique physiographic components, such as relicand endemic populations and community types�Soukupová et al. 1995�.
The study area is heterogeneous in terms of veg-etation and abiotic environmental factors such asgeomorphology, texture, soils, microclimate and dis-turbance regime. Vegetation cover is composed ofgrassland, scrubland, fens and peat bogs. The mosttypical plant species include Pinus mugo �forming al-liance Pinion mughi and Oxycocco-Empetrion her-maphroditi�, creating large dense krummholz; Nardusstricta, a typical grass dominant of dense uniformstands of alliance Nardo-Caricion rigidae at the lo-calities of high snow cover and deeper humus podzolsoils based on granites; Molinia caerulea andCalamagrostis villosa forming acidophilous commu-nities of tall grasses and flowering herbs �allianceCalamagrostion villosae� on wet, wind-protectedsoils with long-lasting snow cover; Deschampsiacespitosa �a dominant of alliance Poo chaixii-Deschampsion caespitosae�, found on mezotrophic toeutrophic deeper, permanently humid soils; peatmosses; and Baeothryon cespitosum subsp. cespito-sum forming oligotrophic communities of boreoarcticand alpine raised bogs with prevalence of shrub formand peat mosses �alliance Oxycocco-Empetrion her-maphroditi�. For details see Table 4, where the fieldbotanical survey according to the Braun-Blanquetfloristic approach �Fišerová 1991� is summarized andthe percentage of the area covered by each vegetationtype given.
260
The study site situated on the Western Plateau cov-ered an area of ca 2 � 3 km. It was selected accord-ing the available field vegetation maps and aerial MSphotography. The plateaus themselves are relativelysmall �Western Plateau: 18.7 km2; Eastern Plateau: 29km2� and the study site therefore covered 30% of theWestern Plateau and 8% of the whole sub-alpine re-gion.
Material and methods
The following material was used in the analysis:– multispectral aerial photography of 1986 �acquired
20.9 1986; scale � 1:22000; channels � 0.48,0.54, 0.66, 0.84 �m�, 1989 �acquired 18.9 1989;scale � 1:22000; channels � 0.54, 0.60, 0.66,0.84 �m� and 1997 �acquired 12.9 1997; scale1:15000; channels � 0.54, 0.60, 0.66, 0.84 �m�;created and provided by the Agency for NatureConservation and Landscape Protection in Prague;
technical parameters – camera MSK-4, focal length125 mm, film material – FOMA �visible spectrum,1986 and 1989�, Aviphot Pan 200PE1 �visiblespectrum, 1997�, I-840 �NIR channel, 1986 and1989�, Kodak Aerographic Infrared Film 2424�NIR channel, 1997�
– panchromatic orthophotographs �1997; createdfrom scanned aerial panchromatic photographs ofscale 1:22500 with overlap of 60%, using a digitalterrain model created by vectorization of topo-graphic maps 1:10000, orientation points of imagesidentified by analytical aerotriangulation in systemORIMA; orthorectification performed on digitalphotogrammetric station Leica-Helava DPW 770,module Mosaic, final pixel resolution 0.5 m�, cre-ated by Geodézie Krkonoše �private company�,provided by the Krkonoše National Park Authority
– field botanical survey �Fišerová 1991; scale1:5000� – vegetation units drawn into 1989 aerialphotographs.
Figure 1. Location of the study site.
261
Aerial photography was chosen instead of the morecommonly used satellite data although it requires amore complicated rectification then the satellite dataand here is a complicated illumination effect acrossthe area combined with the variation in the movementof the aircraft. This approach was adopted since suchdata at a high spatial resolution �0.5 to 1 m� wereavailable �provided by the Krkonoše National Parkconservation authorities�. Analogous aerial photo-graphs �i.e., negatives� were transformed to digitalform by scanning the negative of each channel sepa-rately in resolution 1000 dpi to allow the computerprocessing. The digital type of processing wasselected since digital analyses can easily be repeatedon multitemporal data to evaluate changes over time.Scanned image channels were co-registered byimage-to-image registration and then geo-referencedusing the orthophotographs. Forty to fifty groundcontrol points distributed along whole rectified pho-tograph were collected. The second order of transfor-mation was applied. To minimize the errors caused bymis-registration the final resolution was coarsened topixel size 1 � 1 m. The effect of atmospheric trans-mittance and the angular effect, which would requireadditional data correction, were ignored, as well asthe terrain correction, since the area was generallyflat.
The detailed vegetation analysis was carried out onthe multispectral photographs of 1989. From the fourchannels available �0.54, 0.60, 0.66, 0.84 �m�, Redand Near-Infrared channels proved to be most useful.The digital form provided an opportunity to enhancethe visuality of vegetation types by histogram equal-ization �proved to by particularly useful�, principalcomponent transformation �PCA� and NormalizedDifference Vegetation Index �NDVI�, and to applyunsupervised and supervised classification methods.In the unsupervised approach the Iterative Self-Orga-nizing Data Analysis Technique �ISODATA� was em-ployed with following parameters � 20 clusters, 20iterations. The number of clusters �20� was selectedafter the clustering process from 2 to 30 clusters wasexamined. Higher number of clusters made their in-terpretation comparatively difficult. Resulting clusterswere grouped together according the field verifica-tion. In the supervised approach, Parallelepiped clas-sification �non-parametric method defining classes onthe basis of the lowest and the highest value in everyspectral band� was applied at first to pre-process thedata, to ease the training area selection and to definethe desired land-cover categories possible to distin-
guish on the image. The a priori probability of theclass occurrence determined by the Parallelepiped al-gorithm was then used in the Maximum Likelihoodanalysis, resulting in the final classification map. AMinimum-distance classifier was attempted but re-jected due to the poor results. Descriptions of themethods can be found in Jensen �1996�, Lillesand andKiefer �1999� or Mather �1999�. The digital analyseswere carried out using the Chips software �Chips De-velopment Team 1998�.
The resulting classes of both supervised and unsu-pervised approaches were examined using Chipssoftware to produce scattergrams to determine theconsistency of reflectance values and spectral differ-ence of the class signatures. The Jeffries-Matusitadistance measure of separability �Jensen 1996� wasused to assess the separability of pairs of classes �andtraining sites for supervised classification� on the ba-sis of their signatures.
An accuracy assessment of the 1989 results wascarried out using descriptive and analytical statistictechniques as described in Congalton �1991�. Evalu-ation areas used for the final assessment were posi-tioned randomly over the area and were independentof the training sites used for the supervised classifi-cation. Stratified random sampling was used to ensureall strata are included in the sample �Congalton andGreen 1998�. An overall accuracy as well as ‘user’saccuracy’ �the measure of commission error� and‘producer’s accuracy’ �the measure of omission error�of individual classes were calculated from the confu-sion matrix. The discrete multivariate technique usedin the accuracy assessment was the KAPPA coeffi-cient �Bishop et al. 1975�. After the validation the fi-nal results were smoothed using a 3 � 3 mode filter.Synoptical maps of the study area were thenproduced.
The RS results of 1989 were compared with thefield survey data �FS� of Fišerová �1991�. The FS in-corporated field observation and visual interpretationof the same aerial photographs, classifying the veg-etation at the level of communities or their transitionsaccording to the Braun-Blanquet floristic approach�Braun-Blanquet 1965�. The FS map was digitisedusing the CartaLinx software �Clark Labs 1998�. In-formation contained in the map was utilized for thetraining area selection and classification interpretationof the RS analysis. The RS and FS maps were com-pared in terms of the number of distinguished vegeta-tion types, their distribution, extent, and the mappingaccuracy. The level of correlation between ground
262
and aerial-based vegetation maps was evaluated. RSand FS approaches were compared in regard of thearea covered by each vegetation type as well as onthe cell-by-cell bases in a GIS environment usingArcView GIS �Environmental Research Inst. 1999�.The percentages of corresponding cells and the mostcommonly confused classes were identified in thecorrespondence matrices and it was possible todeduce likely reasons for the differences. To test thepossible distortion of RS/FS comparison caused byregistration errors, the pixel dimension was coarsenedfrom 1 � 1 to 2 � 2, 3 � 3 and 4 � 4 m and thenew RS/FS correspondence matrices were generated.
Interpretation of multispectral data sets of 1986 and1997 focused on the best separable vegetation type –P. mugo, describing its dynamics during 10 years. Theunsupervised classification was applied since in theprevious detail analysis this simple and fast methodproved to be sufficient for such purpose.
Nomenclature of taxa follows Rothmaler et al.�1991�, that of syntaxa Moravec �1995�.
Results
Unsupervised classification
The unsupervised �ISODATA� classification �Figure2� was only moderately successful in determining themain vegetation and land-cover types, with some cat-egories mixed. Fifteen classes were identified. Sincemany classes included the same vegetation/land-cover types, they were merged into six resultingclasses �Table 1�: �1.� pine stands �Pinus mugo domi-nated stands�; �2.� mat-grass communities �Nardusstricta dominated stands�; �3.� tall grass and herbcommunities �dominated by Calamagrostis villosa orMolinia caerulea�; �4.� road-altered vegetation mixedwith tall grass and herb communities; �5.� roads ormat-grass; �6.� roads or short grasses and herbs �De-schampsia cespitosa�. The last three classes weremixed, covering 25% of the area �Figure 3�. Someimportant vegetation types were not separated, suchas mat-grass and roads in the mixed class 5.
The resulting classification was verified in the field.As classification accuracies �Table 2� could not becalculated for the mixed classes, the overall accuracyof 63% and KHAT statistics of 45% is only based onclasses 1 � 3 �covering 75% of the area�. The user’saccuracy ranged from 58 to 74% and the producer’saccuracy from 55 to 72%. The best results, 74% �us-
er’s accuracy� and 72% �producer’s accuracy�, wereobtained for pine stands. The Jeffries-Matusita dis-tance measure of separability �growing from 0 to1.41� showed the highest separability �above 1.2� forclasses 1, 4 and 6, and the value of 0.9 for classes 2/5and 2/3.
Supervised classification
Minimum distance and Maximum Likelihood algo-rithms were tested. The former did not show satisfac-tory results and was therefore rejected �such as inFuller and Parsell 1990�.
Resulting classes of the Maximum Likelihoodclassification �Table 1�:1. Pine stands – Pinus mugo, forming krummholz,
i.e., procumbent shrubs, represents dominatingwood above the timberline. Constrained polycor-mons of P. mugo �Figure 2�, inhabiting shelteredforks and patches of organic and mineral soils ex-pand to surrounding grasslands. Present communi-ties belong to the alliance Pinion mughi, associa-tion Myrtillo-Pinetum mughi and allianceOxycocco-Empetrion hermaphroditi with an en-demic association Chamaemoro-Pinetum mughi. P.mugo is partly natural to the area and partlyplanted. Plantation resulted in closed-canopystands with a regular pattern sharply contrastingwith jagged boundaries of the natural ecosystems�Soukupová et al. 1995�.
2. Mat-grass communities are dominated by Nardusstricta and belong to the alliance Nardo-Caricionrigidae �as. Carici fyllae-Nardetum�. N. stricta re-presents a prevailing dominant of local grasslands�Figure 2� and has a low stress tolerance and highvulnerability to trampling and eutrophication. It isendangered by expanding P. mugo krummholzsince these krummholz favour concurrent gramin-eous species. The class was further subdivided intothree categories �A, B and C� expressing the tran-sition to tall grass and herb communities or thatbetween mineral and peat soils �from A to C�, fromsites dominated by N. stricta �pH 4.5� to those withprevailing Molinia caerulea and N. stricta as acon-dominant �subas. Carici fyllae-Nardetum mo-linietosum caeruleae; pH 3.7–3.9�.
3. Tall grass and herb communities are acidophilousand occur on granite-based wet, wind-protectedsoils with a long lasting snow cover; they aredominated by Calamagrostis villosa, Veratrum al-bum L. subsp. lobelianum �Bernh.� Rchb. and
263
Figure 2. Comparison of the results of three vegetation mapping methods: �a� the field survey �Fišerová 1991�; �b� multispectral aerial photointerpretation – supervised classification approach �Maximum Likelihood algorithm�; �c� multispectral aerial photo interpretation – unsuper-vised classification approach �Iterative Self-Organizing Data Analysis Technique – ISODATA�.
264
Tabl
e1.
Sum
mar
yof
the
aeri
alph
otog
raph
ycl
assi
ficat
ion,
both
unsu
perv
ised
�ISO
DA
TA�
and
supe
rvis
ed�M
axim
umL
ikel
ihoo
d �cl
assi
ficat
ion
appr
oach
es.
Uns
uper
vise
dcl
assi
ficat
ion
�ISO
DA
TAal
gori
thm
�Su
perv
ised
clas
sific
atio
n�M
axim
umL
ikel
ihoo
dal
gori
thm
�
No
Cla
ssN
ame
Cla
ssch
arac
teri
stic
s%
ofar
eaN
oC
lass
Nam
eC
lass
char
acte
rist
ics
%of
area
1pi
nest
ands
Com
pact
stan
dsof
Pin
usm
ugo
proc
umbe
ntsh
rubs
17.1
%1
pine
stan
dsse
eIS
OD
ATA
clas
s1.
18.6
%
2m
at-g
rass
com
mun
ities
Nar
dus
stri
cta
dom
inat
edst
ands
,pr
evai
ling
gras
slan
dty
pe37
.6%
2m
at-g
rass
com
mun
ities
see
ISO
DA
TAcl
ass
2.40
.4%
2.1
mat
-gra
ssco
mm
uniti
esA
site
sdo
min
ated
byN
.st
rict
a�p
H4.
5 �on
min
eral
soils
14.2
%
3ta
llgr
ass
&he
rbco
mm
uniti
esac
idop
hilo
usco
mm
uniti
esof
tall
gras
ses
and
flow
erin
ghe
rbs
onw
etso
ilsdo
min
ated
byC
alam
agro
stis
vill
osa,
Vera
trum
albu
msu
bsp.
lobe
lian
um,
Sene
cio
herc
ynic
usan
dM
olin
iaca
erul
ea
20.3
%2.
2m
at-g
rass
com
mun
ities
Btr
ansi
tion
betw
een
Aan
dC
7.9%
2.3
mat
-gra
ssco
mm
uniti
esC
site
sw
ithpr
eval
ence
ofM
olin
iaca
erul
eaw
ithN
.st
rict
aas
aco
n–do
min
ant
�pH
3.7–
3.9 �
onpe
atso
ils,
tran
sitio
nto
tall
gras
s&
herb
com
mun
ities
18.3
%
4ro
ad-a
ltere
dve
geta
-tio
nan
dta
llgr
ass
&he
rbco
mm
uniti
es
mix
edcl
ass
ofve
geta
tion
inth
ero
advi
cini
tyal
tere
dby
leac
hing
iont
sfr
omal
kalin
ero
ad-b
uild
ing
mat
eria
l,an
dta
llgr
ass
&he
rbco
mm
uniti
es
15.3
%3
tall
gras
s&
herb
com
mun
ities
see
ISO
DA
TAcl
ass
3.23
.1%
4ro
ad-a
ltere
dve
geta
tion
see
ISO
DA
TAcl
ass
4.11
.5%
5ro
ador
mat
-gra
ssco
mm
uniti
esm
ixed
clas
s3.
5%5
road
sha
rden
edby
dolo
mite
and
mel
aphy
regr
avel
,so
me
cove
red
byas
phal
t4.
8%
6ro
ador
shor
tgr
asse
s&
herb
sm
ixed
clas
sof
road
san
dco
mm
uniti
esdo
min
ated
byD
esch
amps
iace
spit
osa
6.1%
6w
ater
pool
san
dw
ater
logg
edde
pres
sion
s0.
3%
7w
etar
eas
peat
bogs
,st
ream
allu
vium
,su
rrou
ndin
gsof
Pin
usm
ugo
shru
bs0.
7%
8un
clas
sifie
d0.
6%
Ove
rall
accu
racy
ofun
mix
edcl
asse
s�
63.1
%;
Kh
atof
unm
ixed
clas
ses
�44
.5%
;O
vera
llac
cura
cy�
78.8
%;
Kh
at�
74.6
%
265
Senecio hercynicus Herborg �alliance Calamagros-tion villosae, as. Crepido-Calamagrostietum villo-sae�, or Molinia caerulea stands of subas. Caricifyllae-Nardetum molinietosum caeruleae.
4. Road-altered vegetation occurs along roads pavedand hardened by allochtonous alkaline buildingmaterial. Alkaline ions are continuously washedinto the originally nutrient poor environment, sup-porting tall herbs and grasses growing on soilsricher in nutrients �this vegetation can be partlyclassified as Bistorto-Deschampsietum alpicolae orCrepido-Calamagrostietum villosae associations�with a high amount of ruderal species �Urtica dio-ica, and indigenous Rumex alpinus�, species oflower altitudes �Senecio ovatus �Gaertn., B. Mey.et Scherb.� Willd., Hypericum maculatum and Cir-sium arvense� and expansive local species �e.g.,Senecio hercynicus, Cirsium helenioides, Veratrumalbum subsp. lobelianum, Deschampsia cespitosa,Calamagrostis villosa�.
5. Roads were hardened by dolomite and melaphyregravel, some covered by asphalt.
6. Water bodies are small pools and waterlogged de-pressions; their size and overgrowth by Sphagnumsp. differ according to fluctuating precipitation.
7. Wet areas are peat bogs, stream alluvium and sur-roundings of Pinus mugo shrubs.
The separability of supervised classification trainingclasses measured by the Jeffries-Matusita distance�growing from 0 to 1.41� showed values above onefor non-vegetation training classes and for the pineclass. Tall grass and herb communities trainingclasses showed possible confusion with mat-grass�0.98� and road-altered vegetation �0.74�, similarseparability was achieved for the resulting classes.
An NDVI analysis was carried out on the 1989photo only. It was used mainly to enhance the visu-ality of classes prior the classification and to enablethe selection of classes possible to discriminate on thephotograph. NDVI values were higher because of ra-diance used instead of reflectance values since an at-mospheric correction of the image was not performed.Vegetation classes of highest mean NDVI were pinestands, tall grass and herb communities and road-al-tered vegetation, while non-vegetation structures re-ceived NDVI values under zero.
The most common classes were mat-grass commu-nities �the most frequent sub-class C�, tall grass andherb communities and pine stands �Figure 3�. The us-er’s accuracy ranged from 62 to 87% and the produc-er’s accuracy from 70 to 88% with the overallaccuracy of 79% and KHAT statistic of 75%. The bestaccuracy was achieved for pine stands and mat-grasscommunities �Table 3�.
Figure 3. Diagrams showing the distribution of classes in both RS data �ISODATA and Maximum Likelihood� and the field survey �FS�.Numbers and letters representing class names are explained in Figure 2 �RS�, and in Table 4 �FS�.
Table 2. Unsupervised classification accuracy, confussion matrix. Numbers correspond to the number of pixels.
ISO classification Pine stands Mat-grasscommunities
Tall grass & herbcommunities
User´s accuracy
Pine stands 71 12 13 74%Mat-grass communities 20 72 32 58.1%Tall grass & herb communities 8 30 54 58.7%Producer´s accuracy 71.7% 63.2% 54.5% 63.1%
266
Comparison of different approaches
The field survey and the RS approach are comparedin Table 4. Eleven vegetation types were recognizedin the field map �Fišerová 1991, Figure 2�. FS veg-etation classes were related to the RS classes �seeTable 4�. Three FS vegetation types dominated by P.mugo contributed to the pine category, two FS classesdominated by Callamagrostis villosa and Moliniacaerulea contributed to the category of tall grassesand herbs. A large group of FS classes could not bedistinguished by RS means or related directly to theRS classes, i.e., Scirpo caespitosi-Sphagnetum com-pacti �classified predominantly as mat-grass or tallgrass and herb communities�, Salicetum lapponum�mostly classified as tall grass and herb communitiesor road-altered vegetation�, stands dominated byCarex rostrata �accordingly classified as tall grassand herb communities�, transitions between M. caer-ulea and Carici fyllae-Nardetum �classified accordingto the character of the transition as mat-grasscommunities or tall grass and herb communities, i.e.,M. caerulea dominated�, transitions between Caricifyllae-Nardetum and Myrtillo-Pinetum mughi �classi-fied as pine stands, tall grasses and herbs or mat-grasscommunities. These classes �covering 17.5% of thearea� were joined into a residual group.
A similar distribution of categories was obtained byusing the three RS approaches �Figure 3�, while theFS showed lower values for the tall grasses and theroad-altered vegetation in favour of the mat-grass�and the residual group�. The classes contributing tothe residual group were excluded from the correspon-dence assessment �Table 5, Table 6�. The resultingcorrespondence values �RS/FS� varied from 24% to75% with the overall correspondence of 59% for the
unsupervised approach and 55% for the supervisedone. The highest FS/RS correspondence values wereachieved for the pine stands �over 70%�.
Enlargement of the cell dimension from 1 � 1 to 2� 2, 3 � 3 and 4 � 4 m had negligible impact onoverall correspondence values and never exceeded1%.
Change detection
Due to the difficulties arising from the nature of themultispectral data �varying quality and exposure, in-accuracy of the geometric correction etc.� only thebest separable vegetation type �pine stands� was se-lected to assess the changes over the time covered byavailable data �1986-1997�. Unsupervised ISODATAclassification approach was employed. Over 11 years,the area covered by P. mugo gradually increased. Itcovered 4.5% of the area in 1986, 6.1% in 1989 and10% in 1997; this represents 40% increase in 1986-1989 and 50% increase in 1989-1997, resulting inmore then doubled area over the whole period �Fig-ure 4�. The average increase calculated for the 11years was 2.8 hectares per year.
Discussion
Aerial data analysis
The present study has shown that aerial data can pro-vide valuable information about vegetation in sub-al-pine region, with comparatively high classificationaccuracy and low class confusion. Still some vegeta-tion types were not possible to identify due to theirspectral or species similarities or small size of
Table 3. Supervised classification accuracy, confussion matrix. Numbers correspond to the number of pixels.
Maximum Likelihoodclassification
Pine stands Mat-grasscommuni-ties
Tall grass &herb com-munities
Road-alteredvegetation
Roads Water Wet areas User´saccuracy
Pine stands 84 2 3 2 1 3 2 86.6%Mat-grass communities 6 95 8 3 0 0 0 84.8%Tall grass & herb com-munities
3 8 79 12 1 0 3 74.5%
Road-altered vegetation 2 6 8 75 4 0 5 75%Roads 0 2 2 5 45 3 0 78.9%Water 0 0 0 0 6 24 0 80%Wet areas 0 4 8 0 0 2 23 62.2%Producer´s accuracy 88.4% 81.2% 73.1% 77.3% 78.9% 75% 69.7% 78.8%
267
Tabl
e4.
Com
pari
son
ofth
egr
ound
and
the
rem
ote
sens
ing
appr
oach
inve
geta
tion
map
ping
�the
field
vege
tatio
nsu
rvey
ofFi
šero
vá19
91,
and
the
aeri
alph
oto
digi
tal
inte
rpre
tatio
n �.
Tra
ditio
nal
field
map
ping
,ac
cord
ing
toth
eB
raun
-Bla
nque
tflo
rist
icap
proa
ch�p
erce
ntag
eof
the
tota
lar
eain
brac
kets
�U
nsup
ervi
sed
clas
sific
atio
n�I
SOD
ATA
�Su
perv
ised
clas
sific
atio
nM
axim
umL
ikel
ihoo
d
A–
Pine
stan
dsC
ham
aem
oro-
Pin
etum
mug
hiH
adac
etV
ána
1967
�11
%�
Pine
stan
ds�c
lass
1.�
Pine
stan
ds�c
lass
1.�
Myr
till
o-P
inet
umm
ughi
Had
ac19
56�1
.1%
�Pi
nest
ands
�cla
ss1.
�Pi
nest
ands
�cla
ss1.
�M
yrti
llo-
Pin
etum
mug
hi–
open
stan
dsw
ithab
unda
nce
ofN
.st
rict
a,tr
ansi
tion
toC
aric
ify
llae
-Nar
detu
m�4
%�
Pine
stan
ds�c
lass
1.�
Pine
stan
ds�c
lass
1.�
B–
Mat
-gra
ssC
aric
ify
llae
-Nar
detu
mJe
ník
1961
�50.
4%
�M
at-g
rass
com
mun
ities
�cla
ss2.
,m
ixed
clas
s5.
,6.
�M
at-g
rass
com
mun
ities
�cla
ss2.
�
C–
Tall
gras
ses
&he
rbs
Mol
inia
caer
ulea
stan
ds�1
0.7
%�
Cre
pido
-Cal
amag
rost
ietu
mvi
llos
aeJe
ník
1961
�0.6
%�
Tall
gras
s&
herb
com
mun
i-tie
s�c
lass
3.,
mix
edcl
ass
4.�
Tall
gras
s&
herb
com
mun
ities
�cla
ss3.
�
Tall
gras
s&
herb
com
mun
i-tie
s�c
lass
3.,
mix
edcl
ass
4.�
Roa
d-al
tere
dve
geta
tion
�cla
ss4.
,48
%�
orta
llgr
ass
&he
rbco
mm
uniti
es�c
lass
3.,
40%
�
D–
Roa
d-al
tere
dve
geta
tion
Bis
tort
o-D
esch
amps
ietu
mal
pico
lae
Bur
ešov
á19
76�4
.7%
�w
ater
�0.2
%�
Roa
d-al
tere
dve
geta
tion
�mix
edcl
ass
4.�
Roa
d-al
tere
dve
geta
tion
�cla
ss4.
�
clas
sifie
dm
ostly
asPi
ne�c
lass
1.�
Wat
er�c
lass
6.�
E–
Res
idua
lgr
oup
Sali
cetu
mla
ppon
umM
atus
zkie
wic
z19
65�0
.3%
�st
ands
dom
inat
edby
Car
exro
stra
ta�0
.4%
�cl
assi
fied
ascl
ass
4.�5
8%
�or
clas
s3.
�17%
�cl
assi
fied
ascl
ass
4.�4
8%
�or
clas
s3.
�37%
�
Scir
poca
espi
tosi
-Sph
agne
tum
com
pact
iW
aren
1926
�5.3
%�
tran
sitio
nC
aric
ify
llae
-N
arde
tum
→M
yrti
llo-
Pin
etum
clas
sifie
das
clas
s4.
�35
%�,
clas
s3.
�29%
�,cl
ass
2.�1
9%�
orcl
ass
1.�1
4%�
Tall
gras
s&
herb
com
mun
ities
�56%
�
mug
hi�8
.8%
�tr
ansi
tion
M.
caer
ulea
→C
aric
ify
llae
-Nar
detu
m�2
.5%
�cl
assi
fied
ascl
ass
2.�3
7%
�,cl
ass
3.�2
7%�
orcl
ass
4.�2
4%�
clas
sifie
das
clas
s3.
�37
%�,
clas
s2.
3�2
9%�,
clas
s2.
1�1
8.5%
�or
clas
s4.
�18%
�
clas
sifie
das
clas
s2.
�47%
�,cl
ass
1.�3
0%�
orcl
ass
3.�1
4%�
clas
sifie
das
clas
s1.
�32%
�,cl
ass
3.�2
4%�,
clas
s2.
1.�2
0%�
orcl
ass
2.3.
�15.
5%�
clas
sifie
das
clas
s2.
�57
%�,
clas
s3.
�23,
5%�
orcl
ass
4.�1
5%�
clas
sifie
das
clas
s2.
3.�3
0%
�,cl
ass
3.�2
9%�
orcl
ass
2.1.
�27%
�
268
patches. Yet the not discerned vegetation types werenot of main importance. Stands of uniform vegetationshowed higher classification accuracy compared tothose with heterogeneous, atypical or transitionalvegetation, which were difficult or impossible to map.As Treitz et al. �1992� pointed out, the more spectrallyunique the class is and the lower species variation itshows, the higher is its classification accuracy.
The supervised classification technique �MaximumLikelihood method�, providing higher overall accu-racy, proved to be more successful and accurate thanthe unsupervised �ISODATA algorithm�. As the unsu-
pervised approach produced mixed classes and didnot differentiate some land-cover types �e.g., road-al-tered vegetation�, it appears to be insufficient for adetailed vegetation study. However as a fast and easymethod it worked well for a simple task of Pinusmugo mapping, showing comparably high classifica-tion accuracy of 74% �user’s� and 72% �producer’s�.The overall accuracy and the KHAT statistic areover-estimated since they represent only the three un-mixed classes, two of them belonging to the mostdistinct vegetation types.
Table 5. Correspondence matrix of the aerial photography interpretation �unsupervised approach� and the field survey. Numbers correspondto the number of pixels.
ISODATA classification �RS� Pine stands Mat-grasscommunities
Tall grass & herbcommunities
RS/FS correspondence
Pine stands 40475 12612 3810 71.1%Mat-grass communities 23978 140789 15527 78.1%Tall grass & herb communities 5368 57776 20112 24.2%FS/RS correspondence 58.0% 66.7% 51.0% 62.8%
Table 6. Correspondence matrix of the aerial photography interpretation �supervised approach� and the field survey. Numbers correspond tothe number of pixels.
Maximum Likelihood �RS� Pine stands Mat-grasscommunities
Tall grass &herb communi-ties
Road-alteredvegetation
Water RS/FS corre-spondance
Pine stands 49383 16534 4726 1275 572 68.1%Mat-grass communities 19069 146528 20418 2164 114 77.8%Tall grass & herb communities 7052 63794 25515 6520 13 24.8%Road-altered vegetation 2984 18371 6929 9811 0 25.8%Water 171 0 0 0 245 58.9%FS/RS correspondence 62.8% 59.8% 44.3% 49.6% 26.0% 57.6%
Figure 4. Changes in the area covered by Pinus mugo over the period of 1986-1997 detected from multispectral aerial photographs. Theblack column gives the total area covered by P. mugo in particular years in hectares, the white column discribes its overall increase.
269
In the supervised approach, class confusion oc-curred in some cases. Mat-grass communities weresometimes misclassified as roads or as P. mugostands. Road-altered vegetation and tall grass andherb communities were confused, which is notsurprising considering that they have a similar speciescomposition. Pine stands were mis-classified as wateror wet areas; P. mugo krummholz often grows on hu-mid soils or surrounded by standing water at stem,which changes the resulting spectral reflectance.
Comparison of different approaches
The ground inventory achieved a higher number ofvegetation classes, and the vegetation units werelarger, with more homogeneous and less complextexture. Not all the categories classified by FS couldbe discriminated from aerial data and vice versa andthe mapped vegetation units differed. It was thereforenot possible to compare the aerial-based vegetationmap directly to the field survey based on phytosocio-logical methods. For example, different plant commu-nity types dominated by P. mugo could not bediscriminated on aerial photographs, because domi-nating shrubs covered under-storey grasses and herbsand contributed strongly to the resulting spectral re-flectance. Transitional classes of field mapping wererelated to photo-interpreted classes that included plantspecies engaged in transition, and could therefore beregarded as correctly classified. The term ‘transi-tional’ is subjective and such stands could hardly beclassified equally by a different analyst. Some classes,such as stands dominated by Carex rostrata and thecommunity Crepido-Calamagrostietum villosae, werenot separated on aerial photographs; however, theirprevailing inclusion into classes representing standsof higher biomass �such as tall grass and herb com-munities or road-altered vegetation� seems to be just.Some classes obtained by photo-interpretation werenot determined by the field mapping �wet areas anddivision of mat-grass communities into three sub-cat-egories�, or were delineated only partly such as theroad-altered vegetation �Bistorto-Deschampsietum al-picolae in Fišerová 1991�.
For the purpose of a correspondence assessment,the FS vegetation types that were impossible to relateto the RS classes were joined into a large residualgroup, covering over 17% of the total area. This factbiased the correspondence and supposedly loweredthe level of correlation; still it is probable that an at-tempt to make questionable relationships between all
RS and FS classes would significantly amplify the er-ror. The best correspondence values resulted frompine and mat-grass classes �both RS classificationtypes�. Higher correspondence values for the unsuper-vised approach, compared to the supervised, werecaused by a lower number of assessed classes. In theunsupervised classification only the three unmixedclasses were taken into account, two of them ratingamong the best discernible. This fact artificially en-hanced the resulting correspondence �as well as theaccuracy� of the unsupervised classification.
As for the supervised classification and the FS, theproportion of water bodies �that are easy to discrimi-nate� was very similar, and those of pine stands andmat-grass communities were approximate �Figure 3�.Slightly larger area of pine stands �RS� seems closerto reality, as ground inventory included only largerpolygons and small individual shrubs were omitted.The RS approach gave comparably higher proportionof tall grasses and herbs �two times higher� and of theroad-altered vegetation �three times higher�. The cat-egory of tall grasses and herbs seems to be overesti-mated by the RS approach and underestimated by FS.The amount of vegetation classified as road-alteredwas more than three times larger when using RS ap-proach, compared to FS. This can be explained by thefact that human-made sites with this vegetation typewere not in the focus during FS, which resulted in itsunderestimation. Still some of the pixels included intothe road-altered vegetation RS class are located fur-ther off the road and belong to tall grass and herbcommunities that show similar characteristics andshare some plant species. Some categories were dif-ficult to compare, as they only partly overlap �e.g.transitional types�.
The results are consistent with those reported onthe comparison of ground- and remote sensing-basedvegetation mapping in arctic �Mosbech and Hansen1994, Spjelkavik 1995� and alpine regions �Frank1988�.
Problems with interpretation of aerial photography
Digital processing minimized inaccuracy originatingfrom the subjective element. Nevertheless, co-regis-tering the channels of the multispectral images priorto RS classification process was not perfect and pro-duced mis-registration error �mentioned e.g., in Cihlaret al. 1996�. As for both types of RS classification�compared in Ferguson 1991 and Joria and Jorgensen1996�, results could not have been field-verified by
270
the author at the same year the aerial data were gath-ered. Some causes of misclassification are thereforeunknown. The vegetation map of Fišerová �1991� isnot completely reliable as it was drawn into raw,non-rectified photographs and was therefore subjectto a geometric distortion. Nevertheless, since the al-teration of cell dimension did not change markedlythe RS/FS correspondence values it seems that theimpact of registration errors was not fundamental.Hereafter there is a question of the FS vegetation mapaccuracy. As Congalton �1991� stated, accuracy ofground surveys has been traditionally accepted with-out any confirmation and digital classifications areoften assessed with its reference. An assumption thatthe ground inventory is 100% correct is rarely validand can bias the digital classification assessment.
RS methods do not produce sharp boundaries be-tween vegetation types, which in reality form a con-tinuum rather then a well-defined mosaic. It can beargued that the classification error is partly caused byan attempt to place a boundary where in reality a gra-dient of change exists �Wood and Foody 1989�.Mixed pixels in transition zones contribute to the het-erogeneity of the resulting map, reflecting naturalconditions �Mosbech and Hansen 1994�.
Change detection
An analysis of vegetation changes from availablemultispectral aerial data failed to provide satisfactoryresults, because the field verification of older data wasnot possible. The results of change detection werealso distorted by problems related to the nature ofaerial data, i.e., the mis-registration and image expo-sure differences �see Bakker et al. 1995 andHolmgren et al. 1997�. Only the best recognizablevegetation type, P. mugo dominated stands, wastherefore chosen for the analysis. The ISODATA al-gorithm, previously proven to be convenient for sucha task, was employed as a fast and simple method,possible to apply over the large area of the NationalPark with satisfactory results. Since all three photo-graphs were taken at the end of the growing season�the second half of September� problems caused bythe shift of the phenological stage and consequentover/under estimation of changes were eliminated.The rapid increase of the area covered by P. mugowas caused mainly by aging of already establishedpine polycormons �growth and expansion ofbranches�.
Relevance to nature conservation
Previous dendrochronological studies �Soukupová etal. 1995� showed that under extreme climatic condi-tions the establishment of P. mugo on the Plateau is adelicate process and progressive spontaneous dis-persal of its population is slow. Still the expansion ofP. mugo represents a potential threat to local herbspecies, valuable from the conservation point of view,and to the unique geomorphologic components, suchas relic polygonal and patterned soils. These uniquesoils were developed in periglacial conditions of gla-cial periods and can be still preserved and activelyshaped if cryogenic processes �periodical soil regela-tion and deflation� continue. Such processes are dis-rupted if the surface is overgrown with shrubs. Thedescription of P. mugo dynamics is thus of an emi-nent interest to the nature conservation authorities andthe resulting maps will be used for conservation pur-poses.
Field inventory of vegetation is a time-consumingand laborious process. RS analysis, conducted in de-tail in one area, represents a methodology that can beextrapolated over larger areas covered by similarvegetation. The possibility of time-series studies ofRS data allow us to follow vegetation changes, andpossibly to separate the natural variation in vegetationattributable to natural factors from that induced byhuman impact.
Acknowledgements
I would like to thank B.U. Hansen, B.N. Rock, P. Py-šek and G. Rees for their help with the research andvaluable comments on the manuscript. Thanks to theKrkonoše National Park Authority, the Agency forNature Conservation and Landscape Protection inPrague, and D. Fišerová for providing the data. TheInstitute of Geography, Copenhagen University cour-teously made available the Chips software.
References
Bakker W.L., Honaker J.J. and Weisberg P.J. 1995. Using aerialphotography and GIS to map the forest-tundra ecotone in RockyMT. NP, Colorado, for global change research. PhotogrammetricEngineering and Remote Sensing 61�3�: 313–320.
Bishop Y.M.M., Fienberg S.E. and Holland P.W. 1975. Discretemultivariate analyses: Theory and practice. MIT Press, Cam-bridge, Massachusetts, UK.
271
Braun-Blanquet J. 1965. Plant sociology: The study of plant com-munities. Hafner, London, UK.
Carmel Y. and Kadmon R. 1998. Computerized classification ofMediterranean vegetation using panchromatic aerial photo-graphs. Journal of Vegetation Science 9: 445–454.
Chips Development Team 1998. Chips 4.3, the Copenhagen ImageProcessing System. University of Copenhagen, Copenhagen,Denmark.
Cihlar J., Ly H. and Xiao Q. 1996. Land cover classification withAVHRR multichannel composites in Northern environments.Remote Sensing of Environment 58: 36–51.
Clark Labs 1998. CartaLinx 1. 04, the spatial data builder. ClarkUniversity, Worcester, UK.
Congalton R.G. 1991. A review of assessing the accuracy of clas-sifications of remotely sensed data. Remote Sensing of Environ-ment 37: 35–46.
Congalton R.G. and Green K. 1998. Assessing the accuracy of re-motely sensed data: Principles and Practices. Lewis Publ.
Ferguson R.S. 1991. Detection and classification of Muskox habi-tat on Banks Island, Northwest territories, Canada, using Land-sat TM data. Arctic. 44�Supp.1�: 66–74.
Fišerová D. 1991. Map of actual vegetation of western Krkonoše.Report, manuscript on file. Agency for Nature Conservation andLandscape Protection of the Czech Republic, Prague, in Czech.
Frank T.D. 1988. Mapping dominant vegetation, communities inthe Colorado Rocky Mountain front range with LandsatThematic Mapper and digital terrain data. Photogrammetric En-gineering and Remote Sensing 54: 1727–1734.
Franklin S.E., Connery D.R. and Williams J.A. 1994. Classifica-tion of alpine vegetation using Landsat Thematic Mapper, SPOTHRV and DEM data. Canadian Journal of Remote Sensing 20:49–56.
Fuller R.M. and Parsell R.J. 1990. Classification of TM imagery inthe study of land use in lowland Britain: practical considerationsfor operational use. International Journal of Remote Sensing 11:1901–1917.
Graetz R.D. 1989. Remote sensing of terrestrial ecosystem struc-ture: an ecologist’s pragmatic view. Ecological studies 79: 5–30.
Homlgren P., Thuresson T and Holm S. 1997. Estimating forestcharacteristics in scanned aerial photographs with respect to re-
quirements for economic forest management planning. Scandi-navian Journal of Forest Research 12: 189–199.
Jensen J.R. 1996. Introductory digital image processing. Prentice-Hall Int., London, UK.
Joria P.E. and Jorgenson J.C. 1996. Comparison of three methodsfor mapping tundra with Landsat digital data. PhotogrammetricEngineering and Remote Sensing 62: 163–169.
Lillesand T.M. and Kiefer R.W. 1999. Remote sensing and imageinterpretation. John Wiley and Sons, New York, USA, 3th ed.
Mather P.M. 1999. Computer processing of remotely-sensedimages: an introduction. John Wiley and Sons, Chichester, UK.
Moravec J. , �ed.� 1995. List of plant communities of the CzechRepublic and their endangerment. Severoceskou prírodou,Litomerice, in Czech.
Mosbech A. and Hansen B.U. 1994. Comparison of satellite imag-ery and infrared aerial photography as vegetation mappingmethods in an arctic study area; Jameson Land, East Greenland.Polar Research 13: 139–152.
Pope P., Van Eeckhout E. and Rofer C. 1996. Waste site character-ization through digital analysis of historical aerial photography.Photogrammetric Engineering and Remote Sensing 62�12�:1387–1394.
Rothmaler W., Jäger E., Schubert R. and Werner K. 1991. Exkur-sionsflora von Deutschland. Volk und Wissen, Berlin, Germany.
Soukupová L., Kociánová M., Jeník J. and Sekyra J. 1995. �eds�,Arctic-alpine tundra in the Krkonoše, the Sudetes. Operacorcontica 32: 5–88.
Spjelkavik S. 1995. A satellite-based map compared to a traditionalvegetation map of arctic vegetation in the Ny-Aalesund area,Svalbard. Polar Record 31: 257–269.
Treitz P.M., Howarth P.J., Suffing R.C. and Smith P. 1992. Appli-cation of detailed ground information to vegetation mappingwith high spatial resolution digital imagery. Remote Sensing ofEnvironment 42: 65–82.
Wood T.F. and Foody G.M. 1989. Analysis and representation ofvegetation continua from Landsat Thematic Mapper data forlowland heats. International Journal of Remote Sensing 10:181–191.
272