URBAN LAND COVER MAPPING USING HYPERSPECTRAL AND MULTISPECTRAL VHRSENSORS: SPATIAL VERSUS SPECTRAL RESOLUTION

Fabio DELL’ACQUA, Paolo GAMBA, Gianni LISINI

Department of Electronics, University of Pavia, ITALY�fabio.dellacqua, paolo.gamba, gianni.lisini�@unipv.it

Working Group III/5

KEY WORDS: Urban remote sensing, hyperspectral imaging, very high resolution sensors.

ABSTRACT

In this paper a discussion about the effects of very high resolution (VHR) in both the spatial and the spectral dimensionsfor urban land cover mapping is proposed. We confirm that VHR spatial sensors are unable to discriminate between someurban materials, since they often come from the same chemical family. However, VHR in the spectral sense sometimescarries too much information, since it differentiates covers made by the same material but with different age or illuminationconditions. Finally, after out test with a supervised classifier, we stress the importance to use the context for a moreaccurate mapping, and offer one simple way to improve the classification using a priori knowledge about the geometricconstraints of the segmented map.

1 INTRODUCTION

Land cover mapping (and consequently land use mapping)in urban areas relies, at the level of detail required by mosturban planners, on very high spatial resolution (VHR) sen-sor. Recently, the availability or the development of newairborne and satellite sensors with very high spectral reso-lution allowed the comparison of data sets from these dif-ferent sources for urban material detection, recognition andcharacterization.

Indeed, hyperspectral data have been used for many urbanapplications. Urban material characterization Marino et al.(2001) has been considered by means of the MIVIS sensor,allowing not only to provide a precise definition of the roofmaterials, but also to build a database of dangerous rooftops for subsequent substitutionand destruction. Similarly,vegetation canopies and tree in urban forests provide in-formation extremely interesting for energy exchange anal-ysis, air quality (micro-climate) and hydrology. Their dis-tribution, height, stress and species have been already de-rived from low altitude AVIRIS measurements in Xiao etal. (1999). As a further example, detection of land coverand sealed parts in an urban area can be accomplished us-ing the multi-technique approach proposed in Heiden etal. (2003), where HyMap data have been analyzed for abetter characterization of urban environments. Finally, adetailed analysis of the spectral properties required by sen-sors to obtain interesting results in urban areas is given inHerold et al. (2003), where it was shown that AVIRIS’sfine sampling of the electromagnetic spectrum provides amuch better result than multi-spectral Ikonos (simulated)data, because the latter lack information on significant por-tion of the infrared wavelength range. However, no discus-sion of the spatial resolution issue is provided in the pa-per, since Ikonos is simulated at the same resolution thanAVIRIS.

One final point to be discussed is therefore if the combina-tion of spatial and spectral VHR leads to significantly bet-

ter results than one of the two characteristics taken alone,and to which extent, if any. So, this paper is devoted to thecomparison of two different hyperspectral data sets, com-ing from the ROSIS and DAIS sensors, and a Quickbirdimage.

The two hyperspectral sensors provide the best spatial avail-able up to date (2.6 and 1.2 m posting, respectively), whileQuickbird mounts the finest multi-spectral satellite sensoravailable (2.8 m). The aim of this paper is therefore toimprove the knowledge of the present possibility for ur-ban land cover mapping right now, considering the effectof both the spectral and the spatial properties of the data.Moreover, it may provide useful information for future im-plementation and choice in sensor design targeted to urbanapplications.

2 NEIGHBOURHOOD-AWARE SPECTRAL CLAS-SIFICATION

The classification tool used for comparison is based on thetechnique recently developed in Gamba et al. (2003) formulti-classification fusion of very fine spatial and spectraldata. A simple scheme of the processing steps is as fol-lows:

� a feature extraction procedure is applied to the data inorder to reduce dimensionality;

� supervised classifications are performed, comparingthe results of different spectral and spatial classifiers;

� all the classification maps are combined using major-ity voting or Linear/Logarithmic Opinion Pools;

� a further geometric refinement step is introduced, andspatial information from the neighborhood of eachpixel is taken into account to improve the results.

(a)

(b)

(c)

Figu

re1:

Firs

ttes

tare

a:a

part

ofth

eto

wn

cent

er,a

sim

aged

(fal

seco

lors

)by

the

DA

IS(a

),R

OSI

S(b

)and

Qui

ckbi

rd(c

).

Seve

ral

feat

ure

extr

actio

nap

proa

ches

have

been

prop

osed

for

the

first

step

.O

neof

the

best

know

nis

disc

rim

inan

tan

alys

isfe

atur

eex

trac

tion

(DA

FE),

(Lan

dgre

be,2

003)

,whi

chis

am

etho

dw

hich

isin

tend

edto

enha

nce

sepa

rabi

lity.

InD

AFE

,a

with

in-c

lass

scat

ter

mat

rix,��

,an

da

betw

een-

clas

ssc

atte

rm

atri

x,��

,are

defin

ed:

��

�� �

�������

(1)

���� �

����������

�������

���

(2)

���� �

�������

(3)

whe

re��

isth

em

ean

vect

orfo

rth

e����

clas

s,��

isth

eco

vari

ance

mat

rix

for

the����

clas

s,an

d�����

isth

epr

ior

prob

abili

tyof

the����

clas

s.T

hecr

iteri

onfo

rop

timiz

atio

nm

aybe

defin

edas

��

tr��

��

����

(4)

whe

retr

()

deno

tes

the

trac

eof

am

atri

x.N

ewfe

atur

eve

ctor

sar

ese

lect

edto

max

imiz

eth

ecr

iteri

on.

The

nec-

essa

rytr

ansf

orm

atio

nfr

om�

to�

isfo

und

byta

king

the

eige

nval

ue-e

igen

vect

orde

com

posi

tion

ofth

em

atri

x���

���

and

then

taki

ngth

etr

ansf

orm

atio

nm

atri

xas

the

norm

al-

ized

eige

nvec

tors

corr

espo

ndin

gto

the

eige

nval

ues

inde

-cr

easi

ngor

der.

How

ever

,thi

smet

hod

does

have

som

esh

ort-

com

ings

.Fo

rex

ampl

e,si

nce

disc

rim

inan

tana

lysi

sm

ainl

yut

ilize

scl

ass

mea

ndi

ffer

ence

s,th

efe

atur

eve

ctor

sse

lect

edby

disc

rim

inan

tana

lysi

sar

eno

trel

iabl

eif

mea

nve

ctor

sar

ene

arto

one

anot

her.

Sinc

eth

elu

mpe

dco

vari

ance

mat

rix

isus

edin

the

crite

rion

,di

scri

min

ant

anal

ysis

may

lose

in-

form

atio

nco

ntai

ned

incl

ass

cova

rian

cedi

ffer

ence

s.A

lso,

the

max

imum

rank

of��

is�

��

sinc

e��

isde

pen-

dent

on�

�.

Usu

ally

��

isof

full

rank

and,

ther

efor

e,th

em

axim

umra

nkof

���

���

is�

��.

Thi

sin

dica

tes

that

atm

axim

um�

��

feat

ures

can

beex

trac

ted

byth

isap

proa

ch.

Ano

ther

prob

lem

isth

atth

ecr

iteri

onfu

nctio

nab

ove

gene

rally

does

noth

ave

dire

ctre

latio

nshi

pto

the

er-

ror

prob

abili

ty.

Inth

eea

rly

nine

ties,

deci

sion

boun

dary

feat

ure

extr

actio

n(D

BFE

)was

prop

osed

byL

eean

dL

andg

rebe

(199

7).T

hey

show

edth

atdi

scri

min

antly

info

rmat

ive

feat

ures

and

dis-

crim

inan

tlyre

dund

ant

feat

ures

can

beex

trac

ted

from

the

deci

sion

boun

dary

itsel

f.T

hey

also

show

edth

atdi

scri

mi-

nant

lyin

form

ativ

efe

atur

eve

ctor

sha

vea

com

pone

ntw

hich

isno

rmal

toth

ede

cisi

onbo

unda

ryat

leas

tato

nepo

into

nth

ede

cisi

onbo

unda

ry.

Furt

her,

disc

rim

inan

tlyre

dund

ant

feat

ure

vect

ors

are

orth

ogon

alto

ave

ctor

norm

alto

the

de-

cisi

onbo

unda

ryat

ever

ypo

int

onth

ede

cisi

onbo

unda

ry.

Thu

s,a

deci

sion

boun

dary

feat

ure

mat

rix

(DB

FM)w

asde

-fin

edin

orde

rto

extr

actd

iscr

imin

antly

info

rmat

ive

feat

ures

and

disc

rim

inan

tlyre

dund

ant

feat

ures

from

the

deci

sion

boun

dary

.It

can

besh

own

that

the

rank

ofth

eD

BFM

isth

esm

alle

stdi

men

sion

whe

reth

esa

me

clas

sific

atio

nac

cu-

racy

can

beob

tain

edas

inth

eor

igin

alfe

atur

esp

ace.

Als

o,th

eei

genv

ecto

rsof

the

DB

FM,c

orre

spon

ding

tono

n-ze

roei

genv

alue

s,ar

eth

ene

cess

ary

feat

ure

vect

ors

toac

hiev

eth

esa

me

clas

sific

atio

nac

cura

cyas

inth

eor

igin

alfe

atur

esp

ace.

The

choi

ceto

perf

orm

diff

eren

tcl

assi

ficat

ions

and

com

-bi

neth

eiro

utpu

tsde

pend

son

thei

rdif

fere

ntbe

havi

ors.

Itis

extr

emel

ydi

fficu

ltto

find

rule

sfor

choo

sing

the

best

clas

si-

fier

inan

ysi

tuat

ion.

It’s

bette

rto

com

bine

thei

rmap

sas

inB

riem

etal

.(20

02).

Her

ew

ere

fer

toth

ew

ell-

know

nM

a-jo

rity

votin

gan

dO

pini

onPo

ols

appr

oach

es(B

ened

ikts

son

and

Kan

ello

poul

os,1

999)

.M

ajor

ityvo

ting

isa

very

stan

-da

rdpr

oced

ure,

and

prov

ides

acl

assi

ficat

ion

map

whe

reea

chpi

xel

isas

sign

edto

the

clas

sto

whi

chit

belo

ngs

inth

em

ajor

ityof

the

map

sto

bejo

intly

cons

ider

ed.T

heL

in-

ear

Opi

nion

Pool

(LO

P)an

dth

elo

gari

thm

icop

inio

npo

ol(L

OG

P)pr

oced

ures

requ

ire

inst

ead

atr

aini

ngse

t,w

hich

inhe

rent

lyde

fines

the

outp

utcl

asse

sto

beco

nsid

ered

inth

efin

alcl

assi

ficat

ion

map

.Fo

ran

ypo

ssib

lepa

ttern

ofth

eou

tput

clas

ses

inth

etr

aini

ngse

tw

eco

mpu

te����

,th

epr

obab

ility

that

the

pixe

lsof

the�-

thcl

ass

ofth

etr

aini

ngse

tare

asso

ciat

edw

ithth

e�-

thpa

ttern

inth

em

ultip

lecl

as-

sific

atio

nse

t.Fi

nally

,the

outp

utcl

ass

isde

fined

bym

ax-

imiz

ing����

���

���� ��� �����

or����

���

��

�� �������

�w

here�

isth

ein

putc

lass

ifica

tion

patte

rn,

�th

enu

mbe

rof

clas

sifie

dim

ages

,an

d �

the

wei

ght

of

shadowshadowbricks

bricks

trees

trees

meadow

meadowasphalt

asphalt

(a)(b)

Figure 2: Spectra of some covers in DAIS data (a), and in Quickbird data (b).

the �-th map (default is � � �). Of course, there could besome patterns that are not present in the training set. Forthese patterns, the output class assignment follows major-ity voting.

The combined maps come from different classifiers, somebased on spectral information (Maximum Likelihood, FisherLinear Discriminant, Fuzzy ARTMAP), others on spatialinformation also (ECHO, Fuzzy ARTMAP with spatial re-classification). The last classifier is a refined version of thefuzzy ARTMAP classification chain introduced in Gambaand Houshmand (2001); Gamba and Dell’Acqua (2003).The output of a pixel-by-pixel classification is used by asecond classifier whose input is a vector representing themapping class percentages in a window around each pixelplus a random value to enhance their discriminability.

After multiple map combination, we applied a further re-classification step, again by means of the fuzzy ARTMAPclassifier, which usually provide better and more preciseresults, since it corrects random errors due to randomlycorrelated noise, i. e. errors, in the maps.

Even if the methodology has been fully presented in previ-ous papers, the usefulness is this research is in comparingthe best results on the same test area and the same test andtraining sets for different data sets. An accurate test on thespectral and spatial properties of the training sample mayhelp not only to understand the different performances, butalso to extrapolate, beyond the particular class legend weused, which may be the best choice for urban land covermapping with these sensors.

Moreover, in this paper we introduce a final step after theabove mentioned classification chain, aimed at a furthergeometric refinement of the classification. This step isbased on the usual methodology applied to panchromaticVHR images, and based on the experience gained withphotogrammetric data processing. Indeed, as correctly statedin recent review papers, (see Guindon (2001) and Gamba et

al. (2003), for instance), that it is time now to start consid-ering that the photogrammetric-based and the classification-based views should be integrated into a general framework.One possible idea, which is somehow a generalization ofJin and Davis (2004), is to match a top-down segmentationstep with a bottom up approach based on geometric fea-tures and cues grouping. In the top-down step, an initialpartitioning of areas of interest within a scene is accom-plished using simple supervised classification schemes. Areduced resolution image may be considered, and texturalfeatures may be used to complement the original informa-tion. In the bottom-up step, each part of the scene andits surroundings are searched for basic patterns of covers,and corrections to misclassifications are made based on thecontext. Alternatively, geometric features are extracted torecognize objects instead of classification segments, andthe class of each object is chosen by majority voting.

Following these line guides, in this work we apply a post-processing step to the classified land cover map, aimed atexploiting some of the constraints for urban covers. Forinstance, we may assume that roof cover pixel belong togeometric structures delimited by orthogonal edges. Thisassumptions leads to a correction procedure that searchesfor segmented regions belonging to a roof cover class andextracts their boundary. Successively, these boundaries arematched with a piecewise linear path with 90Æ angles. Fi-nally, the corresponding regions are re-drawn on the origi-nal map.

3 EXPERIMENTAL RESULTS

We offer results about the town of Pavia, Northern Italy.Four DAIS and ROSIS flight lines over the area, kindlyprovided by the German Space Agency in the frameworkof the HySens project, cover the town and its immediatesurroundings. The Digital Airborne Imaging Spectrometer(DAIS) is a multi-band systems with 80 bands in the visi-ble and infrared wavelength range (from 0.4 to 12.6 �m),

(a)

(b)

(c)

Figu

re3:

Seco

ndte

star

ea:

the

Eng

inee

ring

Scho

ol:

clas

sific

atio

nm

aps

obta

ined

usin

gth

eD

AIS

(a),

RO

SIS

(b)

and

Qui

ckbi

rd(c

)da

ta.

whi

leth

eR

eflec

tive

Opt

ics

Syst

emIm

agin

gSp

ectr

omet

er(R

OSI

S)is

am

ulti-

band

sens

orfo

cuse

don

lyon

visi

ble

and

near

infr

ared

freq

uenc

yba

nds.

The

latte

rcom

pris

es32

fre-

quen

cyba

nds

from

0.45

to0.

85�

m,s

how

ing

high

erpo

ten-

tialf

orve

geta

tion

and

gree

nar

eam

appi

ngth

anth

efo

rmer

one,

whi

ch,

onth

eco

ntra

ry,

has

abr

oade

rra

nge

ofap

-pl

icat

ions

,sin

ceit

prov

ides

info

rmat

ion

upto

the

ther

mal

infr

ared

.To

thes

eda

tase

tsw

ead

ded

elab

orat

ions

mad

eon

aQ

uick

bird

panc

hrom

atic

and

mul

ti-sp

ectr

alim

ages

(4ba

nds,

visi

ble

and

near

infr

ared

),ac

quir

edin

the

sam

epe

-ri

od(J

uly

2002

).

The

fine

spec

tral

reso

lutio

nof

som

eof

thes

eda

tase

tsco

u-pl

esw

ell

with

ave

ryhi

ghsp

atia

lre

solu

tion,

from

2.6

mfo

rD

AIS

to1.

2m

for

RO

SIS

to0.

7m

for

pan-

shar

pene

dQ

uick

bird

data

.

Inth

eur

ban

test

area

ofin

tere

st,t

hefo

llow

ing

resu

ltssh

owan

dco

mpa

real

gori

thm

perf

orm

ance

son

two

diff

eren

ttes

tar

eas,

one

inth

ece

nter

ofth

eto

wn,

cons

ider

ing

typi

calu

r-ba

nla

ndco

vers

,an

don

ein

the

surr

ound

ings

,whe

rebo

thur

ban

and

rura

lco

vers

exis

t.T

his

choi

ceal

low

san

alyz

-in

gth

eov

eral

lan

dcl

ass

accu

racy

valu

esw

ithre

spec

tto

the

spec

tral

and

spat

ialp

rope

rtie

sof

diff

eren

tleg

ends

,and

also

tofo

cus

onth

eef

fect

sdu

eto

typi

calp

robl

emsi

nde

nse

urba

nar

ea,

such

asm

aski

ngby

shad

ows.

The

first

area

isde

pict

edin

fig.1

:pl

ease

note

that

the

area

sar

eno

texa

ctly

the

sam

e,du

eto

diff

eren

ces

inth

efig

htlin

es.

For

exam

-pl

e,du

eto

the

finer

spat

ial

reso

lutio

nan

dth

esu

bseq

uent

redu

ced

swat

hof

the

RO

SIS

sens

or,

only

two

part

sof

the

scen

ear

esh

own

inth

eR

OSI

Sfa

lse

colo

rim

age,

whi

chbe

long

totw

ono

n-ov

erla

ppin

glin

esof

fligh

t.

All

the

supe

rvis

edcl

assi

fiers

used

expl

oitn

on-o

verl

appi

ngtr

aini

ngan

da

test

sets

com

pose

dby

near

ly10

0pi

xels

for

each

clas

s,an

dre

ferr

ing

to7

clas

ses

ofur

ban

cove

rs,

plus

wat

er.

Cov

ers

incl

ude

tree

s,m

eado

ws

and

bare

soil,

as-

phal

tan

dbi

tum

en,

park

ing

spec

ial

cove

ran

dbr

ick

roof

.To

disc

ard

prob

lem

sw

ithsh

adow

edar

eas,

asp

ecia

lcl

ass

for

shad

owpi

xels

was

adde

d.

The

high

esta

chie

vabl

eva

lues

fort

heov

eral

lacc

urac

yam

ong

allt

hete

sts

mad

ear

esh

own

inTa

ble

1.N

ote

that

thes

eva

l-ue

sdo

nott

ake

into

acco

untt

hefin

alge

omet

ric

refin

emen

t,si

nce

this

may

intr

oduc

eso

me

bias

inth

ere

sults

due

toits

depe

nden

ceon

spat

ial

reso

lutio

non

ly.

Thi

sis

the

reas

onw

hyit

will

bedi

scus

sed

sepa

rate

ly,

atth

een

dof

this

sec-

tion.

The

accu

racy

valu

esin

the

Tabl

esh

owth

atfo

rthi

sfir

stte

star

ea,

but

also

glob

ally

spea

king

,sp

atia

lre

-cla

ssifi

catio

nus

ing

AR

TM

AP

neur

alne

twor

ksis

byfa

rth

ebe

stop

tion.

Mor

eove

r,th

ese

valu

esst

ress

the

fact

that

high

spat

ialr

es-

olut

ion

isno

ten

ough

,an

dlo

wer

spat

ial

post

ing

isw

ell-

bala

nced

bym

ulti-

spec

tral

capa

bilit

y.A

sa

mat

ter

offa

ct,

DA

ISm

aps

are

alw

ays

bette

rth

anQ

uick

bird

ones

.M

ore-

over

,it

shou

ldbe

men

tione

dth

atin

Qui

ckbi

rdm

aps

som

eof

the

cove

rsar

eno

tac

tual

lypr

esen

t.Fo

rin

stan

ce,

bitu

-m

enan

das

phal

t,be

ing

virt

ually

indi

stin

guis

habl

edu

eto

the

low

spec

tral

reso

lutio

n,ar

ejo

intly

cons

ider

ed.

Fina

lly,

RO

SIS

map

ping

accu

racy

isba

sed

ofco

urse

onbo

thth

efin

esp

atia

land

spec

tral

reso

lutio

n.

Togi

vean

idea

ofth

edi

ffer

entp

robl

ems

com

ing

from

spa-

tiala

ndsp

ectr

alhi

ghre

solu

tion,

we

offe

rin

fig.2

the

spec

-tr

aof

five

land

cove

rsar

ound

the

area

ofth

eca

stle

ofPa

via,

asth

eyca

nbe

dedu

ced

byth

eQ

uick

bird

and

DA

ISda

tase

ts.

The

two

subs

ets

show

also

the

diff

eren

cein

spat

ial

deta

ils,a

ndth

usal

low

com

pari

ngth

egl

obal

effe

ctof

both

reso

lutio

nson

the

map

ping

capa

bilit

ies

ofth

etw

ose

nsor

s.

As

for

the

seco

ndte

star

ea,

the

Eng

inee

ring

Scho

olof

the

Uni

vers

ityof

Pavi

a,th

ela

ndco

ver

lege

ndus

edw

aspa

r-tia

llydi

ffer

ent.

We

cons

ider

ed9

cove

rs,

i.e.

tree

s,m

ead-

ows

and

bare

soil,

asph

alt

and

bitu

men

,gr

avel

,pa

rkin

gsp

ecia

lco

ver

and

met

allic

roof

s.E

ven

inth

isca

seth

elo

wer

spec

tral

reso

lutio

nof

Qui

ckbi

rdda

tadi

dno

tal

low

disc

rim

inat

ing

bitu

men

agai

nsta

spha

lt,an

dth

etw

ocl

asse

sw

ere

join

tlyco

nsid

ered

.A

gain

,ash

adow

clas

sw

asar

tifi-

cial

lyin

trod

uced

,to

redu

cem

iscl

assi

fied

pixe

ls.

The

clas

-si

ficat

ion

map

sfo

rth

isar

eas

are

show

nin

fig.3

.

Overall accuracy (%) DAIS ROSIS Quickbird

Center 97.5 99.3 91.8Engineering School 87.5 91.4 80

Class. algorithm DAIS ROSIS Quickbird

Center ARTMAP (DAFE) Spatial ARTMAP Spatial ARTMAPEngineering School Majority voting Spatial ARTMAP Spatial ARTMAP

Table 1: Comparison of the best overall classification accuracy values and classification methods for land cover mapsobtained from data by each of the three sensors and referring to the same test areas.

A look at the maps reveals that there are some misclassifi-cations and pixels assigned to the wrong (even if spectrallysimilar) class. We have, for instance, some exchanges be-tween pixels in the bitumen and gravel classes. As a mat-ter of fact, the gravel on top of some of the buildings isposed on a bitumen substrate, so that the spectral responseis somehow mixed. Similarly, bitumen and asphalt are alsoexchanged in some areas. ROSIS data show a better per-formance, but suffer from a stronger dependency on thetraining set. Pure pixels are mandatory and a few mixedspectral response may change dramatically the per-classaccuracy values.

3.1 Geometric refinement of the classified maps

As introduced above, a final step was applied to some ofthe classes. The procedure is based on a boundary extrac-tion for each of the segments of the class of interest, fol-lowed by a piece-wise linear reconstruction of this bound-ary and, finally, a step that poses constraints on the anglebetween subsequent segments. There are therefore threeparameters ruling the procedure: the maximum length ofthe segments and the maximum distance from the originalboundary determine how approximate is its piece-wise lin-ear representation. The tolerance on the angle to be “cor-rected” is used to choose which are the segments whosedirection is to be re-assigned in a clock-wise path alongthe piece-wise linear approximation.

An example of block refinement through geometric con-straints is shown in fig. 4(a). Similarly, the segments ofthe brick roof class in parts of the Quickbirdclassificationmap of the whole town, before and after the geometric cor-rection, are showed in fig. 4(b) and (c), respectively. Wemay observe an improvement in the shape of the objects,that become more recognizable. However, the shapes arenot as regular as expected, as a consequence of the impre-cise piece-wise linear representation of their boundaries.This in turn is due to the small number of pixels definingthe boundaries. We may observe more visible improve-ments when considering, in another part of the town, nearthe Engineering school, the same approach as applied tothe “road” (actually, asphalt) class, as shown in fig. 4(d)and (e). In this example the advantage of using geometri-cal constraints on the shape of the objects is more evident,and results in a regularization of their boundaries.

4 CONCLUSIONS

The present work shows that, with respect to urban andurban/rural land cover mapping, very high resolution in thespectral sense is more valuable than in the spatial sense. Itmay be better to have more bands recorded by a sensor thata more detailed image of the scene.

The main reason for this behavior is the similarity of manycovers due to their very similar chemical components. In-deed, once a sufficient spatial resolution is achieved, sothat in the image there are mostly spectrally pure pixels,the geometric properties of the objects may be improvedexploiting a priori information about their shape. This in-formation is, in fact, valuable but already available for ur-ban areas and objects herein.

Instead, the correctness of assignment of an object to theright land cover class (and possibly the right land use class)is based on the possibility to recognize the spectra of thematerials, which in turn depends on the number and posi-tions of the imaged wavelengths.

ACKNOWLEDGMENTS

The authors want to thank Francesco Grassi for his help inperforming some of the classification tests. The classifica-tions using Quickbird data were performed at the Univer-sity of Siena. Thanks to Andrea Garzelli for running thetests and providing the mapping results.

References

C.M. Marino, C. Panigada, L. Busetto: “Airborne hyper-spectral remote sensing applications in urban areas: as-bestos concrete sheeting identification and mapping”,Proc. of the 1st IEEE/ISPRS Joint Workshop on RemoteSensing and Data Fusion over Urban Areas, Rome, Italy,8-9 Nov. 2001, pp. 212-216.

Q. Xiao, S. L. Austin, E. G. McPherson, and P. J. Peper:“Characterization of the structure and species composi-tion of urban trees using high resolution AVIRIS data”,Proc. of the 1999 AVIRIS Workshop, Pasadena, CA,1999, unpaginated CD-ROM.

U. Heiden,K. Segl, S. Roessner, H. Kaufmann: “Ecologi-cal evaluation of urban biotope types using airborne hy-perspectral HyMap data”, Proc. of the 2nd IEEE/ISPRS

(a)

(d)

(e)

(b)

(c)

Figu

re4:

(a)

anex

ampl

eof

geom

etri

cre

finem

ento

facl

assi

fied

segm

ento

fth

esc

ene;

the

sam

eap

proa

chis

appl

ied

toth

ese

gmen

tsin

(b)

toob

tain

(c),

and

in(d

)to

obta

in(e

).

Join

tW

orks

hop

onR

emot

eSe

nsin

gan

dD

ata

Fusi

onov

erU

rban

Are

as,

Ber

lin,

Ger

man

y,22

-23

May

2003

,pp

.18-

22.

M.

Her

old,

M.E

.G

ardn

er,

D.A

.R

ober

ts:

“Spe

ctra

lre

s-ol

utio

nre

quir

emen

tsfo

rm

appi

ngur

ban

area

s”,

IEE

ETr

ans.

Geo

sc.R

emot

eSe

ns.V

ol.4

1,n.

9,pp

.190

7-19

19,

Sept

.200

3.

P.G

amba

,F.

Del

l’A

cqua

,A

.Fe

rrar

i:“E

xplo

iting

spec

-tr

alan

dsp

atia

linf

orm

atio

nfo

rcla

ssif

ying

hype

rspe

ctra

lda

tain

urba

nar

eas”

,Pr

oc.

ofIG

AR

SS’0

3,Ju

ly20

03,

Vol

.I,p

p.46

4-46

6.

D.A

.Lan

dgre

be,S

igna

lThe

ory

Met

hods

inM

ulti

spec

tral

Rem

ote

Sens

ing,

John

Wile

yan

dSo

ns,

Hob

oken

,N

ewJe

rsey

,20

03.

C.

Lee

and

D.

A.

Lan

dgre

be,

“Dec

isio

nB

ound

ary

Fea-

ture

Ext

ract

ion

for

Neu

ral

Net

wor

ks,

”IE

EE

Tran

s.on

Geo

scie

nce

and

Rem

ote

Sens

ing,

Vol

.8,n

.1,p

p.75

-83,

1997

.

G.J

.B

riem

,J.

A.

Ben

edik

tsso

n,an

dJ.

R.

Svei

nsso

n,“M

ul-

tiple

clas

sifie

rsap

plie

dto

mul

tisou

rce

rem

ote

sens

ing

data

”IE

EE

Tran

s.on

Geo

scie

nce

and

Rem

ote

Sens

ing,

Vol

.40,

n.10

,pp.

2291

-229

9,O

ct.2

002.

J.A

.Ben

edik

tsso

nan

dI.

Kan

ello

poul

os:“

Cla

ssifi

catio

nof

mul

tisou

rce

and

hype

rspe

ctra

ldat

aba

sed

onde

cisi

onfu

-si

on,”

IEE

ETr

ans.

onG

eosc

ienc

ean

dR

emot

eSe

nsin

g,V

ol.3

7,n.

3,pp

.136

7-13

77.

P.G

amba

,B

.H

oush

man

d:“A

nef

ficie

ntne

ural

clas

sifi-

catio

nch

ain

for

optic

alan

dSA

Rur

ban

imag

es”,

Inte

r-na

tion

alJo

urna

lof

Rem

ote

Sens

ing,

Vol

.22

,n.

8,pp

.15

35-1

553,

May

2001

.

P.G

amba

,F.

Del

l’A

cqua

:“I

mpr

oved

mul

tiban

dur

ban

clas

sific

atio

nus

ing

ane

uro-

fuzz

ycl

assi

fier”

,In

tern

a-ti

onal

Jour

nalo

fRem

ote

Sens

ing,

Vol

.24,

n.4,

pp.8

27-

834,

Feb.

2003

.

B.G

uind

on,“

Com

pute

r-ba

sed

aeri

alim

age

unde

rsta

ndin

g:a

revi

ewan

das

sess

men

tofi

tsap

plic

atio

nto

plan

imet

ric

info

rmat

ion

extr

actio

nfr

omve

ryhi

ghre

solu

tion

sate

l-lit

eim

ages

”,C

anad

ian

Jour

nalo

fR

emot

eSe

nsin

g,V

ol.

23,N

o.1.

P.G

amba

,O

.H

ellw

ich,

P.L

omba

rdo:

Edi

tori

alto

the

The

me

Issu

eon

“Alg

orith

ms

and

Tech

niqu

esfo

rM

ulti-

Sour

ceD

ata

Fusi

onin

Urb

anA

reas

”,IS

PRS

Jour

nalo

fPh

otog

ram

met

ryan

dR

emot

eSe

nsin

g,V

ol.

58,

n.1-

2,pp

.1-3

,May

2003

.

Jin,

X.

and

C.H

.D

avis

,“A

utom

ated

build

ing

extr

actio

nfr

omhi

gh-r

esol

utio

nsa

telli

teim

ager

yin

urba

nar

eas

usin

gst

ruct

ural

,co

ntex

tual

,an

dsp

ectr

alin

form

atio

n”,

EU

RA

SIP

Jour

nal

onA

pplie

dSi

gnal

Proc

essi

ng,

Spe-

cial

Issu

eon

Adv

ance

sin

Inte

llige

ntV

isio

nSy

stem

s:M

etho

dsan

dA

pplic

atio

ns,2

004.