Pa e n Recogni ion in AVHRR Images by Means o
Hib yd and Neu o- uzzy Sys ems
Jose An onio Pied a, F ancisco Guindos, Albe o Molina, and Manuel Can on
1 Uni e sidad de Alme ia, Depa men o de Languajes y Compu ación,
Phone +34 950 0140 36, 04120 Alme ia, Spain
{jpied a, guindos, mcan on}@ual.es
[email p o ec ed]
Abs ac . The main goal o his wo k is o imp o e he au oma ic in e p e a ion o
ocean sa elli e images. We p esen a compa a i e s udy o di e en classi ie s:
G aphic Expe Sys em (GES), ANN-based Symbolic P ocessing Elemen (SPE),
Hyb id Sys em (ANN – Radial Base Func ion & Fuzzy Sys em), Neu o-Fuzzy
Sys em and Bayesian Ne wo k.. We wish o show he u ili y o hyb id and neu o-
uzzy sys em in econgni ion o oceanic s uc u es. On he o he hand, o he ob-
jec i e is he ea u e selec ion, which is conside ed a undamen al s ep o pa e n
ecogni ion. This pape epo s a s udy o lea ning Bayesian Ne wo k o ea u e
selec ion [1] in he ecogni ion o oceanic s uc u es in sa elli e images.
1 S uc u e o he Au oma ic In e p e a ion Sys em
Fig. 1 depic s he o e all s uc u e o he sys em ha has been de eloped. In a i s
s ep, he aw image is p ocessed by means o algo i hms such as adiome ic co ec-
ion, map p ojec ion and land masking. These a e well known echniques also used
when he analysis is made by human expe s. Howe e , we don’ make any image
enhancemen like his og am equaliza ion o con as s e ching ha a e app op ia e o
make some ea u es isible o human eye bu ha e no posi i e e ec s when he im-
ages a e going o be p ocessed by digi al sys ems.
The second s ep aims o de ec clouds pixels ha a e opaque o adiance da a
measu ed in he AVHRR in a ed and isible scenes. Cloudy images a e dis o ed in
such a way ha he zone a ec ed isn’ o any alue o ou la e p ocessing, so we
build a mask o 0s ha will exclude hese pixels [2].
The ollowing ask is he segmen a ion ha will di ide he whole image in egions.
The idea is ha each phenomenon o in e es should coincide wi h one o a small se
o he segmen ed egions. The na u e o ocean dynamics makes e y di icul his
p ocess ha is ne e heless undamen al, so we’ e designed an i e a i e knowledge-
d i en me hod o pe o m his pa o he p ocess pipeline [3].
The nex ask is he ea u es o desc ip o s selec ion, which consis s o selec ing an
op imal o sub-op imal ea u e subse om a se o candida e ea u es. The mos
common amewo k o ea u es selec ion is o de ine some c i e ia o measu ing he
goodness o a se o ea u es, and hen use a sea ch algo i hm o ind an op imal o
sub-op imal se o ea u es [4]. We ha e used Bayesian ne wo ks o ea u es selec-
ion in he ecogni ion o oceanic s uc u es in sa elli e images [5].
R. Mo eno Díaz e al. (Eds.): EUROCAST 2005, LNCS 3643, pp. 373 – 378, 2005.
374 J.A. Pied a e al.
Fig. 1. S uc u e o he ocean ea u e ecogni ion sys em
In he las s ep, each egion p oduced in he segmen a ion is analyzed and, i he
ecogni ion is posi i e, i is labeled wi h he iden i ie o he ma ching s uc u e. The
s uc u es o in e es in he Cana y Islands zone as de ined in a e: coas al upwell-
ing, wa m eddies, cold eddies and island wakes. The as majo i y o he egions ha
appea in he segmen a ion a e o no special in e es and hey a e labeled wi h a 0.
We ha e implemen ed a edundan ecogni ion subsys em. I has an ANN-based
Symbolic P ocessing Elemen (SPE) module [6], a ule-based G aphic E.S. (GES) [3],
Bayesian Ne wo k, Hyb id Sys em (A i icial Neu al Ne wo k based Radial Base
Func ion and Fuzzy Sys em based Sugeno) and Neu o-Fuzzy Sys ems (NEFPROX,
ANFIS) pe o ming he same ask. The pu pose is o es di e en me hodologies and
o p o ide a way o alida e and compa e hese esul s.
2 De ailed P ocess
2.1 Fea u e Selec ion by Bayesian Ne wo k
The mos common amewo k o ea u e selec ion is o de ine c i e ia o measu ing
he goodness o a se o ea u es, and hen use a sea ch algo i hm ha inds an op imal
o sub-op imal se o ea u es. Ou goal in his s ep is o apply he heo y o lea ning
Bayesian Ne wo k o he educ ion o i ele an ea u es [1] in he ecogni ion o
oceanic s uc u es in sa elli e images [6].
The expe imen was done o e he ea u e se ob ained in he wo k [6]. The lea n-
ing algo i hms (K2, VNSST, Nai e-Bayes) ha e been e alua ed wi h di e en con-
igu a ions o pa ame e s o selec he bes con igu a ion. Table 1 shows he summa y
Pa e n Recogni ion in AVHRR Images by Means o Hib yd and Neu o- uzzy Sys ems 375
o compa a i e esul s gene a ed om SPE (Neu al Ne wo k) and Lea ning Bayesian
Me hod. This me hod allows choosing a ea u e subse ob aining he same accu acy
a e (abou 80%) ha wi h he ini ial ea u e se .
Table 1. Compa a i e esul s o ea u e selec ion
Algo i hms Numbe o Fea u es
Symbolic P ocessing Elemen 50
K2 Lea ning 15
VNSST Lea ning 15
Nai e-Bayes Lea ning 50
2.2 Classi ica ion
All classi ie s has been designed and es ed wi h he same image da a se .
2.2.1 Hyb id Sys em (A i icial Neu al Ne wo k and Fuzzy Sys em)
Fuzzy modeling is one o he echniques cu en ly being used o modeling nonlinea ,
unce ain, and complex sys ems. An impo an cha ac e is ic o uzzy models is he
pa i ioning o he space o sys em a iables in o uzzy egions using uzzy se s [7].
One o he aspec s ha dis inguish uzzy modeling om o he black-box app oaches
like neu al ne s is ha uzzy models a e anspa en o in e p e a ion and analysis ( o a
ce ain deg ee).
Radial basis unc ion ne wo ks and uzzy ule sys ems a e unc ionally equi alen
unde some condi ions. The e o e, he lea ning algo i hms de eloped in he ield o
a i icial neu al ne wo ks can be used o adap he pa ame e s o uzzy sys ems. We
use he ela ionship be ween RBF neu al ne wo k and uzzy sys em o classi ica ion
pu pose [8].
The ope a ion begins aining a neu onal ne wo k unde ini ial condi ions (equi a-
lence wi h uzzy sys em). Neu al ne wo k ob ains a se o ules, ha a e used o build
he uzzy sys em. Bo h sys ems a e used in classi ica ion p ocesses.
2.2.2 Neu o- uzzy Sys em
Neu o-Fuzzy Sys em e e s o he combina ion o uzzy se heo y and neu al ne -
wo ks wi h he ad an ages o bo h:
1. Manage imp ecise, pa ial, ague o impe ec in o ma ion
2. Handle any kind o in o ma ion (nume ic, linguis ic, logical, e c.)
3. Sel -lea ning, sel -o ganizing and sel - uning capabili ies
4. No need o p io knowledge o ela ionships o da a
5. Reduce human decision making p ocess
6. Fas compu a ion using uzzy numbe ope a ions
We use ANFIS [9] (Adap a i e Ne wo k based Fuzzy In e ence Sys em), which is
a uzzy in e ence sys em implemen ed in he amewo k o adap a i e ne wo k.
ANFIS can se e as a base o cons uc ing a se o uzzy i - hen ules wi h app opia e
membe ship unc ions o gene a e he s ipula ed inpu -ou pu pai s.
376 J.A. Pied a e al.
On he o he hand, we use a gene al app oach o unc ion app oxima ion by a
neu o- uzzy model based on supe ised lea ning. NEFPROX [10] (NEu oFuzzy unc-
ion apPROXima o ) has a simila s uc u e as he NEFCON model, bu i is an ex en-
sion, because i does no need ein o cemen lea ning. On he o he hand i also ex-
ends he NEFCLASS model [11], ha can only be used o c isp classi ica ion asks.
2.2.3 Bayesian Ne wo ks
Bayesian Ne wo ks p o ide an in ui i e g aphical isualiza ion o he knowledge
including he in e ac ions among he a ious sou ces o unce ain y. Bayesian Ne -
wo ks a e models o ep esen ing unce ain y in ou knowledge. We use hem in
ea u e selec ion and classi ica ion.
3 Resul s
The in o ma ion p o ided by knowledge d i en classi ie s a e e u bished wi h he
o iginal segmen a ion o c ea e images (Fig. 2) ha show in a isual way each he la-
beled ocean ea u e o in e es . Each ea u e is ep esen ed by di e en colo ed egion
in a map whe e each colo ep esen s one o he ocean phenomena we a e looking o .
Fig. 2. AVHRR scene (equalized) and ea u e map (o ange:upwelling, ed-g een-blue: wa m
wakes,ligh blue: wa m gy e)
Resul s om sys ems SPEs,GES, HS, BN and NFS, depend mainly on he quali y
o he images. In gene al, a good numbe o aining images a e needed. Fo GES, he
Table 2. Compa a i e esul s o classi ica ion
Classi ie Accu acy Ra e
Hyb id Sys em 60 %
NEFPROX 60 %
ANFIS 60 %
Bayesian Ne wo ks
(K2,VNSST Lea ning)
80 %
S.P.E. 80 %
G.E.S. 95 %
Pa e n Recogni ion in AVHRR Images by Means o Hib yd and Neu o- uzzy Sys ems 377
bes esul s a e achie ed when he human expe p o ides he sys em wi h speci ic
knowledge abou he a ge a ea. When hese equisi es a e me , he sys ems p oduce
posi i e ocean s uc u es ecogni ion, which is shown in able 2.
4 Conclusions
We ha e explo ed he use o Bayesian ne wo ks as a mechanism o ea u e selec ion
in a sys em o au oma ic ecogni ion o ocean sa elli e images. The use o Bayesian
ne wo ks has p o ided bene i s wi h espec o SPE, no only in he educ ion o ele-
an ea u es, bu also in disco e ing he s uc u e o he knowledge, in e ms o he
condi ional independence ela ions among he a iables. In u u e wo ks we plan o
imp o e he accu acy a e o he sys em including mo e a iables. Fu he mo e, we
expec o use models o a oid he disc e isa ion o he con inuous ea u es when lea n-
ing Bayesian ne wo ks.
On he o he hand, ob ained esul s a en´ expec ed o classi ica ion p ocess by
hyb id and neu o- uzzy sys em. They don´ co espond wi h esul s o o he cassi ie s
(EPS, GES and Bayesian Ne wo k). The main p oblem is he numbe o gene a ed
ules, which is excessi e. In u u e wo ks we plan o imp o e he accu acy a e o he
hyb id and neu o- uzzy sys ems including mo e a iables, unning aining pa ame-
e s, op imizing he ob ained s uc u e and au oma ing he cons uc ion o neu o uzzy
sys em.
Finally, he s uc u e o he au oma ic in e p e a ion sys em ha has been in o-
duced, i is a complex and independen s uc u e o he ollowing asks:
1. I e a i e segmen a ion- ecogni ion cycle is made by means o g aphic expe sys-
em.
2. Fea u e selec ion and alida ion o knowledge is done by bayesian lea ning
3. Mul iple classi ica ion allows o ind di e en in e p e a ions o he exis ing
knowledge o alida e he knowledge o he classi ie s.
Acknowledgmen s
This wo k is being pa ially suppo ed by he Spanish Commission o Science and
Technology (CICYT) h ough he p ojec 2004-TIN-05346.
Re e ences
1. Inza I., La añaga P., E xebe ia R. and Sie a B.: Fea u e Subse Selec ion by Bayesian
ne wo ks based op imiza ion. A i icial In elligence, nº 123, (2000) 157–184.
2. To es J.A., Guindos F., Pe al a M. and Can ón M.: An Au oma ic Cloud-Masking Sys em
Using Backp o Neu al Ne s o AVHRR Scenes, IEEE T ansac ions On Geoscience and
Remo e Sensing, ol. 41, nº 4, (2003) 826–831.
3. Guindos F., Pied a J.A. and Can ón M.: Ocean ea u es ecogni ion in AVHRR images by
means o bayesian ne and expe sys em, 3 d In e na ional Wo kshop on Pa e n Recogni-
ion in Remo e Sensing, Kings on Uni e si y, Uni ed Kingdom. Augus (2004).

378 J.A. Pied a e al.
4. Langley P. and Sage S.: Induc ion o selec i e Bayesian classi ie s. In P oceedings o he
Ten h Con e ence on Unce ain y in A i icial In elligence, Sea le. Mo gan-Kau mann,
(1994) 399–406.
5. Yamaga a Y. and Oguma H.: Bayesian ea u e selec ion o classi ying mul i- empo al
SAR and TM Da a. IEEE, (1997) 978–980.
6. To es J.A., Guindos F., López M. and Can on M.: Compe i i e neu al-ne -based sys em
o he au oma ic de ec ion o oceanic mesoscala s uc u es on AVHRR scenes. IEEE
T ans. on Geoscience and Remo e Sensing, ol.41, issue 4, Ap il (2003) 845–852.
7. Zadeh L.A.: Fuzzy Logic. IEEE Compu e , ol. 21 , issue 4 , Ap il (1988) 83–93.
8. Yaochu Jin, We ne on Seelen, Be nha d Sendho : Ex ac ing In e p e able Fuzzy Rules
om RBF Neu al Ne wo ks. Ins i u ü Neu oin o ma ik, Ruh -Uni e si ä Bochum,
FRG. In e nal Repo (2000–2002).
9. Roge Jang: ANFIS Adap a i e Ne wo k based Fuzzy In e ence Sys em. IEEE T ansac-
ions On Sys ems, Man and Cybe ne ics, ol 23, nº 3. May/June (1993).
10. Nauck D. and K use R.: A neu o- uzzy app oach o ob ain in e p e able uzzy sys ems o
unc ion app oxima ion. Fuzzy Sys ems P oceedings, IEEE Wo ld Cong ess on Compu a-
ional In elligence., ol 2 , May (1998) 1106 –1111.
11. Nauck D.: Knowledge disco e y wi h NEFCLASS. Fou h In e na ional Con e ence on
Knowledge-Based In elligen Enginee ing Sys ems and Allied Technologies, ol. 1 , 30
Augus -1 Sep embe (2000) 158–161.