AUTOMATED TECHNIQUES FOR SATELLITE IMAGE SEGMENTATION

Transcription

1 IPR IPT IGU UCI CIG ACG Table of conens Table des maières Auhors index Index des aueurs earch Recherches xi orir AUTOMATD TCHIQU FOR ATLLIT IMAG GMTATIO A. Guarnieri*, A. Veore* *CIRGO (Inerdeparmen Research Cener for Geomaics via Romea Legnaro (Padua Ialy Phone: ; Fax: Commission IV, orking Group IV/7 Y ORD: saellie image classificaion, image segmenaion, clusering ABTRACT In his paper a combined mehod beween classical and auomaic approach for remoe sensing image analysis is presened. Typicall saellie images are used in order o deec he disribuion of vegeaion, soil classes, buil-up areas, roads, and waer body as rivers, brooks, lakes, ecc. Referring for example o Landsa-TM images, he idenificaion of such aspecs is performed hrough he classical approach of image classificaion. Basicall i deals wih he use of pseudocolors and/or combinaions of various specral bands o acquire differen hemaic layers from he images. In his work a furher processing sep is inroduced, namely a segmenaion algorihm is applied o color images in order o improve he image analysis boh from qualiaive and quaniaive poin of view. This algorihm belongs o he class of operaions performed in he auomaized unsupervised analysis of color images. Las recen advances in he field of compuer science and CPU performance have lead o a grea reducion of daa processing imes, allowing herefore o apply also in he field of Remoe ensing more complex algorihms. The proposed segmenaion algorihm is based on a feaure-space approach and implemens wo processing echniques: he hisogram hresholding [1 and he clusering [2. ome ineresing resuls applied o Remoe ensing images will be provided. 1. ITRODUCTIO o far, ipical approach o remoe sensed daa analysis has been based on he use of differen linear combinaions beween available specral bands or creaing new ones. In his phase saellie daa are processed hrough he aciviy of a human operaor, which using RGB filers ries o idenify several classes of elemens appearing on hem. These classes are herefore a way o group homogeneus land feaures, such as urban areas (roads, buildings, ec., vegeaion (woods, culivaed and unculivaed soils and waer areas (rivers, lakes, ec.. These layers are ofen represened in erms of pseudocolor images, in order o beer highligh a specific feaure disribued along he land. The basic concep underlying ha procedure is similar o he image segmenaion, he firs sep for he digial image processing normally adoped in he field of Compuer Vision. This operaion is performed hrough he pariioning of an image us in homogeneous and separaed regions. o far, segmenaion echniques were applied only o gray scale images, hough he color informaion would allow a more complee image represenaion. In fac, he applicaion of his mehod was limied mainly by he compuaional ime spen for color daa processing, larger han he one needed for gray color daa. Toda recen advances in he field of compuer science and CPU performance, have lead o a grea reducion of daa processing imes, allowing herefore o apply hese segmenaion algorihms o he field of Remoe ensing, as well. everal color image segmenaion algorihms are nohing bu he developmen of previous gray color procedures, ohers are insead new ad-hoc echniques for color daa, which ake ino accoun he physic relaionship beween ligh and coloured maerials. uch algorihms work in well defined color spaces, such as RGB, HI or HV. Anyway hese reference frames are no uniform, i.e. color differences of same eni as perceived or measured by human eye, are no convered in similar disances among he poins represening such colors in he above menioned spaces. These problem has been overcome by inroducion of uniform color spaces, such as C.I..L*u*v* and C.I..L*a*b*. Adoped color image segmenaion algorihms can be classified as follows: Feaure-space based, working on he space of coloured figures in he image Image-domain based, i.e. hey analyse he image geomery and color Physics based, involving he physic relaionships beween ligh and maerials In his work, a segmenaion algorihm belonging o he firs class was applied o remoe sensed images in order o improve qualiaively and quaniaively he resul of he classical approach. This means, assuming ha differen classes of land feaures are already idenified hrough classical mehodolog a furher analysis sep is inroduced by applicaion of proposed algorihm o refine he resuls of previous phases. The main advanage of his approach rely on he fac ha he algorihm is unsupervised, herefore i doesn require any a-priori informaion and can be fully auomaized. The paper is srucured as follows: in secion 2 a brief overview of feaure-space based echniques is repored, hen in secion 3 he proposed algorihm is explained. ome resuls of is applicaion o remoe sensed images are presened in secion 4, while secion 5 deals wih he conclusions. 2. TH FATUR-PAC BAD TCHIQU Assuming ha color is a consan propery of he surface of each obec appearing on he image, he image segmenaion can be ymposium on Geospaial Theor Processing and Applicaions, ymposium sur la héorie, les raiemens e les applicaions des données Géospaiales, Oawa 2002

2 addressed hrough wo following differen sraegies: clusering and hisogram hresholding. In he firs echnique, image pixels are firsly mapped on a cerain color space, in order o conver pixels o poins. Then hese poins are grouped in differen ses (clusers on he ground of color informaion of each corresponding pixel. In his wa given he above menioned assumpion, he differen obecs of he image can be discriminaed in erms of hese clusers or cloud of poins. The disribuion of he poins inside each cluser depends mainly upon he color change, due o shading effecs and noise of he acquisiion device. I should be noed ha he clusering echnique belongs o he unsupervised classificaion algorihms, since no a-priori knowledge abou he image is required. An example of clusering implemenaion is provided by he k-means algorihm: i is widely used no only for color image segmenaion bu also for applicaions involving vecorizaion and daa compression. The hisogram hresholding algorihm belongs o anoher class of segmenaion echniques, early applied o gray scale images. In his mehod image pixels are no mapped on a color space, bu raher some ad-hoc hisogram of color figures, such as he Hue, are generaed. Through ha model, obecs on he image will be idenified as peaks of he hisogram, while he background will correspond o is depressions. In he field of color images a hresholding algorihm involves a bi more complex implemenaion, since i has o work in a 3D color space, meaning ha a 3D hisogram has o be aken ino accoun. Furhermore, in his case hisogram profiles become quie agged wih spurious peaks, which make he segmenaion more ambiguous. The proposed mehod is based on a combinaion of he clusering and hisogram hresholding echniques. In summar given a remoe sensed image, he represenaive color are firsly idenified by looking for he maor color groups, hrough he hisogram hresholding of he Hue informaion. Then, he larger clusers in he planes of consan Hue are deermined, hrough he k-means clusering algorihm. 3. TH GMTATIO ALGORITHM The segmenaion process works in he C.I..L*u*v* uniform color space, provided wih euclidean norm L*u*v* = (L* 2 +(u* 2 +(v* 2 1/2. In his space a cilindric coordinae reference sysem was inroduced (H uv *, C uv *, L*, whose Hue angle is defined as H uv *=arcan(v*/u*, he crominance as C uv *=[(u* 2 +(v* 2 1/2 =L* and he sauraion as =([(u* 2 + (v* 2 1/2 /L*. The clusering mehod, based on anisoropic diffusion, is a non-linear filering echnique, which performs a more high selecive smoohing in omogeneus regions and almos null on he egdes, while i reains all he edge-relaed informaion. In his algorihm Hue H(x, and auraion (x, are represened as one complex quani he crominance funcion (x,=(x,exp( H(x,, which is in urn diffused, clusered and segmened. The modeling of Hue and auraion in he same funcion akes ino accoun he physic relaionship exising beween hem. I is well known ha Hue changes are negligible for low values of auraion, bu noiceable for high values. The same operaions are applied o he lighness funcion L*, which is however processed separaely. The combinaion of his wo parallel segmenaion asks leads o a pariioned color image. The overall scheme of developed algorihm is showed in Figure 1. Figure 1: cheme of he segmenaion algorihm in he dashed box The anisoropic diffusion has been numerically implemened hrough he parial derivaive equaion of hea diffusion, as saed below: ( x, / = div[ c( x, ( x, (1 where div is he divergence operaor, while is he gradien compued respec wih he spaial variables. uch equaion can be discreized hrough a square laice [3, wih he he complex Crominance value (x, associaed o he verices and he conducance coefficien c(x, associaed o he arcs, (see Fig. 2, as follows: [ + 1 = [ + c + λ[ c + c + c + (2

3 where 0 λ 0.25 is required for he sabiliy of he numeric scheme,,,, are symbols of he four verices of he laice and symbol δ defines he four differences neares-neighbour: = ( = ( = ( = ( i 1, i+ 1, (3 Finall he image segmenaion is obained by separaely pariioning of Crominance d (x, and Lighness L d *(x, and hen combining he resuls hrough he k-means algorihm. Given hese seings, he proposed algorihm, based on anisoropic diffusion, becomes easy o be implemened, feauring a local behavior, i.e. he amoun and he kind of smoohing are locally deermined by he values of complex crominance and are adoped for each image region. The choice for values of parameers T 1 and T 2 plays a maor role in he seup of clusering process. The firs, T 1, deermines he average disance among he clusers, while he second, T 2, refers o o he average radius of a single cluser, if considered as a circle, and i is no possible o fix an a-priori value suiable for all images. Anyway an esimae of T 1 can be obained from daa disribuion, esimaing he radius of circle ρ ω, in which a cerain percenage of he daa (ω = 95% are conained. Tess peformed on remoe sensed images showed ha seing T 1 =ρ ω /2 and T 2 =ρ ω /4 leads o excellen resuls on a large se of images and, in he same ime, provides a limied number of segmens (ofen less han 7, which well represen he color informaion of he image. Parameer T 2 can be freely se, provided ha T 2 < T 1 in order o avoid he overlapping of clusers. 4. TT AD RULT Figure 2: The basic cell of 4-eares-eighbours laice The conducance coefficien is upgraded a each ieraion as follows: ( ( ( ( The g( funcion can be modeled according o one of he following forms: g( = exp( g( = 1/(1 + 2 (( / A 2 (( / A The firs funcion is bes suied o highligh edges provided wih high conras respec wih he ones o low conras, while he second form discriminaes beer beween large and small regions. In his algorihm he second funcion has been chosen, in order o privilege he generaion of large regions on he image, while he A cosan value is dinamically compued for each ieraion and se as 5% of maximum value of (δ. (4 (5 In his secion some examples of he applicaion of he segmenaion algorihm are repored. The remoe sensed daa were acquired by he Landsa-TM saellie on 12/26/1996 and on 05/03/1997. In he firs phase, saellie images were processed using he classical approach, working on hree daa ypologies: "naural color" images, infrared pseudocolor images and enhanced pseudocolor images. The firs class provides he user wih phoographic likewise viewing of he errain, similar o color airphoos, while he second allows o ge a more in deph analysis, discriminaing beween differen caegories of he same obec class. For insance i is possible o deec differen forms of vegeaion, since he reflecance changes are srongly relaed wih he morphologic srucure of he leaf. nhanced pseudocolor images are obained by combinaion of differen bands and can be useful employed o deec he soil humidi urban and culivaed areas, and so on. Figure 3a shows a Landsa-TM "naural" image aken during spring 1997, as resul of applicaion of RGB fler o TM1, TM2 and TM3 bands. Figure 4a shows a view of Poro Baseleghe lagoon (Venice, orhern Ialy by he mouh of Lemene river, which was aken during winer In his case he specral componens of he image were enhanced hrough following bands combinaions: band TM7/5, where TM7/5=(TM7- TM5/(TM7+TM5, band TM/6 as (TM7-TM6/(TM7+TM6 and finally band TM3/1 as (TM3-TM1/(TM3+TM1. The firs band enhances he reflecance changes due o humidiy and soil composiion, he second discriminaes beween vegeaion and urban areas, while he las band allows o disinguish beween damp or dry soils, on he ground of heir composiion (organic maerial, kind of rocks. Resuls of applicaion of segmenaion algorihm o hese images are showed in figures 3b and 4b, in which clusers were early depiced in pseudocolors bu in his paper have o be convered hem in gray scale. Though i is no easy o appreciae by eye he performance of he algorihm, he resuls can be summarized as follows. Referring o Figure 3, he segmenaion of his image allowed o discriminae perfecly beween he wood area (in red in he resuling segmened color image and culivaed and dry soils (respecively in green and blue, as

4 repored in Figure 3b. As regards he Figure 4b, i shows how well urban areas, he sea and vegeaion areas could be segmened in comparison wih corresponding image of Figure 4a. Figure 4a : nhanced pseudocolor image of Poro Baseleghe Figure 3a : aural color image Figure 4b : egmened image of Poro Baseleghe 5. COCLUIO Figure 3b : egmened naural color image In his work a combined mehod for remoe sensed color images segmenaion was presened. The mehod is based on he clusering and hisogram hresholding echnique, early developed in he field of Compuer Vision for image processing. In his case he proposed algorihm was inroduced as a second sep in he ipical workflow of saellie image analysis in order o improve boh he classificaion resuls of classical approach and he georeferencing of deeced hemaic areas. The advanage of proposed mehod relies in he fac ha i is unsupervised, i.e. i doesn require any exernal human conrol or a-priori informaion. Therefore he procedure can be fully auomaed and can be easily implemened in any GI applicaion.

5 Performance of he algorihm were assessed by comparison beween he resuls of classical Remoe ensing image analysis wih he ones of he presened mehod, which were applied o he same se of saellie images. As showed in previous secion, he segmenaion algorihm could be successfully applied also o pseudocolor images, providing excellen resuls if he image is composed by wide and well defined regions. Unforunael if he saellie image is fragmened, and shows very similar colors disribued in a large amoun of small regions, he segmenaion algorihm provides ofen an ambiguous and unsaisfacory resul. ACOLDGMT This work was developed wih he proec Techniques for auomaic processing of daa acquired by inegraed sysem parly financed by MURT (Ialian Minisry of Universiy and Research in year 2001 as proec of relevan aional ineres. aional coordinaor: Giorgio Manzon head of he Research uni Anonio Veore. RFRC [1 Lucchese L., Mira.., An algorihm for unsupervised color images segmenaion. Proc. of 1998 I 2 nd orkshop on mulimedia signal processing, Redondo Beach, CA, UA, pp [2 Lucchese L., Mira s.., Unsupervised segmenaion of color images based on k-means clusering in he chromaiciy plane. Proc. of I orkshop on conen-based access of images and video libraries (CBAIVL 99, For Collins, CO, UA. [3 Perona P., Malik J., cale space and edge deecion using anisoropic diffusion. I Trans. on PAMI, Vol 12, o 7, pp , July 1990.