Second ISAG, Oxford (UK), 21 -231911993 221

IDENTIFICATION OF FAULTS ON MONTAINOUS AREAS BY ANALYSIS OF SPOT IMAGES.

APPLICATION TO THE EXTRACTION OP STRIME FAULTS FROM A SUB-SCENE OF SOUTHEIXN PERU

Catherine MERING(l) (2) and Jean CHOROWICZ(2)

( 3 ) Laboratoire d'Informatique Appliquée, ORSTOM, 72 Rte d'Aulnay, 93143 Bondy Cedex France (2) Universitt? Pierre et Marie Curie, Paris VI, Dpt de Géotectonique, 4 place Jussieu, 75252, Paris Cedex 05 France Fax: 33( 1)44275085

RESUME: L'exploitation des images satellites B haute résolution en vue d'identifier les structurcs gbologiques se fait généralement par photo-intcrprétation. Les nouveaux développemenls de l'analyse d'image numérique ouvrent la voie 1 une photo-interprétation assistée par ordinateur qui permct d'améliorer la rnisc en bvidcncc dcs traits stru_cturaux, ainsi quc lcur cxtraction. Nous montrons ici un cscmple dc misc en évidcncc et d'cxtraction de faillcs pcrceptiblcs sur unc sdnc dc SPOT Panchromatiquc sur les Andes Centralcs (rkgion d'Arequipa, au Sud du Pérou).

KEY-WORDS: image analysis, remote sensing, recognition of faults, structural map.

INTRODUCTION

One of the objectives of remote scnsing on the Earth is to enable the survcy and mapping of objccts on thc ground surface. The maps obtained from rcmotcly senscd images cm be exploited in various ways like traditional cartography does. Being indecd numerical images, they can easily be combined to any kind of geocoded dala inside a Geographic Information Syslcm: an example would be thè combination of structural maps and obscrved gmoded epicentres for seismic prevision. On these images, gcological structures have no spccifie reflcclance but they have a specific morphology. That is why methods of classification on multispectral data are of no use to map them. On images produced by pssive sensors as the ones of SPOT, forms are dctcctcd thanks to the cas1 shadow rcsulting from grazing illumination on the relief. In thc following example, we try to show how these imagcs can be processed with the methods and algorîthms of Image Analysis to identify structures, and more generally to demonstrate the fcasibility of a "cornputer aided photo- intcrprctation". We,had rccoursc hcrc to binary or grey tonc rtmrpirolo,qicnZ and geode.sic transformations (Scm, 1986), thc last oncs having thc propcrty to prcservc thc contours of the structurcs. Al\ thc Morphologieai Transformations have bcen processed without any modification of the original square digital grid. The algorithms have been addpted to the case whcre the volumc of data is very high: indeed, a Panchromatic Spot sccne has 42 millions of pixels We have thercfore pcrformcd iterative computalion, in which only the strip of pixels having the required width according to the size of the structuring element involved in the transformation, is storcd at each step.

MAPPING OF FAULTS IN MONTAINOUS AREAS

In order to illustrate the mcthod, we have madc the choice of a scene upon Ccntral Andes because, although thc whoic rcgion is highly affccted by cruplions and carlhquakes (Dorbath, W), there is n o exhaustive mapping of the faults. We attcmpt here to map strike faults, easily pcrceptiblc on SPOT images. Strike Faults can bc dcscribcd as dark and sub-rcctilincar lincs with spccific orientation. A largc part of normal or rcvcrse faults correspond on the field to abrupt scarps which have t o be associalcd to the t'au11 throw. Thc darkncss and thc thickncss of thc lines intcrpretcd as faults depends on the gcomorphology or the scarps but also on the *

conditions of the viewing. On Figure 1 (corresponding to a square of size lOkm on the ground) which is a partial victv on the zone of Arequipa (South of Pcru) takcn in July, the deduction of the structures by cxtraction of continuous dark lincs seems easy to be donc by simplc photo-intcrprcbtion. Neverthcless weAtry to lcarn h y v to extract thcm by image analysis, our middlc-date scope bcing to find available automatic scquenccs of trcatments and to reproduce thcm on the sccnes collccte'd above the whole target (southcm Peru, northcrn Chile). The darkness, the shapc, and thc sizc of the objccts may bc pertincnt critcria 10 cxtract f'aults or eicmcnts of faults; on the imagc othcr elemcnts are both lincar and dark the darker and the thickcr oncs arc

222 Second ZSG, Oxford (UK), 211-231911993

Figure 1 Figure 2

Figure 3 Figure 4

Figure 5 Figure 6

Second ISAG, Oxford (UK), 21-231911993 223

actually ravines inside versants in the shadow (sce for example north-wcst of the image). The widcly-known technique of numerical convolutions with gradierzt rnasks (Pratt, 1978) enhances thc contrasts in a givcn direction: in lhe present case, the drawback of such a technique is that both dark and iight lines are enhanccd on thc rcsulting image (sce Fig 2 which is the north-south gradient-image). Our stratcgy was thcn to apply Morphological Transformations on grey tone function having the property to filtcr the darkcr objccts from thc image. By threshdding the rcsulting function, WC should retain two typcs of binary images: "maximal imagcs" arc thc oncs whcrc thc significative enlitics arc conncctcd, rcgardless to the prcscncc o f othcr cntitics, whcn "minimal images" would be those whcre only thc significative cntitics arc prcscnt, evcn if disconncctcd. Theorctically, a Geodesic Dilation of a minimal image into a maximal one should provide a good solution to Our problem. We first apply on the original image a il-minima filter (Beucher, 1987) that docs not takc into account any morphological critcrion, but sclecls only dark objccts having a given contrast with the neighborhood (Fig. 3): al1 types of dark components have then been selected such as large spots or thin lines; the faults are of course compriscd into the resulting binary image (Fig. 4) and thcir shapcs are wcll preservcd. But one has to find thcn thc way of supprcssing undcsircd largc or small black spots. A size clistrilmtiotr analysis has to bc donc to find whether faults have a spccîfic size or not on thc h-minima thresholded function. The robustncss of the mcthod has 10 be testcd on a large sample of scenes upon the target. Other transformations called Top Hat (because thcy allow the sclection of objects according to their tonc, local contrast and thickness) are experimented. WC applied here a Top Hat from Morplwlogical Closing (Fig. 5) and a Top Hal frorn Geodesic Closing (Fig 6) to select dark lines On thc first image (Fig. '5), the contours of the faults are better prcservcd for they are actually dark and continous entitics when Top Hat from Gcodcsic Closing filter (Fig. 6) selects the darker even disconnected elcmcnls (Grimaud, 1991). One solution for finding a minimal image $vas to process a stcpwise "clcaning up" of the binary image from the morphological Top Hat: -smallcst disconnccted entities have been eliminated by a Reconsfruction following a binary Erosion (Fig. 7). -entitiCs the direction of which arc whcthcr diffcrent than thc one ofJhe faults (ic roughly East-West) havc bccn eliminatcd by the following way: a local multidirectional Gradient filter produccs an image of codcs of the associatcd direction of thc highcst local gradient (ie €4 directions on the squarc grid); WC havc sclcctcd thc pixcls having the same codc than the one of both fmlts (Fig 8). Thcn a Geodesic Dilation of the rcsult into the "c1cancd"Top hat image is pcrformed (Fig 9). The remaing big entities on north-west of the image arc very close one to another: they are ravines inside abrupt versants; they have been eliminated by the application o f a mask obtained by a Morphological Closing of size 5 on the resulting image "cleaned up" with a Geodesic Reconstruction (Fig.10) This "minimal image" is then reconstructed into a low thresholding of the original Top Hat (Fig 11). On the rcsulting image (Fig 12) the faults have not been completely connected and their shape is notas well prcservcd as on a h-min filtered image (see Fig 4) Moreover al1 entities othcr than faults have not bccn eliminated. WC think thcse last ones could be elimined by rcgional directional opcntors (Kurdy, 1990) or by further quantitative description of the slructurcs.

CONCLUSIONS Various expcriments of image analysis and morphological transformatons havc bcen made upon a

SPOT image of Central Andes in order to extract strike faults. On this example wc show how satellite images may be filtered until specificd structures are identificd. Plans of transformations to be tested on the scencs of the whole region were also described. The kind of computcr aided photo-interprctation can be easily testcd and reproduced. Furthermore, the results can directly be exploited in order to perform quantitative analysis on the structures or to be inserted into a Geographical Information System..

REFERENCES (Beuchcr, 1987) S-Beucher, J.M. Blosseville, M. Bilodeau, F. Lenoir, S. Espie. Tilan: système de meswe du tr($cpar analyse d'inmges. Rapport commun INRETSICMM. Note interne CMM: N.46/87/MM.Oct. 1987. [Dorbath, 901 L. Dorbath, A. Cisternas and C. Dorbath. Assessrnent of the size of large and great Iustorical earflqztakesinPer~~ Bull. of Seism. Soc. of Amcrica, Vol 80, N"3, 1990, pp551-576, (Grimaud, 1991) M. Grimaud. La Géodesie mtmériqlte en rnorphologie mnthematiqlce. Application à la déléction automatique de microcalcijcation en nzarn~zographie nrtmériqrte. Thesc Ecolc dcs Mincs dc Paris..Déccmbre 1991. (Kurdy, 1990) B. Kurdy. Transformations rnopkologiques directionnelles et adaptatives: Application ~ I L T

sciences des matériaux. Thèse Ecole des Mines de Paris. Scptembrc. 1990. (Serra, 1986) J. Serra. Image Anolysis and Mathetnatical Morphology. V01.2. Theoreticai advances. Acadcmic Press. London. 1986.41 lp.

224 Secorad ISAG, Oxford (UK), 21-23/9/1993

Figure 7 Figure 8

Figure 9 Figure 10

Figure 11 Figure 12