Vers une remise en géométrie automatique des anciennes campagnes aériennesphotogrammétriques

S. Giordano1 A. Le Bris1 C. Mallet1

1 Univ. Paris-Est, LASTIG MATIS, IGN, ENSG, F-94160 Saint-Mandé, France

sebastien.giordano@ign.fr, arnaud.le-bris@ign.fr, clement.mallet@ign.fr

Résumé

Les campagnes aériennes photogrammétriques an-ciennes sont un outil unique et assez peu exploité pour lesuivi des changements 3D d’occupation du sol au coursdu siècle passé. Ces études restent limitées en termesde surface et de nombre de dates concernées, du faitde l’absence de méthode automatique pour géoréferen-cer finement les images anciennes, le principal verrouétant l’identification de points d’appui en nombre suf-fisant. Ce papier adopte une approche photogrammé-trique en 2 temps. L’orientation absolue des images estcalculée automatiquement de manière approchée, et sui-vie d’une détection automatique de points d’appui utili-sés ensuite pour l’amélioration de ce géoréférencement.Orthoimages et MNS récents sont pris comme référencepour toutes les dates. La détection automatique de pointsd’appui se fait par comparaison de cette orthoimage deréférence avec celle calculée à partir du géoréférence-ment approché des images.

Mots Clef

Images aériennes, archives, orientation, points d’appui,modèle numérique de surface.

Abstract

Archival photogrammetric aerial campaigns are a uniqueand relatively unexplored means to chronicle 3D land-cover changes over the past 100 years. However, existingstudies are rather limited in terms of area and number ofdates since no fully automatic method has been propo-sed yet to finely georeference archival images, the majorchallenge being the identification of a sufficient amountof ground references : cartographic coordinates and theirposition in the archival images. This paper uses a photo-grammetric 2-step approach with the automatic compu-tation of a coarse absolute image orientation followed bythe automatic extraction of ground control points usedto refine orientation. It only relies on a recent orthoi-mage+DSM, used as reference for all epochs. The coarseorthoimage, compared with such a reference, allows theidentification of dense ground references.

KeywordsAerial images, archives, orientation, Ground ControlPoints, Digital Surface Model.

1 IntroductionPhotogrammetric archival aerial images were initially ac-quired by mapping, cadastral or military agencies for to-pographic map generation. Stereoscopic configurationswere adopted to offer 3D plotting capacities. These sur-veys have been a common practice in many countriesover the last century. Recently, several countries have di-gitised their film-based photos (e.g., >3 millions imagesin France), and facilitated their access through dedica-ted infrastructures and web services with basic metadataand visualisation capacities. Such images are a uniqueyet unexplored means for long-term environmental mo-nitoring and land-cover change analysis [13]. The timeseries are (i) long (first images : 1920s), (ii) relativelydense temporally (every 2-5 years), (iii) acquired at veryhigh spatial resolution (<1-2 m) and, (iv) generally en-abling to generate 3D information. However, available di-gital images have been generated from old analog photo-graphs, acquired with different cameras, for which onlycoarse image localisation was stored while their use inremote sensing workflows requires a fine georeferencingstep between dates so as to generate orthoimages (andDigital Surfaces Models) that can be spatially compared.Useful metadata is generally distributed, such as lengthand width of the analog photos, focal length of the ca-mera. Flight plans of the aerial survey giving very coarseimage localisation are often distributed with the images.This study focuses on the automation of the georefe-rencing stage and proposes a coarse-to-fine solution, en-abling to provide comparable orthoimages and DSMs.

2 Fine georeferencing of archivalimages

In this paper, the hypothesis is made that a very coarseimage location is supplied for each archival aerial image(∼100m). It therefore focuses on fine georeferencing ofsuch data into a cartographic system. A precise accuracyis expected so as to be able to perform comparisonbetween the products that can be derived (radiometry,

height, labels) at different dates.Non-photogrammetric and photogrammetric geo-referencing approaches can be distinguished. Non-photogrammetric approaches find an empirical transformthat gives the best match between images, while photo-grammetric approaches aim at modelling the geometryof all images (enabling DSM computation).An overview of existing approaches according to theirspatial and temporal characteristics is provided in Fi-gure 1. First, most studies are based on photogrammetricapproaches (6 non-photogrammetric vs. 16 photogram-metric) but no automatic solution is used. For photogram-metric approaches, a majority of studies concern less than5 dates and 100 km2. By increasing size, these papersare : [20, 24, 32, 22, 30, 16, 11, 5, 25, 1, 4]. [21] (9 dates),[27] (10) and [18] (28) are the only authors that haveprocessed more than 5 epochs. [27] consider relativelylarger area (∼ 600 km2) for DSM generation in fores-ted areas. Same conclusions (restricted epochs and areasize) can be made with non-photogrammetric approaches[9, 15, 6], although some of them try to process largerareas [2, 12, 10]. The time interval and image age are notlimiting factor : the average time interval is 30 years andthe strating year around 1945-1955. Spatial (study area)and temporal (number of epochs) limitations can be dueto many factors and probably mainly to the huge manualimplication in the fine georeferencing process.

Korpela et al. (2016)

Nurminen et al. (2015)

Ford (2013)

Micheletti et al. (2015)

Asner et al. (2003)Galster et al. (2008)

Cardenal et al. (2006)Fox and Cziferszky (2008)Necsoiu et al. (2013)

Véga and St-Onge (2008)Walstra et al. (2004)

Chen et al. (2016)Meyer et al. (1996)

Nocerino et al. (2012)Ayoub et al. (2009)

Jao et al. (2014)Aguilar et al. (2013)

Kadmon and Harari-Kremer (1999)

Ellis et al. (2006)

Mondino and Chiabrandoa (2008)Nebiker et al. (2014)

FIGURE 1. Overview of state-of-the-art studies. Papersare plotted according to the area of the study site, thenumber of dates and the time range of the time series.

2.1 Non-photogrammetric approachesFine non-photogrammetric georeferencing methods areusually based on the application of a polynomial trans-form to each archival image of the aerial survey. Thesetransforms are estimated using necessary ground refe-rence, usually Ground Control Points (GCPs), acquiredusing recent satellite imagery [9, 10], recent national ae-rial orthoimages [9, 12] or GPS survey on road inter-sections [2]. These approaches imply the independentidentification of dense ground reference for each archi-val image of the survey (from 6 [12] to 56 GCPs [2]per image. The planimetric accuracy obtained is usuallyfound compatible with multi-temporal analysis (<1m,

[2, 12] – <2m [10]). As finding corresponding GCPs bet-ween recent and historical datasets is a complex task, inparticular in areas facing significant changes, some au-thors [15, 6] have proposed to use standard automatictie-point matching techniques (e.g. [19]). However, thetie-points are only extracted between the archival imagesthemselves to reduce the number of manual GCPs neededto estimate the polynomial transform but no tie-pointshave been extracted between the archival images and aground reference.Non-photogrammetric methods are suitable to obtaingeoreferenced archival orthorectified images with a sui-table planimetric accuracy. However, compulsory manualinteraction remains a major limitation to move to largerspatial or temporal scales.

2.2 Photogrammetric methodsSeveral photogrammetric softwares are adapted to pro-cess archival aerial image surveys [28, 24]. In the follo-wing, the current automation level is assessed for eachstep (interior orientation, relative image orientation, ab-solute image orientation, DSM and orthoimage genera-tion) of standard photogrammetric workflow.

Interior orientation The analog photographs were de-formed and thereafter scanned. The interior orientationcorrects this establishing for each image a mathemati-cal transform between the camera frame coordinate andthe image coordinate system. Is is done using the fidu-cial marks of each image. Calibration certificates can beused to retrieve the fiducial mark physical coordinates[27, 18, 21, 30], but are barely available today. Thus,some authors propose to use coarse calibration informa-tion (calibration models). [1] only use the focal lengthinformation usually saved or mentioned on the photo-graphs, whereas [26] absolve themselves from calibrationcertificates by recovering an approximate informationfrom the images. Many photogrammetric softwares (e.g.,MicMac [28]) propose semi-automatic fiducial mark de-tection to retrieve image coordinates (pixels) of fiducialmarks. The manual digitalisation of the fiducial marks isonly necessary on a single image of the survey.

Relative orientation The relative image orientation iscarried out using automatically detected tie-points andthe interior orientation. No manual intervention is requi-red at this step, since automatic tie-point extraction isa solved problem. During this stage, the calibration ofthe camera of an aerial survey is usually questioned andre-estimated. This self-calibration is done automaticallyusing the tie-points [27, 5, 25, 11, 32]. Nevertheless, itmust be kept in mind that results of self-calibration oftenremains an approximation because of the poorly geome-trically constrained configuration of the image sets. Therelative image orientation has a great impact on the pla-nimetric and altimetric accuracies of the orthoimages andthe subsequent DSM.

Absolute image orientation. Producing an absoluteimage orientation in a cartographic system requires coor-

dinates of ground references. The most common methodis to use Ground Control Points (GCPs) and to determinetheir image coordinates in the aerial photographs. As fornon photogrammetric methods, GCPs are measured ma-nually using ancillary orthoimage and a Digital TerrainModel [27, 26, 25, 30, 9, 11], field GPS [21, 5, 32], orwith both techniques [1]. The required number of GCPsis important compared to the size of the study sites.[27]argue that this important number of GCPs is required totie the images to the ground reference system becausecamera calibration information as well as adequate ap-proximate exterior orientations are often missing. Thus,self-calibration is usually re-performed during the abso-lute image orientation stage.All authors mention the determination of GCP as verytime consuming. Indeed, apart from measuring their co-ordinates, finding these points is very challenging. Theyhave to be stable over time on possibly a long period oftime and precisely identified on the aerial photographs[21]. The planimetric and altimetric accuracies of the ab-solute image orientation depends, among other parame-ters, on the scale (spatial resolution) of the image survey.Most studies obtain on several image surveys <1m to2m planimetric accuracies, depending on the scale of theacquisitions [21, 1, 5, 25, 30, 11] or between 2m to 10m[26]. Thus, absolute image orientation remains the ma-jor lock to full automation or, at least, to reduce humanprocessing times.Very few methods propose solutions to overcome this is-sue. [18] adopts multi-date bundle-block adjustment withrecent images whose absolute orientations are known.He found that automatic tie-points matching is not suffi-cient : manual multi-temporal tie-points had to be identi-fied (279 tie-points for 288 images). [7] and [23] proposeto use linear features characterizing permanent structuressuch as main road axis and building outlines as candidateground reference. Manual digitalisation of these line fea-tures is still mandatory both in the archival images andthe ground reference information in [23]. [7] opted for anautomatic procedure that matches objects of a topogra-phic vector database to lines extracted within the images.However, no bundle block adjustment was performed :the method has to be applied to each archival image inde-pendently.

DSM and Orthoimages. As long as the image model(absolute orientation and calibration) is known, orthoi-mages and DSMs can be calculated. The DSM is laterproduced thanks to a dense matching algorithm. Thus,the DSM can then be used to orthorectify each archivalimage of the aerial survey : true orthoimages can thenbe created. A DSM on archival aerial images was pro-duced in [27, 5, 26, 30, 32, 13]. The important numberof GCPs used in these studies seem to allow to obtainconsistent altimetric DSM accuracies : from a metric ac-curacy [27, 5] to 4m [32]. This altimetric accuracy de-pends among other parameters on the scale of the aerialphotographs. However, when the image model is suffi-ciently good, dense image matching algorithm can be ea-

sily used on archival stereoscopic aerial image surveys.Only [24] compare 3 methods for dense image matchingon different criteria.

2.3 Proposed approach and contributionsThe determination of sufficient ground reference is themajor obstacle to process larger areas and more epochs.This statement concerns both non-photogrammetric andphotogrammetric approaches. It can as well be noticedthat very few studies propose to automate this task. Thus,this paper proposes a workflow providing solutions to ad-dress the problems mentioned above. The automatic de-termination of dense ground references, stable over time,is considered. At the end, a coarse-to-fine strategy bene-fiting from available coarse metadata is proposed.In this paper, a photogrammetric approach is adoptedconsidering that the DSM that can be computed is animportant asset to reach full automation. In the archi-val aerial image processing, automatic tie-point extrac-tion and matching has never been tackled between twoepochs when land-cover changes are important (time in-terval) or when the images of the surveys are very dif-ferent (in terms of radiometry and spatial resolution).This issue is overcome with such a two-step absoluteimage orientation algorithm. First, coarsely georeferen-ced orthoimages and DSM are produced for each yearwithout any additional ground reference information (ex-cept basic metadata, usually provided with the archi-val aerial images). Then, a fine absolute image orienta-tion of the archives is computed using GCPs automa-tically detected between the coarse orthoimages and arecent orthoimage/DSM (GCPs ground coordinates mea-surement). The coarse absolute image orientation and co-arse DSM are inserted so as to automate the identificationof the selected ground references in the archival aerialimages (GCP image coordinates).

3 MethodsFigure 2 shows the photogrammetric workflow proposedin this paper with main inputs and outputs. The work-flow can be divided into 3 main parts. First, the inter-ior and relative orientation steps are based on a standardphotogrammetric method (Section 3.2). Then, an origi-nal two-step (coarse-to-fine) absolute image orientationis proposed (Sections 3.3 and 3.5). Eventually, the au-tomatic GCPs determination provides ground referencesfor a new bundle adjustment (Section 3.4).

3.1 InputsThe necessary inputs of the method are (i) the scanned ar-chival photographs (ii) accompanied with very basic me-tadata information. The metadata can correspond to thedigitalisation process (scan resolution) or the configura-tion of the aerial survey (focal length of the camera, thephysical size (mm) of the photographs and a coarse flightplan). Focal length and the physical size of the photo-graphs are metadata that had usually been archived. A co-arse flight plan is also necessary, albeit it does not need to

Focal lengthPhysical size

ImagesScan resolution

Flight planOrthoimage

DSM

Interior and relative orientation

Interior orientation

Relative orientation

Relative poseCamera calibration

Absolute Orientation

Absolutecoarse

orientation

Absolutefine

orientation

Coarse poseOrtho/DSM

Fine poseCalibrationOrtho/DSM

Automatic GCPs

AutomaticGCPs

determination

GCPs

FIGURE 2. Proposed method. See text for more details.

be very accurate. The planimetric ground coordinates ofthe image nadir (≈ 100m) are already available in mostcases, since they are a basic requirement for the distribu-tion of the images though web-based platforms. Approxi-mate flight height is also necessary to have a coarse 3Dinformation on ground position of the pose of the images.Recent orthoimages and DSMs are also required as a pla-nimetric and altimetric reference. They act as a referencedataset on which archival images are oriented.

3.2 Interior and relative orientationFor interior and relative orientations, state-of-the-art me-thods usually implemented in photogrammetric softwaresare used. First, the interior orientation is performed with asemi-automatic method implemented in the MicMac soft-ware [28]. The image coordinates of the fiducial marksare determined on a single archival image and are thereaf-ter found automatically on the other images of the aerialsurvey using pattern matching techniques. Relative orien-tation (bundle-block adjustment) is performed based onautomatic tie-point extraction and matching. At this stagea first camera calibration is estimated.

3.3 Coarse absolute orientationThe coarse absolute orientation proposed in this paperconsists in applying a geometric transform (3D simila-rity) to the relative image orientations. This transformis estimated comparing the archival image center ground

coordinates of the relative poses and the flight plan. Co-arse absolute image orientation is then obtained. Next,orthoimages and DSM are computed using this coarseimage orientation.

3.4 Automatic ground reference identifica-tion

Several approaches have already been proposed in orderto perform cross-domain homologous points extractionout of multi-modal remote sensing images as for instanceusing Dense Adaptive Self-Correlation descriptors [17].Methods have also been developed so as to deal with dia-chronic images. They have never been applied to archivalaerial images. They are often learning approaches : key-point detection and matching are then trained from refe-rence examples. Among these methods, TILDE [31] for-mulates this problem as a regression model while LIFT[33] adopts deep learning models.The approach proposed in [3] is adopted here. Indeed,it has been specifically developed in order to extract tie-points between images and historical paintings. Thus, itcan deal with both multi-modal and multi-temporal data.This method relies on HoG descriptors [8] and also consi-ders matching as a classification problem. The underlyingidea is to train for each pixel’s HoG descriptor a binaryclassifier "similar descriptor"/ "non similar descriptor"from one positive example and many negative examples.A multi-resolution approach is adopted for keypoint de-tection. A scale-space is computed from the image, andkeypoints are detected at each level as local maxima ofthe keypoint relevance score, assessing how much thispoint is discriminative according to its HoG descriptor.This keypoint relevance score is related to the optimalcost a square-loss SVM. In practice, keypoints are firstextracted out of the recent reference orthoimages. Thus,planimetric ground coordinates are directly associatedwith these keypoints. Height information from the recentreference DSM can then be linked to them. This enablesto retrieve 3D GCPs.They are then matched to the archival orthoimage cal-culated during the previous step. A tile-based approachis adopted. Indeed, the coarse georeferencing of archi-val data makes it possible for each tile to select only thecorresponding subset of possible keypoints detected fromthe reference orthoimage. A dense matching is then per-formed : a scale-space of the archival image tile is cal-culated, and the matching scores between HoG descrip-tors are calculated (i) for each detected keypoint and (ii)for each pixel of each level of this scale-space. This mat-ching score is defined from a Linear Discriminant Ana-lysis classifier. For each reference keypoint, the two bestmatching scores are considered. The ratio between thescores of the first and second best matches is compu-ted [19]. At the end, only a subset of matches with thebest matching scores ratio is kept. A RANSAC filteringis then applied to them. For the remaining matches, pointsdetected in the archival orthoimage are reprojected inthe corresponding archival aerial images, to retrieve their

image coordinates. This re-projection is performed usingthe ground-to-image model given by the coarse absoluteimage orientation and the height information from the ar-chival DSM generated at the previous step.

3.5 Fine absolute orientationAbsolute archival image orientations as well as cameracalibration are re-estimated at this step : the automaticGCPs determined in Section 3.4 are used together withtie-points within a robust least square bundle block ad-justment. The coarse absolute image orientation is usedas the initial solution. The fine absolute orthoimages andDSM are finally computed.

4 Results4.1 Experimental set-upThe method was tested on two study areas. Fabas isan agricultural and forested area where limited urbanchanges have occurred. It is located in the South WesternFrance. Important land-cover changes correspond to theregrouping of agricultural parcels, the creation of an arti-ficial lake and natural forest evolution. The Frejus studyarea is located in the South-East of France. On this area,major urban land-cover changes had occurred. The sitewas mainly dedicated to agriculture with small villagesafter the second World War but has gradually become do-minated by continuous dense and residential urban areas.5 different aerial surveys for each study area were selec-ted (Table 1). Only the panchromatic channel is kept.

date nb scale scan GSD focal height size(dpi) (m) (mm) (m) (mm×mm)

Fabas study area (12 km × 10 km)1942 55 20 000 768 0.66 150 3400 180×1301962 35 25 000 1200 0.53 125 3425 180×1801971 25 30 000 1200 0.64 152 4800 230×2301984 36 17 000 1200 0.36 214 3870 230×2301992 23 25 000 1200 0.53 153 3900 230×230

Frejus study area (3 km × 2 km)1954 19 5 000 1200 0.11 502 2530 300×3001966 15 8 000 1200 0.17 210 1780 180×1801970 19 8 000 1200 0.17 210 1770 180×1801978 10 14 500 1200 0.31 153 2315 230×2301989 12 10 000 1200 0.21 214 2550 230×230

TABLE 1. Aerial survey datasets. nb : number of imagesof the survey ; scale : mapping scale of the survey ; scan :the scan resolution during digitization ; GSD : an esti-mation of the ground sample distance ; focal : the focallength of the camera ; height : the flight height of the sur-vey ; size : the physical size of the archival photographs.

In the proposed method, reference orthoimages and DSMare needed. They are extracted from IGN’s referenceFrench national databases. Orthoimages were producedat a 0.5m GSD from airborne images. An orthoimage of2014 was used for Frejus study site, and a one of 2010 forFabas. DSMs were computed at a 0.5m GSD.

4.2 Orthoimages and DSM assessment

Coarse absolute orthoimages and DSMs. Anexample of the coarse absolute orthoimages and DSMsis given in Figure 3 for a small part of the Fabas studyarea (1984 aerial survey). Visually, the level of detailscontained in the DSM is highly meaningful : roads,forests and hedges are clearly visible on this natural area.The different archival images were correctly mergedtogether to produce the coarse absolute orthoimage.No local displacements were observed between eachorthorectified archival image. Comparing Figure 3a,band Figure 3c,d shows that the scale of the coarseabsolute products is consistent but the planimetricaccuracy appears biased (only on the y-axis (∼120m)on this area). Such shift can also be observed on Figure 5for 1942 survey. The transform was found rigid as asimple first order polynomial transform applied to thecoarse orthoimage with few manual GCPs found inthe reference orthoimage (Figure 3d) is sufficient tocoarsely correct it. The resulting corrected orthoimage isgiven in background orthoimage of Figure 3c). However,a comparison between the coarse absolute and thereference DSM show that the altimetric residuals are stillsignificant and systematic. No polynomial transform wasfound to be able to correct these residuals. This problemis probably due to the difficulty to estimate the cameracalibration only with the tie points.

(a) (b) (c) (d)

FIGURE 3. Results for Fabas 1984 (3.5km×3km). (a) Co-arse absolute DSM; (b) Coarse absolute orthoimage ; (c)Altimetric residuals between coarse DSM and referenceDSM (+60m→ +20m) ; (d) Reference orthoimage.

Fine absolute orthoimages and DSM. Examples ofautomatically extracted GCPs can be seen on Figure 4(Frejus). This example involves a recent orthoimage of2014 and archival imagery of 1954. It can be seen all ex-tracted points do not have the same precision but eventhe worst of them could generally at least provide an-chors to improve a coarse georeferencing. It can also benoted that GCPs were generally detected near man-madestructures (roads and buildings), which could lead to in-homogeneous density in natural areas. These GCPs werethen used within a robust least square bundle adjustment,

allowing to filter some errors. The number of remainingGCPs as well as the mean absolute residuals are providedin Table 2.Orthoimages calculated from archival images have beenfirst visually assessed on both areas : they were displayedtogether with current orthoimages in a chessboard view toassess the continuity and the alignment of persistent land-scape features. They were also superimposed with recentreference building and road vector databases. Represen-tative examples of such views are shown on Figures 6 and7 for both study areas. The 2D registration was generallysatisfactory, corresponding to what could be expected ac-cording to the quality of the different input data.At first glance, the quality of DSMs seems suitable : thedifferent ground/above ground objects can be correctlydistinguished. In addition, when considering the diffe-rence between the reference DSM and the calculated ar-chival DSM, errors due to systematisms were stronglyreduced (<2 m in unchanged areas). The difference isshown in Figure 8 for the 1971 Fabas study area. Howe-ver, some problems still occur. For instance, some lineardiscontinuity jumps can still be observed, revealing someremaining calibration errors. Stronger gaps can be obser-ved near the borders of the study site, where the bundleadjustment was less constrained. Nevertheless, it must bekept in mind that both study areas are not simple cases.For instance, an important part of Frejus area is coveredby the sea where not tie-points and GCPs can be found,thus limiting the possible constraints during the calibra-tion process.

date nb mean(x) mean(y) mean(z) variance(x) variance(y) variance(z)(m) (m) (m) (m) (m) (m)

Fabas study area (12 km × 10 km)1942 3275 4.52 4.13 1.85 12.48 10.65 5.651962 5688 4.21 4.04 1.86 13.07 13.08 5.341971 3029 2.78 2.87 1.52 6.49 6.78 2.921984 3438 6.12 3.78 2.63 76.43 14.93 10.311992 6222 3.15 3.46 1.99 6.79 7.07 3.60

Frejus study area (3 km × 2 km)1954 1133 1.01 0.87 1.85 0.51 0.57 3.511966 5116 1.15 0.83 1.60 0.66 0.45 2.931970 4661 0.86 0.93 1.84 0.94 0.86 4.301978 8032 2.56 1.29 2.97 4.19 1.55 6.781989 18670 1.02 1.02 1.42 0.72 0.84 2.74

TABLE 2. Number of automatic GCPs extracted (nb) andcomparison of their 3D ground coordinates (after fine ab-solute georeferencing vs. current reference).

5 Conclusion and perspectivesAn almost fully automatic workflow for the photogram-metric georeferencing of archival images was proposedbased on a coarse-to-fine absolute image orientation ap-proach. Homologous point extraction between each ae-rial survey and a current master reference was performedautomatically using GCPs. The computation of the co-arse absolute orthoimages makes this step much easier.Besides, advantage was taken from the coarse absoluteimage orientation and the computed DSM to perform thismatching step on true orthoimage and then to automa-tically determine the position of the homologous pointin the archival images. Archival orthoimages and DSMs

FIGURE 4. Frejus 1954. Examples of GCPs automaticallyextracted from 1954 orthoimage (left) and current orthoi-mage (right)

.

FIGURE 5. Fabas 1942. "Coarse absolute" 1942 orthoi-mage displayed together with the recent reference orthoi-mage.

FIGURE 6. Fabas 1942. "Fine absolute" 1942 orthoimagesuperimposed with the current orthoimage and a recentbuildings and roads vector database (yellow).

FIGURE 7. Frejus 1954. "Fine absolute" 1954 orthoimagesuperimposed with the current orthoimage and a recentbuildings and roads vector database (yellow).

-200 m -120 m -5 m 5 m

FIGURE 8. Fabas 1971. Difference between referenceDSM and archival DSMs calculated from coarse (left)and refined georeferencing (right), respectively.

could then be calculated. Their planimetric registrationappeared to be very satisfactory. 3D information is impro-ved but still remains imperfect. The quantitative analysisprovided in this paper is not sufficient. Indeed, it must bekept in mind that the extracted GCPs have varying geo-metric precision. Indeed, even for correct matches, theirplanimetric depends on the level of the scale-space atwhich they are detected. Besides, their height is estima-ted from a DSM. Its precision is related to the one of theDSM. Thus, some quite strong residuals can in fact cor-respond to the truth (land-cover changes) and not to an er-roneous bundle adjustment result. A rigorous planimetricand altimetric accuracy assessment will be proposed inthe near future. 3D check points are intended to be measu-red by human operators soon. Additional improvementsare still necessary, especially to extract more numerousand robust GCPs. First, more GCPs could be extractedusing other point-based methods, as for instance deeplearning based ones [33]. A specific architecture couldbe trained, benefiting from available georeferenced or-thoimages at different dates. Other approaches could alsobe investigated. Some buildings are stable features acrosstime, and thus, as the DSM was improved, it can be consi-dered to perform 3D registration between archival DSMand reference DSM over patches around them, taking intoaccount a quality score to assess whether a change occur-red. The approach of [7] could also be used to match seg-ments detected from the archival orthoimage to the roadsand buildings from topographic databases. Such ground

control lines could then be integrated in the bundle ad-justment. Eventually, the planimetric accuracy obtainedcan be consistent enough to train land-cover classificationmodels using recent land-cover databases. The automaticGCPs could then be labeled with a rather coarse land-cover class (e.g., vegetation/non-vegetation), and conse-quently filtered out for the final adjustment.

Références[1] Aguilar, M., Aguilar, F., Fernández, I. and Mills,

J., 2013. Accuracy assessment of commercial self-calibrating bundle adjustment routines applied to ar-chival aerial photography. The PhotogrammetricRecord 28(141), pp. 96–114.

[2] Asner, G., Archer, S., Hughes, R., Ansley, R. J. andWessman, C., 2003. Net changes in regional woodyvegetation cover and carbon storage in Texas Dry-lands, 1937–1999. Global Change Biology 9(3),pp. 316–335.

[3] Aubry, M., Russell, B. C. and Sivic, J., 2014.Painting-to-3D model alignment via discriminativevisual elements. ACM Transactions on Graphics33(2), pp. 14 :1–14 :14.

[4] Ayoub, F., Leprince, S. and Avouac, J.-P., 2009. Co-registration and correlation of aerial photographsfor ground deformation measurements. ISPRS Jour-nal of Photogrammetry and Remote Sensing 64(6),pp. 551–560.

[5] Cardenal, J., Delgado, J., Mata, E., González, A.and Olague, I., 2006. Use of historical flightfor landslide monitoring. In : Spatial Accuracy,pp. 129–138.

[6] Chen, H.-R., Tseng, Y.-H. et al., 2016. Study of au-tomatic image rectification and registration of scan-ned historical aerial photographs. InternationalArchives of the Photogrammetry, Remote Sensingand Spatial Information Sciences 41(B8), pp. 1229–1236.

[7] Cléry, I., Pierrot-Deseilligny, M. and Vallet, B.,2014. Automatic georeferencing of a heritage of oldanalog aerial photographs. ISPRS Annals of Pho-togrammetry, Remote Sensing and Spatial Informa-tion Sciences II-3, pp. 33–40.

[8] Dalal, N. and Triggs, B., 2005. Histograms of orien-ted gradients for human detection. In : CVPR.

[9] Ellis, E. C., Wang, H., Xiao, H. S., Peng, K.,Liu, X. P., Li, S. C., Ouyang, H., Cheng, X. andYang, L. Z., 2006. Measuring long-term ecologicalchanges in densely populated landscapes using cur-rent and historical high resolution imagery. RemoteSensing of Environment 100(4), pp. 457–473.

[10] Ford, M., 2013. Shoreline changes interpreted frommulti-temporal aerial photographs and high resolu-tion satellite images : Wotje atoll, marshall islands.Remote Sensing of Environment 135, pp. 130–140.

[11] Fox, A. J. and Cziferszky, A., 2008. Unlocking thetime capsule of historic aerial photography to mea-sure changes in antarctic peninsula glaciers. ThePhotogrammetric Record 23(121), pp. 51–68.

[12] Galster, J. C., Pazzaglia, F. J. and Germanoski, D.,2008. Measuring the impact of urbanization onchannel widths using historic aerial photographsand modern surveys. Journal of the American WaterResources Association 44(4), pp. 948–960.

[13] Giordano, S., Le Bris, A. and Mallet, C., 2017.Fully automatic analysis of archival aerial imagescurrent status and challenges. In : Joint Urban Re-mote Sensing Event (JURSE).

[14] Hermosilla, T., Wulder, M., White, J., Coops, N.and Hobart, G., 2015. Regional detection, charac-terization, and attribution of annual forest changefrom 1984 to 2012 using landsat-derived time-seriesmetrics. Remote Sensing of Environment 170,pp. 121–132.

[15] Jao, F.-J., Chu, H.-J. and Tseng, Y.-H., 2014. His-torical image registration and land-use land-coverchange analysis. Environments 1(2), pp. 181–189.

[16] Kadmon, R. and Harari-Kremer, R., 1999. Studyinglong-term vegetation dynamics using digital proces-sing of historical aerial photographs. Remote Sen-sing of Environment 68(2), pp. 164–176.

[17] Kim, S., Min, D., Ham, B., Ryu, S., Do, M. andSohn, K., 2015. DASC : Dense Adaptive Self-Correlation descriptor for multi-modal and multi-spectral correspondence. In : CVPR.

[18] Korpela, I., 2006. Geometrically accurate time se-ries of archived aerial images and airborne lidardata in a forest environment. Silva Fennica 40(1),pp. 109.

[19] Lowe, D., 1999. Object recognition from localscale-invariant features. In : ICCV, pp. 1150–1157.

[20] Meyer, P., Staenz, K. and Itten, K., 1996. Semi-automated procedures for tree species identificationin high spatial resolution data from digitized colourinfrared-aerial photography. ISPRS Journal of Pho-togrammetry and Remote Sensing 51(1), pp. 5–16.

[21] Micheletti, N., Lane, S. and Chandler, J., 2015. Ap-plication of archival aerial photogrammetry to quan-tify climate forcing of alpine landscapes. The Pho-togrammetric Record 30(150), pp. 143–165.

[22] Mondino, E. and Chiabrandoa, R., 2008. Multi-temporal block adjustment for aerial image timeseries : The Belvedere glacier case study. Inter-national Archives of the Photogrammetry, RemoteSensing and Spatial Information Sciences 37(B2),pp. 89–94.

[23] Nagarajan, S. and Schenk, T., 2016. Feature-basedregistration of historical aerial images by area mi-nimization. ISPRS Journal of Photogrammetry andRemote Sensing 116, pp. 15–23.

[24] Nebiker, S., Lack, N. and Deuber, M., 2014. Buil-ding change detection from historical aerial pho-tographs using dense image matching and object-based image analysis. Remote Sensing 6(9),pp. 8310–8336.

[25] Necsoiu, M., Dinwiddie, C., Walter, G., Larsen, A.and Stothoff, S., 2013. Multi-temporal image ana-lysis of historical aerial photographs and recent sa-tellite imagery reveals evolution of water body sur-face area and polygonal terrain morphology in Ko-buk Valley National Park, Alaska. EnvironmentalResearch Letters 8(2), pp. 025007.

[26] Nocerino, E., Menna, F. and Remondino, F., 2012.Multi-temporal analysis of landscapes and urbanareas. International Archives of the Photogram-metry, Remote Sensing and Spatial InformationSciences 39(B4), pp. 85–90.

[27] Nurminen, K., Litkey, P., Honkavaara, E., Vasta-ranta, M., Holopainen, M., Lyytikäinen-Saarenmaa,P., Kantola, T. and Lyytikäinen, M., 2015. Automa-tion aspects for the georeferencing of photogram-metric aerial image archives in forested scenes. Re-mote Sensing 7(2), pp. 1565–1593.

[28] Pierrot Deseilligny, M. and Cléry, I., 2011. Apero,an open source bundle adjusment software for au-tomatic calibration and orientation of set of images.International Archives of the Photogrammetry, Re-mote Sensing and Spatial Information Sciences38(5/W16), pp. 269–276.

[29] Song, X.-P., Sexton, J., Huang, C., Channan, S. andTownshend, J., 2016. Characterizing the magnitude,timing and duration of urban growth from time se-ries of Landsat-based estimates of impervious co-ver. Remote Sensing of Environment 175, pp. 1–13.

[30] Véga, C. and St-Onge, B., 2008. Height growth re-construction of a boreal forest canopy over a periodof 58 years using a combination of photogramme-tric and lidar models. Remote Sensing of Environ-ment 112(4), pp. 1784–1794.

[31] Verdié, Y., Yi, K., Fua, P. and Lepetit, V., 2015.TILDE : A Temporally Invariant Learned Detector.In : CVPR.

[32] Walstra, J., Chandler, J., Dixon, N. and Dijkstra, T.,2004. Time for change-quantifying landslide evolu-tion using historical aerial photographs and modernphotogrammetric methods. International Archivesof Photogrammetry, Remote Sensing and Spatial In-formation Sciences 35(B4), pp. 475–480.

[33] Yi, K., Trulls, E., Lepetit, V. and Fua, P., 2016.LIFT : Learned Invariant Feature Transform. In :ECCV.

[34] Zambanini, S. and Sablatnig, R., 2017. A Hough vo-ting strategy for registering historical aerial imagesto present-day satellite imagery. In : ICIAP.