Skip to navigation – Site map

HomeSectionsCartography, Images, GIS1996Evaluation of natural surfaces in...

1996
13

Evaluation of natural surfaces in a dry tropical environment by remote sensing and ground survey

Evaluation des surfaces naturelles par télédétection et enquête de terrain en milieu tropical sec
Vincent Godard

Abstracts

A study was conducted in a Sahel pastoral area in Mauritania from 1987 to 1989 which involved the application of satellite sensor images to ground surveys, to determine the surface areas of different types of land occupancy and in particular those suitable for raising livestock. In the preparation phase, satellite sensor images were used in conjunction with the sampling plan that monitored the land occupancy inventory. During the ground survey phase, images representing enlarged survey units were used to complement the survey records. Finally, in the analytical phase, remote sensing was used not only to draw ground inventory maps but also to correct, by regression, the estimation of the surface areas of the various categories of land occupancy detected. The findings of our study show that remote sensing, while allowing for a reduced sample size during the ground survey phase, provides for a high level of accuracy in the estimation of surface areas.

Top of page

Author's notes

Il existe une version approchante en français de cet article sous la référence : GODARD (V.) - 1992 - Évaluation des surfaces naturelles par télédétection et enquête de terrain en Mauritanie. STATECO, n° 71, septembre 1992 : 27-58.

Full text

This investigation was made possible thanks to he Institut d'Elevage et de Médecine Vétérinaire des pays Tropicaux (IEMVT), the Breeding Department of the Centre de Coopération Internationale en Recherche Agronomique pour le Développement (CIRAD), and of the Ministère français de la Recherche et de l'Enseignement Supérieur (MRES). The autors would also like to thank Michèle GIBOIRE and Elizabeth VILLERE for their patience in the editing and translation of this article.

© CYBERGEO 1996

Introduction

1In order to estimate the surface area of plant formations covering several hundreds or several thousands square kilometers, an extensive knowledge of the different land cover is necessary. Since it is not possible to survey and explore such vast areas in their entirety, sampling and then extrapolating the results obtained to the whole in question area is the most efficient method to use. Remote sensing is a valuable tool in this process since it provides a global, instantaneous and exhaustive view.

2Many researchers have already applied remote sensing to problems similar to ours, for agroecological maps (Bruneau 1986, Pouyllau 1990), pastoral resources maps (De Wispelaere 1991). However, they did not use a probabilistic ground sampling plan. Remote sensing cannot be used alone to estimate surface area of various categories of land cover. It is essential to design a ground sampling scheme representative of the entire area under study, otherwise the extrapolation of the results to the area will be diminished by an unknown (though real) error, detrimental to any further use of the inventory results. A ground survey, like any land cover inventory, is the basis of data collection, and remote sensing provides complementary information, thus improving the surface area estimations of the various categories resulting from the ground survey.

3In Mauritania, and others countries of the Sahel, reliable statistics on plant formations are essential for raising livestock. The people in charge of land managment lack of these types of statistics. Our task was to devise and implement a method using ground surveys combined with satellite sensor data to estimate surface areas of plant formations, so that eventually the Sahel people will be able to adapt livestock density to the available resources and to avoid over - or under-grazing.

4In the United States and somes countries of the European Union, semi-operational programs are carried out using this methodology for agricultural environments (Houseman 1979, Bryant 1983, Fournier 1986, IAT 1989). There is, however, no such program in tropicals areas.

The Study area

5With slightly more than one million square kilometers, the Mauritanian Islamic Republic (RIM) is one of the largest countries on the African continent, and also one of the most arid. Two thirds of the territory are part of the Sahara Desert. Only the southern fringe, located in the Sahel, is outside the hyper-arid area and allows the inhabitants such activities as animal breeding and, rarely, agriculture.

6Our investigation was carrried out in this southern region, inside three test areas (Figure 1) -- one located near Rosso (region of Trarza), another near Kaedi (region of Gorgol) and the last near Kiffa (region of Assaba). Only the results of Rosso are presented in this paper.

7Our field of investigation can be defined as follows (Leroux 1983 p.578). In general, annual rainfall is between 100 and 500-600 mm. Summers are hot and characterized by a good deal of evaporation, and winters are cool. The maximum temperatures are experienced during the intervening periods. The aerological environment is equally divided between the Harmattan in winter and the monsoon in summer. However, as in many other Sahel regions, the climate has been getting more arid in the last 30 years, as is clearly shown by the rainfall records.

Figure 1 - Location of the investigation areas

8The rainy season (from July to September) accounts for 90% of the rainfall (Godard 1991 p.28). As our inventory concerned only vegetal formations, we wanted to investigate plant growth at the stage of maximum development before it withered. The optimum dates for images and investigations were just before mid-October (Godard 1991 p.51).

9In these regions of the north Sahel, the vegetation is a steppe "strengthened" by loose shrubbery and discontinuous plants (Rougery 1988 p.72). Thicker, thorny bushes can sometimes be found in the sloping areas with favourable water conditions such as the wadi beds. Typically the trees and shrubs are from 3 to 6 m tall, with a significant majority of thorny leguminous plants such as acacias. The herbaceous stratum is dominated by perrenial graminaceae of the Aristida and Cenchrus genus.

10Because of the scarcity of rainfall, rain-dependent crops are absent from the Rosso pilot area. Only subsidence crops or irrigated crops exist in sufficient quantities in this region to be part of the sampling plan. However, because of the dates of the images, only the irrigated crops were visible on the data used. The subsidence crop areas were still flooded at the time (October and November 1987) and could be identified only during the control phase in the dry season (March and April 1988).

11Located in the cenosoic Senegalo-Mauritanian basin, outcrops are mainly quaternary in the Rosso pilot area. The dieri, a Toucouleur word describing the areas that are never flooded by the Senegal river, is primarily covered by the red sand dunes of the "Ogolian". Similarly, the wallo, the high water bed, is made up of elements rising in tiers between the Nouakchott terraces and the current or sub-current deposits (Michel 1973).

12The Rosso area is therefore made up of dunes and interdunes oriented NE-SW and, along the river, of mostly argilaceous and hollow areas. Outside the rainy periods which temporarily flood the interdunes and revive the wadi, there is no real difficulty in moving about the areas, nor in reaching the investigation units.

Method

Preparation of the ground survey

13For the preparation of the ground survey, we based our sampling scheme on an analogical satellite sensor image (SPOT1 HRV2) recorded on November 2, 1986. Therefore, it was impossible to take spatial autocorrelation into account. Though systematic sampling gives a better assessment of the entire area of investigation (Webster 1989 p.68), we used non-aligned systematic sampling so as to avoid errors from periodic phenomena in the landscape, but did not use stratification.

14We decided to adapt ground survey methodology developed by Ph. Fournier, as there were no existing references for tropical environments. Since it is impossible to locate accuratly sample pixels on the field, surfaces should be sampled instead of points (Fournier 1985 p.11). We therefore selected one-level sampling, wich consists in selecting at random the coordinates of a pixel, delineating surface developed from this pixel and investigating the surface. The resulting surface, the survey unit, is a square with fixed dimensions, called a segment, whose size must be defined. The advantage of probabilistic sampling is that it gives an estimation of the surface area of each land cover category, and more importantly, gives an idea of the accuracy of the estimation.

15Without information on spatial autocorrelation, we chose a sampling rate between 1 and 2% (Fournier 1986 p.14) since it allows for a good representation of the study area. Also, given time constraints, it was adviseable to allot one half-day maximum per segment for the ground survey (Fournier 1985 p.11). Therefore, we chose segments of 36 ha, squares, with a side-length of 600 m.

16Since we used non-aligned systematic sampling, we were able to position a grid randomly on the satellite sensor image of the survey area. Given a grid scale equal to ten times the length of the side of a segment (10 x 600 m), sampling one segment per cell is equivalent to sampling at a 1% rate, two segments per cell, 2%, etc. All the cells of the grid are arranged in groups of four, each group making up a block. The x and y coordinates of the segment are different for each cell of a block but identical from one block to another (Figure 2). The sampling is named as systematic and non-aligned.

Figure 2 - Location of the segments

Implementation of the ground survey

17Once the sampling scheme was defined, the ground inventory could begin. In the field, the investigator was provided with 1:50,000 scale color maps showing the composition of the survey area, on which the segments were drawn. The investigator had also 1:10,000 scale color enlargements showing the color composite. These enlarged prints complement the record card, which made it possible to locate accurately every element of land cover investigated.

18The inventory was carried out on natural or cultivated plots found inside the segment. On the record card, the investigator sketched the plots and noted what type of land cover was found on each plot. This is not always easy to do in a natural environment since there is no clear delineation of the different land cover categories. The proportion of each category per segment was then calculated (Godard 1991 pp.275-279) and used as a ground reference to control the classifications and to calculate the first estimations of surface area, even before the classifications monitored by the ground survey had been carried out. An abstract of the developed method is proposed (Figure 3).

Figure 3 - Synopsis of the method

PROCESSING AND ANALYSIS OF DATA

- Surface area estimates with ground survey and without correction by regression

19The following formulas were developed by the Service Central des Enquêtes et Études Statistiques (SCEES) for temperate agricultural environments (Fournier 1986) but are also valid for tropical climates. The formulas can be used in two particular cases: either to provide provisional estimations very quickly, or if there is no recent satellite sensor data, to correct the surface area estimations resulting from the ground survey.

20- Surface area estimation of category k

21If satisfactory digital satellite sensor data are available, it is possible to improve the estimations of surface area and their accuracy with supervised classification. In order to be comparable with the landscape that was investigated, the satellite sensor data used in the analysis must be recorded at the same season as the data collected during the survey phase in the field. Unless the landscape has changed significantly, the data can of course have been recorded a few years before the mission. We used level 1B SPOT1 HRV2 (KJ 023-317) data recorded on November 6, 1987.

Surface area estimates using regression

22The classifications used for making maps of land cover are of the "maximum likelihood" type, and are two-fold. First they provide a visual document, the map, that makes it possible to immediately locate the spatial distribution of the various categories. Second, they provide the mapped surface area of each of the categories. The surface areas, obtained through classification, provide statistics that can be used to correct the estimations obtained in the field.

23Although we did not intend to discuss the mapping aspect in this paper, we should add that the quality of classifications is very important, so as to provide a satisfactory correction of the estimations of surface area. As discussed later, the accuracy of estimations is greatly increased, with an overall accuracy of the classifications close to 87%. The overall accuracy of the classifications is obtained by dividing the sum of the properly classified pixels, measured on the diagonal of the matrix of confusion, by the sum of the verified pixels. Other indices read on the matrix of confusion were used to select the optimum classifications, then to determine their strong and weak points, category by category (Godard 1991 pp.300-306).

24- The regression estimate

25The formulas we will now present come for the most part from the Théorie et pratique des sondages (Desabie 1966 pp.215-219). Unlike the previous formulas, these formulas are more complexe. In addition to the proportion of land cover estimated during the ground survey (the proportion being actually converted into pixels), the formulas include statistics obtained through classification for both the segment level and the survey area level.

26It is not possible to use the results of the supervised classifications as reliable land cover statistics, since classification is never perfect, and a direct estimation of the surface areas on the map is altered by an unknown bias due to the selection of the initialisation plots during classification and to the classification method used (Meyer-Roux 1981 pp.10-12). However, information provided by the satellite sensor is indispensable since it takes into account the whole study area.

27A linear regression is carried out between the ground data, category by category at segment level, and the homologous data obtained by classification, also at the segment level. The regression estimation factor allows to take into account the behaviour of the sample in relation to the whole area, relative to the satellite sensor data, and to correct the estimation resulting from the ground survey alone.

28- Surface area regression estimate Yreg k at segment level

29- Surface area regression estimate Yreg k for the pilot area

30- Variance of the regression estimate at the segment level

31The variance of the regression estimate is calculated to determine the accuracy of this estimation. As mentioned by J. Meyer-Roux (Meyer-Roux 1981 p.12): "If Xk has no variance, Yreg k does have one, since the line of least squares used to define it is only an estimation of the unknown real line.". Actually, Xk is calculated on the classification, at the level of the pilot area, and therefore has no variance.

32The variance associated with estimation can be determined as follows:

33where

34It should be noted that the above variance estimation formula applies to large samples. Our study, however, only covered 20 segments per pilot area, and perhaps we should have used a variance estimation formula adapted for a reduced sample size. For brevity's sate, we will only give the formulas and results for larger samples. This method is intended to be used in areas larger than our pilot areas, using a sample of at least 60 segments.

35However, when comparing the coefficients of variation factors of both estimation methods, the results are not very different. The variance estimation calculated for smaller samples will of course give less accurate results.

36- Variance of the regression estimate Yreg k at the level of the pilot area

Results

37The first results presented are those obtained at the end of the ground survey, without rectification by satellite sensor digital data. The simplest way to estimate the accuracy of the results is to analyse the coefficients of variation (CV). The coefficient of variation is obtained by dividing the standard deviation by the surface area estimated, then converting the quotient into a percentage, which can be used to evaluate and compare the relative accuracy of each category.

38We have divided the coefficients of variation into three groups. The first group contains the categories whose coefficient of variation is from 10% to 30%. In Rosso, this corresponds to bare sandy, loamy soils. They cover an area generally larger than 20% of the pilot area. The second group contains categories with a coefficient of variation of about 50%. This corresponds to bare clayey soils. Whose surface area is less than 10% of the pilot area. Lastly, some categories of land cover have a coefficient of variation larger than 75%. Their surface areas seems generally to be less than 1%.

Table 1 - Surface areas estimated by using the ground survey without correction by regression

39An analysis of these results, without correction by regression, shows that the sampling plan we used is satisfactory only for the categories with a large surface area, typically more than 20% of the pilot area. For the categories with a surface area of less than 5%, the estimations are too inaccurate. However, they give a fair enough idea to make the provisional estimations at the end of the inventory, which can be useful. The problem is completely different when recent satellite sensor digital data, or even data of the same date as the ground survey, is not available to correct the estimations. In this case, one has to make do with only these estimations, which may be far from the optimal.

Table 2 - Surface areas estimated by regression

40However, when satellite sensor digital data that is recent enough to correct the estimations of the ground survey is available, the accuracy of the results is quite different. It can be seen that a little less than 5% of the surface area of the pilot area is not allocated at the end of the calculation of the estimations by regression. As a rule, for publication other than research, this remainder associated with the calculation method is distributed among all the existing categories.

41As can be seen from the above table, the relative accuracy, as shown by the coefficient of variation, has increased noticeably. Without the regression factor, there was no category with a coefficient of variation lower than 15%, but there are now four categories (Table 2). Similarly, there is only one category left with a coefficient of variation higher than 30%: the bushy steppes of the sand banks.

42The rather low accuracy of this category is caused primarily by inaccurate mapping (Godard 1991 p.303). When mapping is used to correct the estimations obtained from the field, the increase in accuracy is insufficient. The coefficient of variation decreases from 60.9%, without regression (Table 1), to 31.5% with regression (Table 2).

43Similarly, if we compare the overall increase in accuracy due to the use of regression with the use of satellite sensor data to correct the ground survey estimations, we find that the coefficients of variation have considerably diminished. If we calculate the average of the coefficient of variation, we find it decreases from 61.5% without regression to 14.5% with regression.

44Categories that have benefited the most are those for which there is a good correlation between the terrain and the classification and, particularly interesting in this case, the categories with a small surface area, which the ground survey may have over-or underrepresented, due to a low sampling rate. An example is the watered bushy steppes, whose coefficient of variation decreases from 130.4% (Table 1) to 19.9% for a surface area of 1.8% (Table 2), an surface area largely under-estimated by the ground survey.

45By using remote sensing, we have been able to readjust the estimations of the surface area of irrigated crops. At the end of the ground survey, irrigated crops were said to comprise 2.7% of the Rosso pilot area, with -- true enough -- a coefficient of variation of 90% (Table 1), However after correction by the satellite sensor data, the surface area estimated was reduced to 0.7% (CV of 16.4%, Table 2). Remote sensing therefore makes it possible to attenuate the consequences of a sampling plan ill adapted to categories of small surface area, not representative of the landscape, but very significant in terms of resources.

46As it is possible to verify the accuracy of a sampling by strata compared to elementary sampling, it is also possible to verify the accuracy of estimations by regression compared to an "elementary" estimation. If we define an elementary estimation without recourse to remote sensing, we consider that the efficiency Ef of remote sensing can be measured by the following ratio (Meyer-Roux 1981 p.13):

47where r is the coefficient of correlation obtained by regression between ground data and classification.

48An efficiency Ef of 3.6 corresponds to a coefficient of correlation of 0.85, that of the grassy steppes of sand banks and interdunes (Table 3). This means that without regression, 3.5 times as many segments would be needed to make an inventory of the category and to obtain the same relative accuracy: 94.4%, corresponding to a coefficient of variation of 15.6%.

Table 3 - Efficiency of remote sensing

49We find that (Table 3), depending on the coefficient of correlation r, the improvement in accuracy for 20 segments, as compared to elementary sampling, ranges from 1.5 times for the watered bushy steppes to 50 times for categories such as flooded surface areas or steppes with gum-trees. The average accuracy is 28 times higher than without correction by regression for all categories. The average coefficient of correlation (r) is 0.90. In other words, without using satellite sensor data to correct estimations, 28 times more segments would have been necessary to obtain the same level of accuracy.

50The high level of efficiency of the method, as compared with results obtained in a temperate agricultural environment, may come as a surprise. By comparison, the results obtained by R. Pastorelli and J.C. Porchier (Pastorelli 1989 p.55) concerning land cover in Indre (a French "département") are less effective, and their coefficients of correlation are generally lower. According to the autors, the explanation may be due to the fact that the land cover nomenclature is too precise compared to the possibilities of radiometric discrimination on the date that the imagery was recorded, i.e. too late in the season.

51In conclusion, when analysing either the coefficients of variation or the efficiency of an estimation by regression against an rudimentary estimation (without correction by remote sensing), the contribution of satellite sensor digital data is obvious. This data is essential to the simplification of the sampling plan and the improvement of the accuracy of the estimation of surface areas.

Conclusion

52To correct the estimations of surface areas obtained by probabilistic sampling of the ground, using satellite sensor digital data has many advantages. The number of survey units to be investigated can be significantly reduced while the accuracy of the estimations is much improved.

Figure 4 - Performing the inventory

53When the relative accuracy of the two estimation procedures is compared in respect to their coefficients of variation, the average gain in accuracy is significant. The coefficients of variation decrease on an average from 61% to 14%, since it is possible to correct estimations by regression. In terms of efficiency, on an average, 28 times more segments would have been necessary to obtain the same accuracy if satisfactory satellite sensor digital data had not been available.

54Obviously, the 20 segments per survey area we were able to use are too few when satellite sensor digital data is not available. With a few exceptions, the accuracy is much too low to be satisfactory. In addition, a small sample entails other problems which must be borne in mind. Remote sensing, on the one hand, can show only what is visible through radiometric measurements and, on the other hand, what the ground survey has identified. A small sample, therefore is very likely to show only the main categories in the landscape. The small size of some categories can be balanced by a very characteristic radiometry which will indicate an "omission" in the sampling plan. This applies to categories like irrigated crops; as their spectral behavior is very distinct, they can easily be located. This is less true for categories which, though not common, are only a variation of their surroundings.

55In 1987, our study was the first application of segment survey combined with remote sensing in a dry tropical environment. As we had no digital data to realize our sampling plan, we could not determine the size of segments through analysis of spatial autocorrelation of satellite sensor data. Many geostatistical analysis were carried out on remote sensing data with variograms (Godard 1994) and the optimal size of segment seems to be between 200 and 300 m. A segment of 600 m sides gives to much work without a real accuracy gain in. Therefore, it is possible to obtain the same results in less time.

56We would also advise, in operational ground survey phase, a sample of 30 to 60 segments per pilot area. This makes it possible, if the number of categories and their distribution permit, to obtain fairly accurate estimations for the major categories when satellite sensor images are not available. A larger sample size also makes possible a larger selection for the initialisation and verification of classification. We would also recommend consulting an automatic classification, so as to evaluate partially the homogeneity and the size of the parcels before conducting the ground survey, which gives a good preliminary description of the thematic overlapping at the segment level. A GPS-type satellite locating tool can be very helpful in the field to get one's bearing and to actually find the segment.

Top of page

Bibliography

BRUNEAU M., KILLIAN J., PAIRINTRA C., SAVARY G. - 1986 - Identification et dynamique des milieux agricoles dans le nord-est de la Thaïlande (Udon Thani - Sakon Nakhon). Talence, CEGET-CNRS, 75 p. (Travaux et Documents de Géographie Tropicale n°54, Série Cartosat n°1).

BRYANT C.R., RUSSWURM L.H. - 1983 - Les méthodes d'échantillonnage aréolaire selon les besoins et objectifs de la surveillance de l'utilisation des terres. Environnement Canada, Direction Général des terres, 76 p. (doc. de travail n°24).

DESABIE J. - 1966 - Théorie et pratique des sondages. Paris, Dunod, 480 p.

DE WISPELAERE G., PEYRE de FABREGUES B. - 1991 - Évaluation et suivi des ressources pastorales par télédétection spatiale dans la région du Sud-Tamesna (Niger). Maisons-Alfort, IEMVT, 93 p. (+ 1 vol. d'annexes).

FOURNIER Ph. - 1986 - Enquête sur l'utilisation du territoire effectué en 1985 par la méthode des segments. Paris, SCEES, Série S, 63 p. (Méthodes et Applications Scientifiques n°13).

FOURNIER P., MEYER-ROUX J. - 1985 - Observations terrain et télédétection. Cahiers de Statistique Agricole. n°1/6 (janvier-février) : 5-11.

GODARD V. - 1991 - Utilisation conjointe de la télédétection et de l'enquête de terrain lors des inventaires d'occupation du sol. Recherche appliquée au Sahel sud-mauritanien. Maisons-Alfort, IEMVT. Thèse de doctorat de l'EHESS, 433 p.

GODARD V. - 1994 - Apport de l'analyse variographique pour déterminer la taille et l'espacement des unités d'échantillonnage lors d'un inventaire d'occupation du sol en milieu naturel. Bulletin de la SFPT. 136 (1994-4): 33-44.

HOUSEMAN E. - 1979 - L'échantillonnage d'aires dans l'agriculture. Washington DC, Statistical Reporting Service, US Dept. of Agric. 88 p. (SRS n°20).

IAT - 1989 - Inventaires régionaux. Flash Télé-Agri-News. Inst. des Applic. de la Téléd., Service Spécialisé Agriculture (CCRCCE), Ispra (Italie), 2/89 : 2-15.

LEROUX M. - 1983 - Le climat de l'Afrique tropicale. Paris, ed. Champion, 633 p.

MEYER-ROUX J., SIGMAN R., CRAIG M., OZGA M. - 1981 - Cartographie et statistique en télédétection. - Cahier de Statistique Agricole. n°6/6 (novembre-décembre) : 9-14.

MICHEL P. - 1973 - Les bassins des fleuves Sénégal et Gambie. Etude géomorphologique. Paris, ORSTOM, 3 tomes, 752 p. + cartes (Mémoire ORSTOM n°63).

PASTORELLI R., PORCHIER J.C. - 1989 - Premiers résultats sur l'utilisation des terres à l'échelle départementale à partir des données du satellite SPOT. Cahiers de Statistique Agricole. n°3/6 (mai-juin) : 51-56.

POUYLLAU M., LE MEN H., POUYLLAU D., et al. - 1990 - Cartographie agro-écologique des llanos occidentaux du Venezuela à petite et moyenne échelle : paysages et évolutions 1979, 1982, 1987 (Landsat et SPOT). Actes de la table ronde internationale : Apport de la télédétection spatiale à l'étude des paysages et des systèmes agraires tropicaux. CEGET-CNRS, Talence 25-27 octobre 1989. Bordeaux-Talence, CNRS : 23-46.

ROUGERY G. - 1988 - Géographie de la biosphère. Paris, Armand Colin, 288p.

WEBSTER R., CURRAN P.J., MUNDEN J.W. - 1989 - Spatial Correlation in Reflected Radiation from the Ground and Its Implications for Sampling and Mapping by Ground-Based Radiometry. Remote Sensing of Environment. 29 : 67-78.

Top of page

References

Electronic reference

Vincent Godard, « Evaluation of natural surfaces in a dry tropical environment by remote sensing and ground survey  », Cybergeo: European Journal of Geography [Online], Cartography, Images, GIS, document 13, Online since 14 December 1996, connection on 28 March 2024. URL : http://journals.openedition.org/cybergeo/5265 ; DOI : https://doi.org/10.4000/cybergeo.5265

Top of page

About the author

Vincent Godard

Université de Paris 8 & Centre de Biogéographie-Écologie, UMR 180 CNRS, ENS Fontenay-St-Cloud, Le Parc, 92211 Saint-Cloud Cedex, France
vgodard@univ-paris8.fr

By this author

Top of page

Copyright

CC-BY-4.0

The text only may be used under licence CC BY 4.0. All other elements (illustrations, imported files) are “All rights reserved”, unless otherwise stated.

Top of page
Search OpenEdition Search

You will be redirected to OpenEdition Search