ETUDE COMPARATIVE DE DEUX APPROCHES DE CARTOGRAPHIE DE L’OCCUPATION DU SOL EN MILIEU RURAL : PAR PIXEL ET PAR OBJET - HAL-SHS - Sciences de l'Homme et de la Société Accéder directement au contenu
Article Dans Une Revue Photo-Interprétation. European Journal of Applied Remote Sensing (PIEJARS) Année : 2018

LAND COVER MAPPING FROM SATELLITE SPOT 5 XS - HRG1 IMAGE, COMPARATIVE STUDY FROM TWO CLASSIFICATION METHODS : PER-PIXEL AND OBJECT-BASED

ETUDE COMPARATIVE DE DEUX APPROCHES DE CARTOGRAPHIE DE L’OCCUPATION DU SOL EN MILIEU RURAL : PAR PIXEL ET PAR OBJET

Résumé

Extended Abstract: Appeared in the 1970s, the per-pixel classification approach has been widely used as a method of mapping land cover using satellite imagery. Many studies have thus been carried out by the application of this approach. In Côte d'Ivoire, the per-pixel approach is the most used by academic researchers for mapping and analysis of the spatio-temporal dynamics of land cover. With the advent of very high spatial resolution satellite images, a new approach to mapping has been developed since more than 10 years. This approach, referred to as "object-based", has been tested and found to be more efficient and promising by many researchers throughout the world. However if this approach is much used in some countries, it is not the case in many African countries, particularly in Côte. This study therefore proposes to explore and test the object-based approach in a case study. In addition, we compared the obtained result to that of a classical per-pixel classification, using as input the same satellite image, namely SPOT5. The sector of Eloka-Palmeraie, a very heterogeneous space composed of a diversity of plantations and forests, is the experimental framework for this study. It is located to the east of Abidjan, between latitudes 5°16'11" and 5°22'21" North, and longitudes 3°50'39 and 3°44'36" West (Plate 1). This space enjoys a very wet subequatorial climate leading to the development of a very dense forest vegetation with a notable presence of mangrove and very dense hydromorphic vegetation formations, composed of swampy forests. The resident population is in majority rural and its main activity is agriculture. The two classification approaches (by "object" and by "pixel") have been used for the mapping of the land cover of this area. The mapping of the land cover by object approach has been implemented using different methods of segmentation, classification and validation, and then back by modifying the parameters up to reach a satisfactory result. This way, several levels of segmentation have been made; but at each level of segmentation performed, a visual assessment of the result is done in order to check the conformity of the segmentation with the typology previously established. Therefore after several combinations, parameters as follow: scale 30, form 0.3; color 0.7; compacity and smoothing 0.5, have been retained. This combination of settings has resulted in a segmentation of the image into "objects-image" more or less consistent with our knowledge of the reality and with the field data. After the phase of segmentation, followed the step of the classification of objects-image created previously by segmentation. For validation purpose, 100 sample points (including 10 by theme) were first selected and then validated in the field. The choice of these samples has been done by visual analysis of the image followed by a field campaign. Subsequently, a "buffer" of 5 meters has been drawn around each sample point. The vector layer (polygon) obtained has been coded and imported in eCognition to automatically generate sample objects which constitute the basic elements of the classification of objects. The “maximum likelihood" was the algorithm used for the classification of objects-image. This algorithm affects each object-image to a type of land cover on the basis of the samples formed. Once the classification was completed, the objects constitutive of each class of land cover have been merged into more homogeneous classes. Concerning the mapping by the pixel approach, we used a supervised classification with the same classification algorithm " maximum likelihood". Also, the training samples used for the classification of objects-images have also been used for the classification by the pixel approach of the SPOT 5 image. After the different steps of classification of the image by the application of the two approaches, the evaluation of the results has subsequently been made before their validation. This evaluation has been made using the data collected in field. These sample data are different from those used during the classification process. This assessment has led to the calculation of indicators of performance of a classification that are the overall accuracy and the kappa index. Thus, with the overall accuracy of classification of 91, 56% and a kappa of 91% (Table 1), the result of the approach by object appears much better than that of the approach by pixel, whose overall accuracy of classification and the index of Kappa are respectively 78,67% and 81% (Table 2). These results demonstrate that the classified image, outcome of the approach by object presents a level of higher compatibility with the field data than the approach by pixel. This good performance of the approach by object is reflected visually by a cartographic result more homogeneous, more developed and also fits to a very high level with the field data; the result of the approach by pixel presents a very high level of heterogeneity (Table 2). In the detail, several differences and disparities have also been observed at the level of the land cover categories. Thus, the class "swampy forest" with the overall accuracy of 95%, appears more homogeneous at the level of the object approach than that per pixel which is very heterogeneous with a precision of 75% (Table 3). This situation is explained by the fact that several pixels of the class "swampy forest" have thus been assigned to other classes at the time of the classification by the approach by pixel. It is the same with the classes "food crop", "bushes / fallow", “residual forest" whose accuracy of classification remain low and far below those of the approach by object. The very low accuracy of classification (less than 65%) of the classes "bare soil /plowed fields" and " built areas" shows that many confusions occurred in the assignment of the pixels to these classes in the classification by the approach by pixel, which is not the case for the approach by object for which the accuracy of classification are respectively 85% and 90%. Furthermore, the superimposition of the result of classification of the approach by pixel, to that of the object approach suggests that only 65% of the objects in the study area are identical (Table 4). This gap in correspondence indicates that the two approaches lead to the production of spatially different cartographic results. In summary, taking into account in the approach by object the textural, shape and neighborhood parameters in addition to the spectral signature contribute to the good performance of this approach. Thus, mapping using the approach by object appears as a kind of automation of visual interpretation conducted by photo-interpretation. However, if this approach allows to obtain a very good cartographic result, its implementation remains, very complex, laborious and time consuming.
Apparue dans les années 1970, l’approche pixel par pixel a pendant longtemps été utilisée comme unique méthode de cartographie de l’occupation du sol à partir des images-satellites.. Depuis une décennie, une nouvelle approche de cartographie a vu le jour avec l’avènement des images satellitaires à très hautes résolution spatiales. Pour traiter ces images, cette nouvelle approche nommée "orientée-objet", est jugée plus performante et prometteuse par plusieurs auteurs. Malheureusement, elle est très peu utilisée dans les études de cartographie de l’occupation du sol en Côte d’ivoire. Cette étude se propose d’explorer cette nouvelle approche de cartographie et de comparer le résultat obtenu à celui d’une classification classique pixel par pixel plus couramment utilisée. Le secteur d’Eloka-palmeraie, localisé à l’Est d’Abidjan constitue le cadre d’étude. Après application des deux approches de cartographies sur une image SPOT5, deux cartes d’occupation du sol ont été obtenues. Une évaluation visuelle et statistique à partir des indices de performance d’une classification, a révélé que la carte d’occupation du sol obtenue par l’approche orientée-objet présente un contenu sémantique bien meilleur et plus conforme à la réalité de terrain que celui de l’approche pixel par pixel. Aussi, la confrontation des résultats cartographique révèle qu’il existe un écart important de compatibilité entre ces derniers. La prise en compte au niveau de l’approche orientée-objet des paramètres texturaux, de forme et de voisinage, en plus de la signature spectrale, sont autant d’élément qui concourent à la bonne performance de l’approche orientée-objet.
Fichier non déposé

Dates et versions

hal-01786605 , version 1 (06-05-2018)

Identifiants

  • HAL Id : hal-01786605 , version 1

Citer

Armand Kangah, Charles Kramo, Célestin Hauhouot. ETUDE COMPARATIVE DE DEUX APPROCHES DE CARTOGRAPHIE DE L’OCCUPATION DU SOL EN MILIEU RURAL : PAR PIXEL ET PAR OBJET . Photo-Interprétation. European Journal of Applied Remote Sensing (PIEJARS), 2018, 53, pp.19-29. ⟨hal-01786605⟩
139 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More