DÉTECTION DES CUVETTES OASIENNES DU CENTRE-EST DU NIGER PAR CLASSIFICATIONS D’IMAGES-SATELLITE SPOT5-THX - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Photo-Interprétation. European Journal of Applied Remote Sensing (PIEJARS) Année : 2018

DETECTION OF OASIAN BASINS OF EAST-CENTRAL NIGER BY CLASSIFICATIONS OF SPOT5-THX IMAGES

DÉTECTION DES CUVETTES OASIENNES DU CENTRE-EST DU NIGER PAR CLASSIFICATIONS D’IMAGES-SATELLITE SPOT5-THX

Résumé

Extended abstract: Located in a semi-arid zone, oasis basins of east-central Niger (plate 1) are high-potential agro-ecosystems (plates 2 and 3). These oasis basins are covered by dense vegetation consisting mainly of Hyphaene thebaica, Phoenix dactylifera, Digitata andasonia reflecting an oasis-like area, thus the name of oasis basin. Oasis basins represent the last areas where farming is possible throughout the year. These areas have experienced many investments from several NGOs and projects in these recent years. However, the oasis basins are still unknown and / underused by development actors due to their extreme diversity and also due to their huge number. In this context, research has been initiated to preserve the oasis basins and assess their socio-economic roles. In order to achieve this goal a thorough knowledge of the variable characteristics of these basin would be necessary. Given their large number and diversity, remote sensing, with its synoptic vision, is a relevant tool to locate and identify and characterize (set up an inventory) these oasis basins. SPOT5-THX images (2.5m resolution) were used in this study (table 1). This study evaluates the performances of three image classification methods in detecting oasian basins in the administrative unit (« Départment ») of Gouré using a confusion matrix tool (table 2). The methods are pixel analysis, textural analysis and object analysis (plate 4). A visual analysis of image mosaics was performed to obtain a preliminary categorization of the study area and to choose a sampling design. A field campaign led to the selection of a set of regions of interest, one part of which was used to develop the classifications and another part to validate the results. As part of this study, four geomorphological classes were defined (oasis basin, dry valley, dune edifice and hill). Table 3 summarizes the number and areas of the basins obtained by the three methods. The analysis by object yielded fewer oasis basins in number while the pixel approach identified more oasis basins and with areas that are far larger than those of the two other methods. This is explained by a strong confusion between the oasis basins and other classes like dry valley, hill and dune edifice (plate 6). The distribution of oasis basins according to their size (plate 7) shows a very high proportion of oasis basins covering less than 0.5 ha. Although the oasis basins with scopes lesser than 0.5ha represent more than fifty percent, those that have an area greater than 1O ha represent more than half of the areas obtained for each method (plate 8). Evaluation of the results (tables 4-6) shows that the approach by object (global precision was 97%, Kappa coefficient was equal to 0.97) is more efficient than the textural method (overall precision 90%, Kappa coefficient 0.90) and the pixel-based approach (overall precision 82%, Kappa coefficient 0.82). Indeed, given the importance of the detected oasis basins and large differences between the results of the three approaches, the performances of the three methods have been validated through a model of prediction to determine the exact number of oasis basins which approaches the ground reality. The oasis basins probability detection (ability of a method to identify oasis basins observed in the field and photo-interpretation) of the three methods gives satisfactory results: 98.3% pixel analysis, textural analysis 98.9% and 99.2% object analysis. However, the false positive rates, which represent the probability that the oasis basin is detected by a method when it does not exist in reality in the field are important (pixel analysis: 31%, textural analysis: 22% and object analysis: 7%). Oasis basins observed in the field or by photo interpretation and which were not detected by the alternative method (false negatives) are fewer at the analysis by object. The critical success index (CSI) has been computed, taking into account the rate of false positive and false negative results, the quality of the agreement between the estimate of the oasis basin identified by a method and true oasis basins observed in the field or by photo-interpretation, apart from random errors. The object analysis gave the highest CSI value (i.e. 0.94). It is unquestionably the method whose results are similar to those observed in the field. Analysis by texture is better than that by pixel. One can even get a further improved performance of the textural analysis by combination the fourteen textural attributes of HARALICK et al. (1973). This study used a single textural attribute i.e. the variance. Our results confirm several studies showing that the performance of image classification by object analysis is greater than textural analysis and simple pixel based analysis. But some confusion remains between mapped classes. This confusion can be explained by the classes identified in the study, which are geomorphological units. Yet, these geomorphological units have similar land cover. For example, a temporary water can be located atop a hill, in a dry valley or in an oasis basin. The study provides a first original information, namely the number, the position and the surface of all the oasis basins on the study area. The total number of oasis basins in the Department of Gouré is estimated at 11,300 with an error more or less than 0.06.The inventory of these resources is a first compulsory step for the characterization of oasian basins to be used for the evaluation of their agricultural potential.
Situées en zone semi-aride, les cuvettes oasiennes du centre-est du Niger sont des agroécosystèmes à haute potentialité. Cependant, elles sont souvent peu connues par ces acteurs du développement à cause de leur extrême diversité et de leur nombre. Cette étude évalue les performances de trois méthodes de classification d’images dans la détection des cuvettes oasiennes (analyse pixellaire, analyse texturale et analyse par objet) à partir des images SPOT5-THX de 2,5m de résolution spatiale. L’évaluation des résultats montre que l’approche par objet (indice critique de succès égal à 0,94) est beaucoup plus performante que celle utilisant la texture (indice critique de succès égal à 0,78) et l’analyse par pixel (indice critique de succès égal à 0,69). L’étude fournit une première information originale, à savoir le nombre (11300 cuvettes oasiennes détectées avec une erreur de plus ou moins 6%), la position et la surface de l’ensemble des cuvettes sur la zone d’étude.
Fichier non déposé

Dates et versions

hal-01786336 , version 1 (05-05-2018)

Identifiants

  • HAL Id : hal-01786336 , version 1

Citer

Mahamadou Karimou Barké, Bernard Tychon, Issa Ousseini, André Ozer, Charles Bielders. DÉTECTION DES CUVETTES OASIENNES DU CENTRE-EST DU NIGER PAR CLASSIFICATIONS D’IMAGES-SATELLITE SPOT5-THX. Photo-Interprétation. European Journal of Applied Remote Sensing (PIEJARS), 2018, 53, pp.32-46. ⟨hal-01786336⟩
98 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More