Compilation de données radar et optiques pour la cartographie des classes d'occupation du sol aux environs du système lacustre de Bizerte (Tunisie du Nord) - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Photo-Interprétation. European Journal of Applied Remote Sensing (PIEJARS) Année : 2016

Compilation of radar and optical data for land cover mapping in the area of the lake system of Bizerte (North of Tunisia)u

Compilation de données radar et optiques pour la cartographie des classes d'occupation du sol aux environs du système lacustre de Bizerte (Tunisie du Nord)

Résumé

Radar, owing to its geometry acquisition (directional and side viewing) and longer wavelengths, provides unique information on the Earth’s surface. It is complementary to optical remote sensors. The response of a scene to the wave emitted by a radar sensor is dependent primarily on a large number of factors including moisture, roughness, amount of vegetation, geometric structure of the scene, and, secondarily, on parameters of the system itself (wavelength, angle of incidence, polarization). Many studies have focused on the physical understanding of polarimetry, including backscatter, when characterizing a target (Huynen, 1970). The purpose of this work is to assess the potential of radar data for the characterization of a complex lacustrine area and to evaluate the complementarity between radar and optical data for the identification of land use/land cover types. The study area is the lake system of Ichkeul-Bizerte, one of the important components of Tunisian coastal landscapes. It is located in northern Tunisia between latitudes N 37° 19’ and N 37° 5’. The area belongs to a collapse zone surrounded by small hills and consequently the catchment is receiving various kinds of materials, inducing complex physical, chemical and biological processes in the water system. In the present study, multifrequency remote sensing data are used. A preprocessing phase was performed on the radar data, including geometric corrections and speckle reduction. Several speckle reduction filters were tested on radar bands (Median, Lee Kuan and Frost). Two evaluation criteria were used : a visual test and a statistical criterion. Color composite images have been produced, combining HH, HV and VV channels (plate 1). Intensity variations of the radar signal can be observed, which can be explained by wind-induced roughness differences of lake surfaces (Bizerte lake and Ichkeul lake). Mean backscatter is maximum in HV channel, for built-up and forest classes. Statistical and visual results of filtered images allowed us to retain the Enhanced Frost filter with a 3 x 3 pixels window. In addition to reducing speckle, this filter keeps the contours and details of information contained in radar channels. Using a 9 x 9 filter improves the separability between built-up and forest classes. Afterwards, radar images (ASAR and Palsar) and optical images (Landsat) have been re-sampled to the same pixel size of 15 m. The identification of the different scene components is based on their textural and spectral behavior. Textural parameters were calculated using the approach based on the co-occurrence matrix developed by Haralick et al. (1973). Least correlated parameters have been choosen. Thus, eight texture parameters were selected: standard deviation, mean, uniformity, dissimilarity, entropy, correlation, second angular momentum and contrast. These texture parameters are calculated for each of the radar bands CVV, LHV et LHH, unfiltered to preserve textural information. The optical data texture analysis was done using a mathematical morphology method that relies on processing techniques insensitive to radiometric variations (Kombé, 1999). Landsat ETM1 channel served as input image for morphological treatments. This approach included three phases : extraction of small structures, determination of an urban mask and reconstruction of urban patches. The evaluation of the potential of the various radar and optical data (spectral and textural) for identification and classification of land cover classes is made by means of multivariate statistical analysis. This analysis is used to the selection of variables and to study the optimal combinations of parameters to better discriminate cover classes. The linear combination method used is the Principal Component Analysis (PCA) which consists of a mathematical processing based on the analysis of the covariance of the image or the correlation matrix of several sets of data (Bonn and Rochon, 1992). Data integration was made in four steps, starting from the unfiltered data and going to processed data using different operator types. When a new parameter is added, it is checked whether this latter increases the total discriminant power obtained, based on the MPCC (Mathematical Programming with Complementary Constraints) and Kappa parameters. The results of performance classifications made after the four stages of application of the statistical analysis are presented at table 2. This gives us an overall classification accuracy of 75% from step I; this means that, on the basis of microwave multipolarisation information only, we are able to map only the main land cover classes. A filtered channel has improved the discrimination of different classes. The land cover classes are better identified with welldrawn boundaries and a high rate of elimination of isolated pixels. The average of correctly classified pixels (MPCC) is 86.7%. The textural channels improved separation between different classes of soil occupation. The percentage of correctly classified pixels reached 94.2%. The adding of the ETM1 derived channel obtained following the morphological processing in the linear combination has allowed the discrimination of urban structures on the classified image and increased the percentage of correctly classified pixels, reaching 96.2%. The final document generated by these treatments consists of a land use/land cover map. Altogether, eight classes are determined (plate 2). The evaluation of the results of the different steps is done based on MPCC parameters and on the Kappa index. The comparison of the land cover map and available cartographic document shows that the map derived from remote sensing data shows better cartographic fineness: it highlights local heterogeneity in the various classes.
Le but de ce travail est de présenter une méthodologie d’analyse et d’interprétation d’images satellitaires multisources radar RSO (ASAR et PALSAR) et optiques (Landsat ETM) visant à intervenir dans le domaine de la cartographie des classes d’occupation de sol dans les environs d’un système côtier méditerranéen qui est le système lacustre de Bizerte-Ichkeul. Les potentiels des données satellitaires ASAR, PALSAR et Landsat-7 sont comparées pour cartographier les principales classes d’occupation de sol du domaine d’étude. Les données radar polarimétriques de différents types (en bandes C et L ; en polarisations parallèles et croisées) brutes et filtrées, de même que des indices de texture (moyenne, écart-type, second moment angulaire, entropie, etc.) sont extraits de ces données. Des attributs géométriques extraits des données numériques par application des opérateurs de la morphologie mathématique (ouverture, fermeture, chapeau haut de forme) sont intégrés aux données précédentes. Après application de l’analyse statistique multivariée, seuls les paramètres les plus significatifs sont retenus et utilisés en quatre étapes dans un modèle de combinaison linéaire pour la mise en évidence des principales composantes du paysage côtier.
Fichier non déposé

Dates et versions

hal-01629460 , version 1 (06-11-2017)

Identifiants

  • HAL Id : hal-01629460 , version 1

Citer

Ibtissem Amri, Mohamed Chedly Rabia, Jean-Paul Rudant. Compilation de données radar et optiques pour la cartographie des classes d'occupation du sol aux environs du système lacustre de Bizerte (Tunisie du Nord). Photo-Interprétation. European Journal of Applied Remote Sensing (PIEJARS), 2016, 51 (2), pp.2-13. ⟨hal-01629460⟩

Collections

UNIV-EIFFEL
140 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More