Extraction automatique d'objets nécessaires à la cartographie des interfaces habitat-forêt en région française méditerranéenne | Automated feature extraction on quickbird image required to map wildland urban interfaces (WUI) in the french mediterranean region

In the french mediterranean region, forest fire risk increases because of dynamics of land cover: fuel load accumulation due to agricultural fallows and non exploited forest, urbanisation expansion. Urbanization joined to the forest extension phenomenon, generates new spatial configurations called wildland urban interfaces (WUI) (Jappiot et al., 2002; Lampin et al., 2006). WUI concerns integrate "natural" vegetation connected to urban systems which bring out both components of forest fire risk: hazard (breaking out probability, distribution) and vulnerability (Blanchi et al., 2002). As these interfaces should extend in the next years, assessing forest fire risk in the WUI is a need for wildfire prevention and land management. To characterize wildland urban interfaces, satellite images allow us to work on large and spectral homogeneous areas that was not possible on aerial photos. Traditional classifications per-pixel were adapted to high resolution images (10 m) but do not let working on textural information appearing on very high resolution satellite images. Characterizing and mapping WUI need to extract involved shapes like houses or roads for urban features, but also involved textures like scrublands or forest with different densities of trees for natural features. A lot of papers uses feature extraction or segmentation programs on homogeneous areas to detect specific objects: man-made (Sithole and Vosselman, 2006), burnt area (Mitri and Gitas, 2004) or fuel mapping (Gitas, 2006). In our case, we need to extract these objects in heterogeneous contexts where natural, agricultural and artificial features are interconnected. This paper presents a methodology, using remote sensing in pre-fire planning: automated feature extraction to map an accurate and reliable land cover required to characterize WUI at a large scale. The study area is located in the Meyreuil district near Aix en Provence (South of France). A Quickbird image taken in June 2006 was acquired. A principal component merge between 0.6 panchromatic and 2.4 multispectral Quickbird images was done to retain the spectral information of the four Quickbird MS bands. A feature extraction program, Feature Analyst 4.1® for Erdas Imagine®, was tested in the framework of the FireParadox European research program. Using multiple spatial attributes (size, shape, texture, pattern, spatial association) with spectral information, this software improves considerably automated detection of involved land cover structures. Vegetation classes integrate the texture of the objects: arrangement of pixels with different radiometry, shadows, etc. Mineral objects can be easily identified only if they have a specific shape (that is not the case of roads here), otherwise it is difficult as algorithms used by Feature Analyst® are unknown.