Recognition of urban vegetation by hyperspectral airborne high-resolution VNIR imager (Kaunas, Lithuania)

National audienceThe management of urban green spaces became crucialbecame crucial with the growing of the environmental and climatic issues nowadays; insofar as they provide multitude of services for the well-being of citizens, and contribute to the environment regulation. Remotely sensed data offers an abundance of environmental and urban information needed for the urban planners and urban architects to characterize, manage, and monitor urban green spaces. These data could brings complementary information’s at large scale (e.g. fast species recognition, vegetation health, green spaces quantification) and consolidate field inventories data Multi-platform remote sensing sensors exist, with different spatial/spectral resolutions and different spectral domains. The choice of the VNIR sensor takes To choose a sensor, one must take into account the allocated budget, the study scale (individual trees scale, groups of trees scale, green frame scale) and the seasonality aspect.Multi-spectral Satellite imagery are limited for the recognition and characterization of the urban objects, indeed, few bands are generally available (i.e. 4 to 10 spectral bands usually), with a limited spatial resolution (e.g. presence of mixed pixels) and some other problematics related to clouds presence. . An interesting alternative to satellite imagery could be the use of hyperspectral imaging (HSI) , which offers a better spectral characterization at the pixels level (from 50 to hundreds of contiguous bands) . HSI offers gives precise spectral characterization of each pixel, and generates a better spatial resolution especially with airborne platform (i.e. meter to centimeters). On the other hand, HSI introduces specific issues related to correlation between bands (i.e. information redundancy), to sensor noise (i.e. usually higher than the multispectral case), and to the complexity of image processing and image analysis because of the size and number of spectral bands (i.e. usually hundreds of bands to process). An Airborne VNIR 64 bands hyperspectral imager with 0.5 meter GSD and FWHM of ~10 nm is experimented for the recognition of urban vegetation species of Kaunas city (Lithuania). The large umbrella of green spaces biodiversity (parks, green open spaces, and urban forests), urban structures (XVIe, XIXe, Inter-war Art Nouveau, Soviet, post-soviet urbanisms) and more specifically of tree species, allow us to consider Kaunas as a promising test zone to assess the feasibility of urban vegetation identification by hyperspectral imagery. The maps, biogeographic and individual tree species geo-localized databases available as well as the sampling techniques gathering the validation of the methodologies implemented for the vegetation and species recognition. The generated maps, biogeographic and individual tree species databases were integrated in a Geographic Information System (GIS) as well as the sampling techniques that served for complementary validation (i.e. georeferenced photos).The vegetation mapping will concern three identification scale:; (1) green frame scale for assessing the recognition of vegetation pixels, (2) group of species scale (i.e. deciduous, coniferous and short grass), and (3) individual species recognition. The methodology developed and tested validated is composed of four main steps: (1) image pre-processing, (2) creation of a vegetation mask, (3) automatic classification based on spectral library (i.e. image and laboratory spectra used for the training process) and, (4) validation of the results. The pre-processing step includes band co-registration , mosaic build, calibration by MODTRAN model followed by a Minimum Noise Fraction (MNF) transformation to reduce the image noise. The second step consists of the elimination of the non-vegetation pixels (water bodies, shadow and urban fabric) to reduce misclassifications. The third step is a classification based on spectral SVM with a RBF kernel. The classifier is applied over the full spectral range and over a reduced spectral range to assess if a band reduction step contributes to the enhancement of the classification mapping. The validation and accuracy level estimation of the generated maps is done using the validation data included in the GIS and collected over Kaunas city. Keywords: hyperspectral, green frame, vegetation species, SVM, classification, band selection, green spaces monitoring