Investigating the suitability of Sentinel-2 data to derive the urban vegetation structure

Abstract

Urban green is indispensable from an urban ecological and social point of view and fulfils important functions such as dust binding, temperature reduction, wind damping or groundwater recharge. Especially for bioclimatic modeling, knowledge of size, structure and green volume of the urban vegetation is essential. Manual mapping of vegetation structures is timeconsuming and cost-intensive and can only ever be carried out in locally limited study areas. Active and passive remote sensing technologies in combination with automated methods for information extraction offer the opportunity to record the green structure in urban areas differentiated according to vegetation types. The new globally and freely available data provided by the European Copernicus Program raises the question whether these data are suitable for mapping and quantifying the urban green structure, including an accuracy estimation. Previous studies on the usability of Sentinel-2 data for vegetation analysis were essentially limited to crop and tree species classification in open space. The approach presented here thus considers for the first time the application of this data in a purely urban environment. Here we present a modeling approach based on multiple regression models. A Sentinel-2A scene from July 4, 2015 covering the greater Dresden area served as the input data set. After atmospheric correction of the satellite image scene 10 spectral channels were available. A high-resolution vegetation cover model with a grid width of 50 cm was available as a reference data set for the entire study area (City of Dresden, Germany). This takes into account the vegetation classes deciduous trees, conifers, shrubs, low (grassy) vegetation and arable land. Thus the area share of these vegetation types could be determined aggregated for each pixel of the satellite image scene. In addition, vegetation indices (NDVI and others) were calculated using suitable channels. For the prediction of each vegetation class, estimation equations were drawn up and evaluated with regard to their quality. Especially for deciduous and coniferous trees, satisfactory model quality values could be obtained, so that the green component prediction in these cases represents a useful basis for the determination of the green structure at the building block level in urban areas.