Identification of urban objects using spectral library combined with airborne hyperspectral imaging
Résumé
The development of the hyperspectral imagery, gives an opportunity to characterise more precisely terrestrial surfaces, landscapes and the geographic objects because of the high spectral resolution of the sensors. The hyperspectral imagery is increasingly used in many fields of applications: geology, agriculture of precision, gas detection, etc., or urban planning and management as in this research. The characterisation of the urban objects is constrained by a specific spatial and geographic context.
The urban environment is characterised by a high spatial, objects and spectral heterogeneity between the object’s urban surfaces; and the uses of “classical” classification techniques is often long and inefficient . To resolve the problem of the classification of the urban objects , a method of classification based on the use of urban object’s spectral library has been developed. The uses of this approach could be useful for the detection of some specific objects , by spectral classification. The goal is to discern efficiently urban objects in a high heterogeneous geographic context. The airborne hyperspectral image used for the application in Kaunas city centre (Lithuania) dated of July 2015. The acquisition was made with the Finnish VISNIR hyperspectral imager RIKOLA LTD. The airborne imager was fixed below a single-seat aircraft, and was configured to produce 16 bands, with a spectral range of 500-900nm, and a spatial resolution of 1 meter. The mean aircraft altitude was equal to 1000 meter during the flight mission. Kaunas city centre is characterised by a heterogeneity of urban structures, built-ups, roofing’s, roads and vegetation species. They are results of mixing of post-soviet urbanisation, soviet urbanism and architecture, 14-18e century’s old centre district, 19e Russian century buildings and a large interwar Art Nouveau urban space. The urban object’s identification method based on the uses of external spectral library concerns the vegetation species, and manmade materials recognitions . The city main vegetation species are: deciduous trees, coniferous trees and green grass. Regarding the manmade materials, the city is composed essentially of tile, metal, asbestos roofing, asphalt roads, pavements, and stony roads.. The spectral library was generated on Kaunas during the summer 2015 and is called the “KUSL” spectral library. The vegetation species are hardly recognizable using spatial classification methods (i.e. almost the same colour and texture, no form recognition is possible between trees). Indeed, a spectral classification could be more useful to detect the species, thanks to their spectral signatures . The detection of the vegetation is made by the implementation of KUSL vegetation spectra into the Spectral Angle Mapper classifier (SAM). For each species, five training spectral signatures from the KUSL library are extracted and implemented in the SAM classifier. The region of interest is composed of variety of vegetation species; green grass, coniferous and deciduous trees. Our classification method allows the identification of green grass in Traffic circles, parks, and individual houses. Around the green grass surfaces, some deciduous trees are also detected (Pines and Fir), finally three different coniferous trees (example: Chestnut, Linden, Aspen) spread along the city; in the parks, individual houses, and along the streets, are detected.The extraction of manmade materials from our 16 bands Kaunas hyperspectral image is based on the same method (SAM classification using the KUSL spectral library). In addition, some material reference spectra have been directly extracted from the hyperspectral images, and used in the classification process because the KUSL library is limited in terms of roofing materials. Only five roofing objects essentially of tile and metal roofing exist at this moment. The database is under building. The tile roofing are correctly detected using the KUSL library, some metal painted roofing are detectable, as the red painted metal roofs. The asphalt is also recognised, but highly correlated with bitumen coated and asbestos roofing . The pavements are detectable but correlated with tile roofing. The intersection mask between tile class and road (stony and pavement) class resolves the problem of confusion. Other painted metal roofing are identified with the use of the reference spectra extracted from the image. The validation of the classification results were done using recent pictures of urban objects taken over Kaunas, other validation strategies are under-development especially for the vegetation species. Even with a limited spectral bands number, the spectral classification results using the airborne 16 bands image coupled with KUSL spectral library are significant, with an efficient discrimination of urban vegetation species, and some man made materials, as tile, metal roofing and asphalt. Nevertheless, an increase in the spectral bands could enhance considerably the spectral classification results and the recognition of urban objects. A new campaign of acquisition with 50 spectral bands is planned in summer 2016 with the enrichment of the KUSL urban spectral signatures database.