Improvement of Buildings’ Roof Detection using Spectral-Spatial Information in Hyperspectral Images

The precise and up-to-date information about urban features is being increasingly demanded for various applications in the modelling and planning of the urban environment. Hyperspectral imagery is one of the most efficient tools for obtaining the physical and spatial information in such a complex area. Indeed, the hyperspectral systems are able to collect valuable image data, which can be used for detection, classification, identification and quantification of human or natural phenomena. Target detection technique is one of the most reliable and efficient approaches aiming to extract information about a class or classes of interest. In this paper, we propose an analytical framework to produce an accurate land-cover map of buildings’ roof using highly spatial hyperspectral imagery. The proposed approach consists of several steps. The first step attempts to reduce the undesirable noise and spectral correlation of hyperspectral images by using the MNF (minimum noise fraction) algorithm. In the second step, different spectral matching techniques are employed to obtain the target maps. The spatial information is also used to improve detection accuracy. Finally, the ANFIS (adaptive neuro-fuzzy inference system) method is used to fuse the detection results and arrive at a land-cover map of buildings’ roof materials. The results of quantitative and qualitative assessments of the experiments show that the proposed algorithm can lead to acceptable accuracy for mapping of urban features.