Generic spectral library framework for urban land cover mapping with optical remote sensing imagery

Abstract

Spectral libraries link surface reflectance characteristics to thematic cover type interpretation. Despite their potential, spectral libraries are rarely used beyond their original application or analysis. In this paper we introduce the concept of a Generic Urban Spectral Library (GUSL). A GUSL is a thoroughly labelled collection of multi-site, -sensor and -temporal spectral libraries that supports urban mapping. The GUSL is envisioned as an open data source equipped with tools that facilitate deployment for different mapping purposes and image types. We introduce the GUSL concept, present a preliminary implementation containing eight spectral libraries and discuss two use case experiments performed with the spectral libraries and the GSL-tools currently included in the GUSL. The first use case focuses on GUSL expansion with image-derived spectral endmembers. From a 200 by 200-pixel hyperspectral urban image, we extract 229 endmember spectra. By comparing the extracted spectra to the spectra already included in the experimental GUSL, we demonstrate how the GUSL can reduce the burden of endmember labelling and identify spectra that are novel to the GUSL. In the second use case the experimental GUSL is used for mapping urban land cover from airborne hyperspectral data in two cities (Munich, Brussels) for which local spectra are included in the GUSL, and in a third city (Pavia) for which no local spectra are present in the library. Generalized material group and more detailed artificial material type mapping yield average user and producer accuracies ranging between 0.72 and 0.96. When mapping is performed on the Pavia image, the obtained material group accuracies remain reasonable, i.e., around 0.80. Our experimental results confirm the potential of the GUSL for urban library building and mapping. Future research avenues are proposed to move towards an operational GUSL.