Mapping spectrally similar urban materials at sub-pixel scales

Abstract

Future space-borne imaging spectrometers could enable global comparative analyses of urban composition. In particular, the high spectral resolution of imaging spectrometry could improve the discrimination of materials that have similar spectral signatures but are functionally dissimilar, such as turfgrass and trees. However, the amount of reflected energy needed to measure narrow-band reflectance with acceptable signal-to-noise ratios means that space-borne imaging spectrometry data will be collected at relatively coarse spatial resolutions, potentially limiting its usefulness for mapping urban composition. In this study, we use Airborne Visible Infra-Red Imaging Spectrometer-Classic and -Next Generation imaging spectrometry acquired in the summer of 2014 over the Santa Barbara, California area to quantify sub-pixel urban composition at fine (4m) and coarse (18m) spatial resolutions. We develop and compare spectral libraries of single- and multiple-resolution endmembers, and use Multiple Endmember Spectral Mixture Analysis to estimate sub-pixel fractions of spectrally dissimilar materials (vegetation, impervious, pervious) as well as pairs of spectrally similar materials (turfgrass and tree, paved and roof, non-photosynthetic vegetation and soil) at both resolutions. Fractions were validated using 1m orthophotography. Overall, fractional accuracy was affected by the spatial resolution of the spectral library and image, and the (dis)similarity of the measured classes. Spectral libraries of multiple-resolution endmembers performed better than single-resolution libraries, likely because they increase within-class variance by capturing multiple levels of material variability that occur across spatial scales. A positive relationship was observed between pixel size and the number of sub-pixel materials, however significant pixel mixing occurred at 4m resolution, with an average of 48% of all pixels modeled by more than one endmember. Fractional estimates produced by the best performing libraries at 4m and 18m resolution correlated with validation fractions, with mean R2>0.89 for spectrally dissimilar classes and mean R2>0.76 for spectrally similar classes. These results demonstrate the scalability of fractional estimates of urban materials using imaging spectrometry, suggesting its potential for future global urban analyses.