A Novel Adaptive Scheme for Evaluating Spectral Similarity in High-Resolution Urban Scenes

The analysis of high-spatial-resolution urban images for object recognition must deal with variable illumination conditions and many spectrally similar materials in the built environment. Spectral similarity measures have the potential to contribute to the effective analysis of urban scenes, however, without readily available surface reflectance conversion, the characteristics of existing spectral measures may lead to unacceptable performance. To better account for these spectral imaging scenarios for an urban environment, a simplified in-scene radiometric calibration approach is presented to preserve data collinearity, and a novel spectral similarity measure based on the geometric characteristics of the Mahalanobis distance is developed to incorporate both spectral direction and spectral magnitude. With a minimum of human input to define representative pixels, the experimental results demonstrate through the analysis of ROC curves the potential advantages of the novel distance measure when applied to the identification of materials in urban images.