Various spectral data preprocessing approaches have been used to improve endmember extraction for urban landscape decomposition, yet little is known of their comparative adequacy for impervious surface mapping. This study tested four commonly used spectral data treatment strategies for endmember derivation, including original spectra, image fusion via principal component analysis, spectral normalization, and the minimum noise fraction (MNF) transformation. Land cover endmembers derived using each strategy were used to build a linear spectral mixture analysis (LSMA) model in order to unmix treated image pixels into fraction maps, and an urban imperviousness map was generated by combining the fraction maps representing imperviousness endmembers. A cross-map comparative analysis was then performed to rank the four data treatment types based on such common evaluation indices as the coefficient of determination (R²) and root mean square error (RMSE). A Landsat 7 ETM+ multispectral image covering the metropolitan region of Shanghai, China was used as the primary dataset, and the model results were evaluated using high-resolution color-infrared aerial photographs of roughly the same time period. The test results indicated that, with the highest R² (0.812) and the lowest RMSE (0.097) among all four preprocessing treatments, the endmembers in the form of MNF-transformed spectra produced the best model output for characterizing urban impervious surfaces. The outcome of this study may provide useful guidance for future impervious surface mapping using medium-resolution remote sensing data.