A comparison of two spectral mixture modelling approaches for impervious surface mapping in urban areas

Urban change processes that have been occurring over the past decades are affecting the human and natural environment in many ways, and have stressed the need for new, more effective urban management approaches. In this context, mapping man-made impervious surfaces has been the focus of attention as impervious surfaces can be used as a general indicator to quantify urban change and its environmental impact. Despite the currently available digital imagery from high-resolution satellite sensors such as Ikonos and Quickbird, or from airborne cameras, spectral unmixing approaches applied on medium-resolution data from sensors such as Landsat Thematic Mapper (TM)/Enhanced TM Plus (ETM+) or Syst?me Probatoire d' Observation de la Terre-Haute Résolution Visible (SPOT-HRV) offer interesting perspectives to map impervious surfaces for large spatial extents. Several techniques for subpixel impervious surface mapping have been examined previously but there is a lack of comparative analysis. Our objective was to compare two spectral mixture analysis (SMA) models: the linear spectral unmixing model and the multilayer perceptron (MLP) model. Both models were implemented in a multiresolution framework, where reference data for model training were obtained from a high-resolution land-cover classification (derived from Ikonos imagery), while the models themselves were applied on medium-resolution data (Landsat ETM+). As a secondary objective, the effect of spectral normalization on the performance of both models was assessed. The MLP model clearly performed better than the linear mixture model. The average absolute error of the impervious surface proportion estimate within each medium-resolution pixel was 10.4% for the MLP model versus 12.9% for the linear mixture model. Spectral normalization was used to improve the results obtained by the linear mixture model, with the mean absolute error (MAE) for impervious surfaces decreasing from 14.8% to 12.9% after normalization. Its effects on the MLP model appeared to be insignificant. The outcome of this study can help to provide guidance for the selection of an approach to estimate continuous impervious surface fractions from medium-resolution data.