Automated mapping of sub-pixel impervious surface area from landsat imagery

The past few decades have seen rapid, global urbanization. Remotely sensed imagery is the best source of information about the extent of urbanization, but extracting urban extent from remotely sensed imagery is often an intensive, supervised task for analysts to perform. This project presents a fully automated method to extract impervious surface area (ISA), an important component of urban expansion, from Landsat TM and similar sensors. These moderate resolution sensors have a multi-decade collection archive, sub-monthly revisit rate and have served as a model for other national and commercial programs. The unsupervised methodology proposed herein, termed the PEEL process (pre-processing, endmember extraction and labeling), is an SMA (spectral mixture analysis) technique that uses as inputs endmembers that have been labeled by a SVM (support vector machine) classification through the fusion of the PanTex GLCM-based texture measure and endmembers drawn from the SMACC (sequential maximum angle convex cone) algorithm. Labels are provided to endmembers with an overall accuracy of 94% across 13 Landsat scenes from different sensor types and of several regions and urban forms. Multiple unmixing methods are tested, with BNMESMA (brightness normalized multiple endmember spectral mixture analysis) performing the best with a RMSE of 0.276. Caution is given regarding the value of RMSE as a metric for comparing method accuracy and more detailed error metrics are introduced. The method is shown as a viable template for mapping ISA across multiple scenes and as a useful framework for analyzing large archives of imagery with a common, automatable methodology.