Mapping sealed surfaces from CHRIS/Proba data: A multiple endmember unmixing approach

Previous work on spectral unmixing of medium-resolution multispectral data for mapping of sealed surfaces has pointed out the limitations of the approach, which are mostly related to the confusion between sealed surface materials and spectrally similar non-artificial land-cover types. Use of hyperspectral data may improve the accuracy of sealed surface mapping in urbanized areas. In this paper the potential of multiple endmember unmixing for sealed surface mapping from hyperspectral CHRIS/Proba data is examined using a modeling scenario based on endmembers for four major classes: grey sealed surfaces, red sealed surfaces, bare soil and vegetation. A reference database was developed for validating the sub-pixel fractions using 25 cm resolution aerial photographs. The average proportional error for sealed surfaces, vegetation and bare soil is around 15%. Defining a model selection criterion that favors the use of models with few endmembers leads to a substantial improvement of the accuracy of the unmixing.