T. De Roeck, T. Van de Voorde, F. Canters

Cartography and GIS Unit, Geography department Vrije Universiteit Brussel, Brussels, Belgium



Sealed surfaces prevent water from infiltrating into the soil. They have a negative impact on environmental conditions in urbanized watersheds. Mapping and monitoring the extent of sealed surfaces has therefore become important in many applications. A promising technique to obtain information about the distribution of sealed surfaces is to extract its occurrence directly from satellite imagery. To exploit the advantages of both medium and high-resolution data one can make use of a multi-resolution strategy based on sub-pixel classification. In this study, alternative sub-pixel classification models were developed and tested on a Landsat ETM+ image for Brussels and surroundings. Training and validation of the models was accomplished by extracting detailed information about land cover from a classified Ikonos image covering part of the area.


When classifying multispectral high-resolution satellite data (Ikonos, Quickbird) of urban areas, traditional per-pixel classification methods do not produce satisfying results. Besides the appearance of structural clutter in the classified image, certain land-cover classes that are important for mapping sealed surfaces are difficult to distinguish due to the limited spectral resolution of the high-resolution imagery. Object-oriented classification is a promising technique to tackle these problems. Instead of single pixels, this method classifies image objects, obtained by image segmentation, making use of information not only referring to the spectral characteristics of the object, yet also to its texture, context and shape. In this study a hierarchical, rule-based classification approach was applied to produce high-quality reference data for the training and validation of sub-pixel classification models. At each step in the hierarchical procedure a set of rules was defined to separate one land-cover class from the remaining classes, using an optimal subset of object-based features.


The high-resolution classification result obtained this way was used as a reference for training and validation of sub-pixel classification models. By counting the amount of sealed Ikonos pixels present within the pixels of the co-registered Landsat image, the proportion of sealed surfaces was obtained for a random sample of medium-resolution pixels. Alternative models were developed to map the multi-spectral signature of each pixel in the training set onto the proportion of sealed surfaces at the 30 meter resolution. The accuracy of the models was assessed by comparing sealed surface proportions produced for an independent set of Landsat validation pixels with the high-resolution reference classification. An average proportional error of about 10% was obtained, which is comparable to the results found in similar studies.