Research project S0/00/050 (Research action S0)
Context and objectives
The overall objective of the project is to propose a strategy for improved man-made object extraction (road structures, buildings) from very-high-resolution satellite imagery of urban areas. The strategy is based on the development of novel indices that measure the geometric activity in the neighbourhood of a pixel, and on the use of these indices in a per-pixel, knowledge-based classification framework. The strength of the proposed method is the close interaction between multi-spectral classification and accuracy assessment methods on the one hand and computer vision techniques on the other. Specific objectives of the research are:
a. to develop geometric activity indices as an alternative for the more traditional texture indices that better describe the geometry of man-made structures in very-high-resolution satellite imagery;
b. to propose a strategy for selecting the most suitable geometric activity indices to be included in a multivariate classification framework;
c. to assess the improvement in classification accuracy of man-made objects obtained by making use of the selected geometric activity indices.
Methods were developed and tested on Ikonos and Quickbird imagery for two urban test sites situated south and southwest of the city centre of Ghent (Watersportbaan Ghent and Sint-Denijs-Westrem/De Pinte).
Methodology
Several types of geometric activity features were defined based on the use of corner detectors, ridge detectors and morphological operations. All features were examined in detail in terms of their ability to detect man-made structures in very-high-resolution imagery. Since numerous variants of each feature can be generated for different window sizes and scales, Multiple Discriminant Analysis (MDA) was implemented under Matlab for the fine-tuning and generation of summary features. To assess the potential of the geometric activity features for mapping of urban areas different classification scenarios were set up. Two popular learning algorithms (decision tree classifiers and multi-layer perceptrons) were used to test the added value of including geometric activity summary features in the classification, next to spectral variables. The use of geometric activity features was also compared with the use of object-based features generated by eCognition software. Assessment of classification performance was based on exhaustive ground truth obtained by visual interpretation of the imagery, use of large-scale aerial photographs and complementary field checks.
Results
Because the current state-of-the-art corner detectors produce very high false detection rates, results produced by these detectors proved to be unsuitable for deriving meaningful geometric activity indices for man-made object classification. For deriving a final set of geometric activity indices attention was therefore focused on information provided by ridge detectors and morphological operators. Several ridge based features were defined giving a per-pixel indication of the presence of linear structures. The main problem with this kind of features, however, is the unreliability near borders. Therefore, from initial classification experiments with spectral and ridge features, only little improvement was achieved in distinguishing between roads and other man-made objects (buildings in particular). Since morphological features are less affected by these problems and contain similar information, much attention was given to the development of morphologically-based indices. Methods were developed to derive morphological signatures giving an indication of the minimum and maximum size of an object at different scale levels. The combined use of ridge features and morphological features led to significant improvements in the classification accuracies of man-made object classes, especially for roads. Compared to a scenario where only spectral information is used, accuracy improvements of 12 to 15% were observed. Although the use of object-based features generated by eCognition software also led to an improvement of classification accuracy for roads, the gain in accuracy was less than with geometric activity features. Combining the use of geometric activity features with object-based features leads to best results for building detection, although the accuracy for the most prominent building class remains low (around 60%).
