|
|
Detecting man-made structures in urban areas using multi-spectral and geometric classification methods Authors: J. C-W Chan, R. Bellens, S. Gautama and F. Canters Publication Date: Jul. 2006
Abstract: Mapping man-made objects from Very High Resolution (VHR) imagery is a very difficult task, especially in urban areas. To produce maps from remotely-sensed data for a complex urban scene, contextual as well as textural information are needed. This study investigates whether textual/contextual information in the form of geometric activity (GA) features of the target classes can enhance classification of man-made objects in an urban environment.
Geometric activity features such as ridge features and scale-based morphological features are proposed for detection of man-made objects in the urban area using VHR imagery such as IKONOS and Quickbird. Ridge features are defined for detecting linear objects such as roads and paths. Scale-based morphological features are developed for objects with explicit structures such as buildings and houses. Since numerous variants of these features can be generated for different window sizes and scales, Multiple Discriminant Analysis was implemented under Matlab for the fine-tuning and generation of summary features.
To assess the potential of the GA features for mapping of urban areas different classification scenarios were set up. Two popular learning algorithms (decision tree classifiers and multi-layer perceptrons) were implemented to test for the utilization of extra input features. GA features were compared with object-based features generated from eCognition® for per-pixel classification. In order to have a good understanding of the performance, a full ground truth was produced to generate the confusion matrices.
Our results found GA features effective for classification of man-made objects in the urban area using VHR imagery. Accuracy improvements for important classes such as roads and buildings are significant with both learning algorithms. The accuracy of the road class increased 12ᆣ%, to 70%, and the class buildings with dark roof (the most important class among building related classes) improved by 10% after inclusion of GA features, though the accuracy is still low at 52%. In terms of class and overall accuracies, GA features and object-based features are comparable. However, with GA features better classification results are attained in terms of coherence of objects, especially for road networks. The best accuracies are attained by combining GA and object-based features. Overall accuracies reached 75ᇣ% and average accuracies of man-made object classes reached 72ᇞ%. The use of combined features, unfortunately, produces a less coherent classification with patchy objects. Also the use of a binary hierarchical classification strategy was tested. Initial results did not show significant improvements. However, more efforts are needed to reach definite conclusions.
Based on the promising experience and results, GA features should be further investigated in terms of optimization and training sample extraction. The need for extensive training samples using VHR data remains an issue for features generation as well as classification. Binary-hierarchical classification strategies based on advanced classification algorithms can further be investigated. A final goal of most studies related to VHR remote sensing is to generate end products that can readily be used in a GIS environment. To date, this is still a big challenge. One of our original objectives, to vectorize the classification result using confidence level information obtained from the output of soft classification, was not achieved due to low accuracies. Future research could focus on this issue. Emphasis on operational aspects of object detection and labelling is indeed important to narrow the gap between the production of classification results and their practical use.
|
|
