Integration of object-based image analysis and data mining techniques for detailes urban mapping using remote sensing
Urban areas consist of a wide range of man-made and natural features, which lead to a high level of spectral and spatial confusions in detecting roofing materials and other urban land cover features. These heterogeneities and confusions result in erroneous thematic maps that use spectral information...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/57895/1/FK%202015%2079RR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Urban areas consist of a wide range of man-made and natural features, which lead to a high level of spectral and spatial confusions in detecting roofing materials and other urban land cover features. These heterogeneities and confusions result in erroneous thematic maps that use spectral information only. Reliable and new approaches that combine the spectral and spatial characteristics of urban land covers are necessary to extract accurate detailed maps of urban features in different regions. Airborne hyperspectral images with very high spatial and spectral discrimination capabilities can be considered to characterize urban land-cover classes;however, the frequent use of airborne hyperspectral images is infeasible because of the limited spatial coverage and high cost of data acquisition. Consequently, pan-sharpened WorldView-2 (WV-2) multispectral images with a spatial resolution of 0.5 m were used as the main data for this research, whereas hyperspectral images were included as supplementary data to indicate the productivity of the proposed approach in decreasing the dimensionality of such images. Object-based image analysis (OBIA) was implemented to delineate urban surface materials. OBIA was performed in a rule-based structure that requires selecting and identifying rules. In a subsection of the thesis, OBIA was supported by including ancillary information, such as LiDAR data. The reproducible and transferable novel models were proposed based on (1) user-defined OBIA rule sets and (2) data mining (DM) techniques. In the first case, the rule sets were manually developed from the first study site and then transferred to the second study site, which had a wider coverage. Overall accuracies of 88.05% and 87.78% were achieved for the first and second study sites, respectively. In the second case, available training data from the first study site were used to conduct the DM task. The proposed OBIA rule sets were automatically organized from the C4.5 algorithm to form a decision tree structure that explores a wide range of spectral, spatial, and textural attributes. The rulebased classification of the first study site obtained an overall accuracy of 87.90%. Finally, the generated model was validated in the second study site to prove that its performance was reproducible and applicable to an area with a wider geographic coverage. This transferability analysis was performed without including any training data from the second study site and an overall accuracy of 85.16% was achieved. Apart from the analysis of the transferability of OBIA rule sets, this thesis contained two subsections. First, the OBIA of a WV-2 image and LiDAR height information was employed to improve urban surface material classification. This data integration resulted in an improvement of 7% in the overall classification accuracy of the WV-2 image. Second, the dimensionality of OBIA attributes in hyperspectral images was reduced. Numerous OBIA attributes were explored using the C4.5 algorithm on the Universiti Putra Malaysia (with 20 bands) and Kuala Lumpur (with 128 bands) hyperspectral images. These images obtained overall accuracies of 93.42% and 88.36%, respectively. The manually developed OBIA rule-sets achieved the transferable land cover classification in different areas. In this research, to eliminate the manual developments of the rule-sets, the supervised DM technique was used to identify the appropriate selection of attributes for object-based classification. This algorithm represents the decision tree knowledge model, enables fast classification of intra-urban classes, and disables subjectivities related to the interaction with analysts. The proposed integration of DM algorithm and OBIA provides the opportunity to generate the transferable OBIA rule-sets based on the available training area which can be re-used in other study areas. The generated rule sets can be applied in different WV-2 images to extract similar land-cover environments by providing an automated procedure. Furthermore, supervised DM-based DT overcomes complexities related to a high level of dimensionality in hyperspectral OBIA attributes and the bias of analysts in creating rule sets. The OBIA enhancement of WV-2 image was performed by adding LiDAR height information as an ancillary data. LiDAR data was effective to improve the productivity of OBIA and reduce the complexity of the OBIA rulesets. The detailed land cover maps of this study could support the environmental applications related to water quality assessment, urban microclimate and urban health assessment. |
|---|