University of Twente Student Theses
Integration of Spatial and Spectral data of very high Resolution imagery for building footprint detection using Super Resolution Mapping
Wijeratna, Nanthamuni Arachchge (2011) Integration of Spatial and Spectral data of very high Resolution imagery for building footprint detection using Super Resolution Mapping.
|
PDF
2MB |
| Abstract: | Building footprint detection from VHR remote sensing images is an important application to supply fundamental data for GIS application and topographical mapping. Automatic detection of shape and size of the building is a challenging task due to the spectral limitation of the VHR MS images and the spatial limitation of the VHR panchromatic image. The integration of spectral and spatial data of VHR MS and panchromatic images is a solution for above limitation. The integration of those data can be done using the image fusion techniques and the MRF based SRM techniques. As the image fusion affects the original reflectance data or the DN value of the image, MRF based SRM is better to preserve the original reflectance value of the images in data integration. Also the MRF based SRM is sensitive to the shape and size of the objects. Therefore this study is carried out to detect the building footprint with the integration of spectral and spatial data of VHR images using MRF based SRM. The study area for the research is Lampuuk village in Indonesia and images are a 4m spatial resolution MS image with four spectral bands and 1m spatial resolution panchromatic images of KOMPSAT -2. This method is based on the MRF based SRM technique following soft classification. Soft classification is applied to the VHR MS image to get the land cover proportion images. Then the initial SRM is generated using the proportion images produced from the soft classification and the scale factor 4. The initial SRM is optimized with the posterior probability of the pixel. According to the MRF and Gibbs equivalence the energy is optimized instead of optimization of probability. The maximization of posterior probability is equivalent to the minimization of posterior energy. The posterior energy is modelled using contextual information and the likelihood energy is modelled using the class statistics from MS and panchromatic images. Then the optimization was done with Maximum A Posterior (MAP) solution which is reached with simulated annealing (SA) algorithm. The optimization with SA is compared with the Iterated Conditional Modes (ICM). Finally the validation of the method is done in pixel based and object based analysis. This method was compared with the conventional MLC. The pixel based accuracy assessment of the SRM optimized with SA shows the user accuracy 68%, producer accuracy 65%, overall accuracy 87% and the kappa value 0.584. Those values of the SRM optimized with ICM are 69%, 64%, 87% and 0.581 respectively. The same measures from MLC with fused image are 50.86 %, 68.69%, 62.11% and 0.483 respectively. The object area based accuracy assessment of SRM with SA showed the over identification 0.436, under identification 0.23 and total error 0.493. Those from SA with ICM are 0.419, 0.24 and 0.483 respectively. The same measures from MLC with fused image are 0.550, 0.252 and 0.605 respectively. Then the building object wise validation also done and it showed that the MRF based SRM method detected 276 building footprints out of 292 building footprints in the reference image. According to that MRF based SRM has detected 95% of the buildings in the study area. According to two types of accuracy measures it can be concluded that both the SA and ICM algorithms produced almost the same accurate SR maps and detected the same percentage of buildings in the study area. Secondly it can be concluded that MRF based SRM provides more accurate results than image fusion for the integration of spectra and spatial data of VHR images in building detection. Third conclusion is that the MS image with panchromatic image provides more accurate SR map for the building footprint detection. The overall conclusion of this study is that MRF based SRM is more accurate than the conventional MLC for the building footprint detection from VHR satellite images. Key words:- Super Resolution Mapping (SRM), Markov Random Field (MRF), Soft classification, Linear Spectral Unminixg, Maximum Likelihood Classification, Maximum A Posterior solution (MAP), Simulated Annealing (SA) and Iterative Conditional Modes (ICM). |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/92777 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page