HR: 1340h
AN: H13C-0453    [Abstracts]
TI: BAYESIAN NETWORK STRUCTURE LEARNING FOR URBAN LAND USE CLASSIFICATION FROM LANDSAT ETM+ AND ANCILLARY DATA
AU: * Park, M
EM: mhpark@seas.ucla.edu
AF: University of California, Los Angeles, Department of Civil and Environmental Engineering, 405 Hilgard Ave., Los Angeles, CA 90095 United States
AU: Stenstrom, M K
EM: stenstro@seas.ucla.edu
AF: University of California, Los Angeles, Department of Civil and Environmental Engineering, 405 Hilgard Ave., Los Angeles, CA 90095 United States
AB: Recognizing urban information from the satellite imagery is problematic due to the diverse features and dynamic changes of urban landuse. The use of Landsat imagery for urban land use classification involves inherent uncertainty due to its spatial resolution and the low separability among land uses. To resolve the uncertainty problem, we investigated the performance of Bayesian networks to classify urban land use since Bayesian networks provide a quantitative way of handling uncertainty and have been successfully used in many areas. In this study, we developed the optimized networks for urban land use classification from Landsat ETM+ images of Marina del Rey area based on USGS land cover/use classification level III. The networks started from a tree structure based on mutual information between variables and added the links to improve accuracy. This methodology offers several advantages: (1) The network structure shows the dependency relationships between variables. The class node value can be predicted even with particular band information missing due to sensor system error. The missing information can be inferred from other dependent bands. (2) The network structure provides information of variables that are important for the classification, which is not available from conventional classification methods such as neural networks and maximum likelihood classification. In our case, for example, bands 1, 5 and 6 are the most important inputs in determining the land use of each pixel. (3) The networks can be reduced with those input variables important for classification. This minimizes the problem without considering all possible variables. We also examined the effect of incorporating ancillary data: geospatial information such as X and Y coordinate values of each pixel and DEM data, and vegetation indices such as NDVI and Tasseled Cap transformation. The results showed that the locational information improved overall accuracy (81%) and kappa coefficient (76%), and lowered the omission and commission errors compared with using only spectral data (accuracy 71%, kappa coefficient 62%). Incorporating DEM data did not significantly improve overall accuracy (74%) and kappa coefficient (66%) but lowered the omission and commission errors. Incorporating NDVI did not much improve the overall accuracy (72%) and k coefficient (65%). Including Tasseled Cap transformation reduced the accuracy (accuracy 70%, kappa 61%). Therefore, additional information from the DEM and vegetation indices was not useful as locational ancillary data.
DE: 6344 System operation and management
DE: 1899 General or miscellaneous
DE: 1800 HYDROLOGY
DE: 1860 Runoff and streamflow
SC: Hydrology [H]
MN: 2004 AGU Fall Meeting