Detecting the Socioeconomic Conditions of Urban Neighborhoods through Wavelet Analysis of Remotely Sensed Imagery

• Main Author: Zhou, Guiyun
• Other Authors: Nina Lam
• Format: Others
• Language: en
• Published: LSU 2006
• Subjects: Geography & Anthropology
• Online Access: http://etd.lsu.edu/docs/available/etd-08072006-095041/

Wavelet analysis is an efficient approach to studying textural patterns at different scales. Artificial neural networks can learn very complex patterns in the data and could be an efficient classifier. However, whether wavelet analysis, in combination with artificial neural networks or other classifiers, can be used to detect the social-economic conditions of urban neighborhood is a key research question that needs further study. The hypotheses of this study were: 1) neural networks yielded higher classification accuracy than linear discriminant analysis and the minimum-distance classifier based on wavelet measures of urban land covers; 2) wavelet textural measures could be used to efficiently discriminate among urban neighborhoods of different social-economic conditions; 3) image resolution had great influences on the discrimination of urban neighborhoods; and 4) window size had great influences on the discrimination of urban neighborhoods. In addition, two technical problems related to the application of textural approach, including the edge effect and image segmentation problem, were examined. The results show that the new approach developed to reducing edge effects consistently achieved higher accuracy than the traditional moving-window approach. The post-segmentation integration scheme in the region-based splitting-and-merging segmentation procedures reflected all the segmented clusters identified by two or more textural measures and was helpful in identifying homogeneous regions in an image. Regarding the four hypotheses, (1) The minimum-distance classifier performed the worst. Neural networks were found to generally yield slightly better results than discriminant analysis but the difference was not statistically significant. The first hypothesis was shown to be invalid. (2) With a window size of 85m by 85m, an overall accuracy of 93.00% was achieved using band 2 and an overall accuracy of 96.83% was achieved using combination of band 2 and band 3. (3) The 1-foot resolution subsets were found to yield higher classification accuracy than the 0.9m resolution subsets and the 2.7m resolution subsets for band 2 and band 3 for the six neighborhoods in Baton Rouge, Louisiana. The differences were generally over 5%. (4) Window size was found to have great influences on the discrimination of urban neighborhoods. The larger the window size, the higher the classification accuracy.