Doctor of Philosophy (PhD)
Geography and Anthropology
Urban land-use/land-cover classification is entering a new era with the increased availability of high-resolution satellite imagery and new methods such as texture analysis and artificial intelligence classifiers. Recent research demonstrated exciting improvements of using fractal dimension, lacunarity, and Moran’s I in classification but the integration of these spatial metrics has seldom been investigated. Also, previous research focuses more on developing new classifiers than improving the robust, simple, and fast maximum likelihood classifier. The goal of this dissertation research is to develop a new approach that utilizes a texture vector (fractal dimension, lacunarity, and Moran’s I), combined with a new genetic Bayesian classifier, to improve urban land-use/land-cover classification accuracy. Examples of different land-use/land-covers using post-Katrina IKONOS imagery of New Orleans were demonstrated. Because previous geometric-step and arithmetic-step implementations of the triangular prism algorithm can result in significant unutilized pixels when measuring local fractal dimension, the divisor-step method was developed and found to yield more accurate estimation. In addition, a new lacunarity estimator based on the triangular prism method and the gliding-box algorithm was developed and found better than existing gray-scale estimators for classifying land-use/land-cover from IKONOS imagery. The accuracy of fractal dimension-aided classification was less sensitive to window size than lacunarity and Moran’s I. In general, the optimal window size for the texture vector-aided approach is 27x27 to 37x37 pixels (i.e., 108x108 to 148x148 meters). As expected, a texture vector-aided approach yielded 2-16% better accuracy than individual textural index-aided approach. Compared to the per-pixel maximum likelihood classification, the proposed genetic Bayesian classifier yielded 12% accuracy improvement by optimizing prior probabilities with the genetic algorithm; whereas the integrated approach with a texture vector and the genetic Bayesian classifier significantly improved classification accuracy by 17-21%. Compared to the neural network classifier and genetic algorithm-support vector machines, the genetic Bayesian classifier was slightly less accurate but more computationally efficient and required less human supervision. This research not only develops a new approach of integrating texture analysis with artificial intelligence for classification, but also reveals a promising avenue of using advanced texture analysis and classification methods to associate socioeconomic statuses with remote sensing image textures.
Document Availability at the Time of Submission
Secure the entire work for patent and/or proprietary purposes for a period of one year. Student has submitted appropriate documentation which states: During this period the copyright owner also agrees not to exercise her/his ownership rights, including public use in works, without prior authorization from LSU. At the end of the one year period, either we or LSU may request an automatic extension for one additional year. At the end of the one year secure period (or its extension, if such is requested), the work will be released for access worldwide.
Ju, Wenxue, "A Genetic Bayesian Approach for Texture-Aided Urban Land-Use/Land-Cover Classification" (2008). LSU Doctoral Dissertations. 572.