Discriminant Analysis with Spatial Weights for Urban Land Cover Classification
Soe Myint and
William Stefanov Additional contact information Elizabeth Wentz: GeoDa Center for Geospatial Analysis and Computation; Arizona State University, http://geography.asu.edu/wentz Soe Myint: GeoDa Center for Geospatial Analysis and Computation; Arizona State University, http://www.public.asu.edu/~smyint/
Classifying urban area images is challenging because of the heterogeneous nature of the urban landscape resulting in mixed pixels and classes with highly variable spectral ranges. Approaches using ancillary data, such as knowledge based or expert systems, have shown to improve the classification accuracy in urban areas. Appropriate ancillary data, however, may not always be available. The goal of this study is to compare the results of the discriminant analysis statistical technique with discriminant analysis with spatial weights to classify urban land cover. Discriminant analysis is a statistical technique used to predict group membership for a target based on the linear combination of independent variables. Strict per pixel statistical analysis however does not consider the spatial dependencies among neighbouring pixels. Our study shows that approaches using ancillary data continue to outperform strict spectral classifiers but that using a spatial weight improved the results. Furthermore, results show that when the discriminant analysis technique works well then the spatially weighted approach performs better. However, when the discriminant analysis performs poorly, those poor results are magnified in the spatially weighted approach in the same study area. The study shows that for dominant classes, adding spatial weights improves the classification accuracy.