Remotely sensed indicators or urban land use intensity: Comparison of sub-pixel analysis techniques

The goal of this dissertation is to investigate novel methods of remote sensing/geographic information system (GIS) technologies to improve the accuracy of mapping urban land cover. Medium spatial resolution remotely sensed imagery is comparatively very cheap, but has a critical drawback “mixed” pixels (i.e., mixtures of impervious surface, lawn and tree cover with a single pixel) in the complex urban landscape. Accordingly, there are two major research areas that I propose to address: (1) Improving the specificity and accuracy of remotely sensed indicators of human land use, with a focus on impervious surface, lawn and urban tree cover; and (2) Testing the utility of newly available high (IKONOS) and medium (Landsat ETM) resolution remotely sensed image data for such purposes. While previous studies have focused on the estimation of impervious surface, this study is the first to thoroughly investigate the lawn and tree cover as separate urban green space components.; I tested three different sub-pixel analysis methods: Linear Mixture Model (LMM), Fuzzy c-means Clustering (FCM), and Self-Organizing Map Neural Network (SOM). Overall, the SOM method provided the best estimates of the three land cover components: impervious surface estimated ranged from ±4∼12%, lawn ranged from ±8∼11%, and tree ranged from ±11∼19% as compared to reference data. The linear mixture assumption of the endmember spectra of LMM is upheld to a large extent as evidenced by the rather high accuracy of impervious surface estimation, but the spectral reflectance of lawn and urban tree are not linearly mixed. LMM and FCM do not correctly estimate pure pixels of lawn and urban tree, while SOM_LVQ estimates these pure pixels quite accurately. Providing higher spatial resolution by the merging of higher spatial resolution panchromatic and lower spatial resolution multispectral Landsat ETM imagery did not improve the estimation of urban land cover components.; The results of this study provide comprehensive information of the utility of sub-pixel analysis for the estimation of urban land cover components and suggest that the comparatively accurate land cover estimation of urban land cover components is attainable from medium resolution satellite imagery. These results are significant in that they demonstrate that medium resolution remotely sensed imagery such as Landsat ETM can provide a cost effective image data source for urban monitoring.