Urban Small Area Population Estimation Based On High-resolution Remote Sensing Data

Small area population statistics is not only essential for understanding political, socioeconomic, cultural, and environmental issues at the city level but also meaningful for the construction of "Smart Earth". In the era of numerous earth observation systems on orbit, remote sensing has been widely used as an important means for population estimation in urban areas. Recent availability of sub-meter resolution remote sensing images makes the estimation of fine-scale population data possible, which is urgently needed in many fields to cope with the rapid urban development. Following this line of research, the thesis proposed to use residential buildings as an operating linkage between remote sensing data and small area population. That is, the geometric information of urban residential buildings is extracted through some computer-based algorithms from the high-resolution remote sensing imagery, and estimation models can be built to represent and test the relationship between population and the derived building geometry.This thesis attempted to accomplish two major tasks towards high-resolution remote sensing based population estimation for residential districts. The first task involved developing proper methods to extract the footprint and height of residential buildings from high-resolution RS imagery. Several existing approaches were examined, and their strength and weak points as well as applicability to this task were identified. In view of the requirements of this project and data characteristics, a multivariate image segmentation method based on the object-oriented view was chosen to better delineate building outlines by minimizing the salt-and-pepper effects due to heterogeneous spectra commonly found on building roofs. On the other hand, building heights were measured using the satellite metadata about solar height and the building shadows on the image. The estimated building area and height were validated with field-collected sample data for quality control and quality assurance purposes.The second task is related to model construction for small-area population estimation. Previous works in this area were reviewed, and pros and cons of major models were identified. Combined with field-collected sample data, three variants of the population estimation model were constructed by formulating a relationship between per-capita living space characterized by extracted building geometry (i.e. area and height) and the actual occupancies in the building. The model results were compared among the three variants with respect to estimation errors, which were further analyzed in relation to building types (divided as single-story or multiple-story) and heights (represented by the number of stories).The best model variant seemed to relate to the fact that it was population density rather than the absolute number of residents that is more sensitive to the extracted building geometry.The results indicate that the population in small urban area, which taking neighborhood committee as unit, can be estimated accurately using the population estimation model proposed here based on high-resolution remote sensing data. And the automation of estimation processes, to a certain extent, is improved by building information extraction and classification.