Authors: Jacob Arndt*, Oak Ridge National Laboratory, Jeanette Weaver, Oak Ridge National Laboratory, Dalton Lunga, Oak Ridge National Laboratory, St Thomas LeDoux, Oak Ridge National Laboratory, Sarah Tennille, Oak Ridge National Laboratory
Topics: Geographic Information Science and Systems, Remote Sensing, Land Use
Keywords: machine learning, classification, remote sensing, urban, gis
Session Type: Paper
Presentation File: No File Uploaded
Urban land-use information such as the locations of industrial areas, commercial areas, and different types of residential areas often serves as a foundational datasource in studies that estimate intercity population density, electricity usage, energy consumption, socio-economic trends, and more. However, developing accurate urban land-use datasets in an automated fashion is a challenge due to the difficulty of extracting effective features from high resolution satellite imagery that model the defining attributes of the target land-use type. Urban land-use types are composed of structural patterns that are best described using spatial characteristics such as building size, orientation, density, uniformity, ease of access, and presence of road networks. These defining characteristics require contextual information beyond the scope of a single pixel and can be quantified using multiscale textural features which describe the spatial variation of pixel tones within localized areas of different scale. Textural features are powerful descriptors of the complex patterns that occur in urban areas and can be used with machine learning algorithms to help distinguish between different types of urban land-use in satellite images. Here, we evaluate the usefulness of various low- and high-level multiscale textural features derived from very high-resolution satellite images for classifying different types of land-use in five large cities with varying climate, cultural, and geographic characteristics.
To access contact information login