Mapping Urban Land Cover: A Machine Learning Approach Using Landsat
and Nighttime Lights
Ran Goldblatt and Gordon Hanson
School of Global Policy and Strategy, UC San Diego
The revolution in geospatial data is transforming how we study the growth and development of cities. As improved satellite imagery becomes available, new remote-sensing methods and machine-learning approaches have been developed to convert terrestrial Earth-observation data into meaningful information about the nature and pace of change of urban landscapes and human settlements. Urban areas can be detected in satellite imagery using machine-learning approaches, which typically rely on reference ground-truth data that mark urban features, either for training or for validation. Reference data are fundamental not only for mapping and assessing cross-sectional urbanization across space, but also for classification of urbanization over time. However, because they are expensive to collect, large-scale reference datasets are scarce. We present a low-cost machine-learning approach for pixel-based image classification of built-up areas at a high-resolution and large scale. Our methodology relies on data infusion of nighttime and daytime remotely sensed data for automatic collection of ground truth data, which we use for supervised pixel-based image classification of built-up land cover. We demonstrate the effectiveness of our methodology, which is implemented in Google Earth Engine, through the development of accurate 30m resolution maps that characterize built-up land cover in three diverse countries: India, Mexico, and the U.S. Our approach highlights the usefulness of data fusion techniques for studying the built environment and has broad implications for identifying the drivers of urbanization.
Thursday, November 2, 2017
11:00AM AP&M 2402
Center for Computational Mathematics9500 Gilman Dr. #0112La Jolla, CA 92093-0112Tel: (858)534-9056