Monitoring urban growth by means of multi-temporal time-series of optical Landsat data

Abstract

Rapid urban expansion in cities across the globe is forcing the development of appropriate methods to monitor the status as well as the historical evolution of urban areas to support applications like urban planning and management. Here, remote sensing data have proven to be an effective tool to delineate urban areas, since they provide satellite imagery for large areas at a relatively high temporal frequency with the additional advantage of free access to data archives comprising data back to 1972. This study proposes a novel fully automated classification system based on Support Vector Machines (SVM) to derive urban extent maps. For this purpose, multi-temporal Landsat time-series namely temporal statistics consisting of selected spectral and temporal indices, have been employed for the 8 sites included in the urban supersites initiative of the Group on Earth Observation (GEO) and the investigation periods 2002–2003 and 2013–2015. The proposed methodology includes (1.) the pre-processing of Landsat scenes and calculation of the temporal statistics, (2.) the enhancement of the Global Urban Footprint (GUF), which is applied for an automated and random collection of training samples, (3.) the collection of training points for a set of configurations to overcome impacts of randomness, (4.) the application of a majority voting strategy to obtain a final urban extent map, and (5.) the implementation of an extensive accuracy assessment. The derived results report the automated SVM based classification system to be quite promising and in addition it proved to be very robust since it resulted in high accuracies throughout all study areas. In general, the obtained overall accuracy and Kappa coefficient is always higher than 91.38 % and 0.827, respectively.