An automated method for time-series human settlement mapping using Landsat data and existing land cover maps

Abstract

Automation of satellite-based human settlement mapping is highly needed to utilize historical archives of satellite data for urgent issues of urban development in global scale. We developed an automated algorithm to detect human settlement from Landsat satellite data. A machine learning algorithm, Local and Global Consistency (LLGC), was applied with improvements for remote sensing data. The algorithm enables to use existing coarse-resolution land cover maps as a training dataset so that any manual process is not required for preparation of training data. In addition, for better robustness against uncertainty in satellite data, we proposed a method to combine several LLGC results for several dates in a certain period from a target date into a single human settlement map with a pixel-based median composition among the input LLGC results. Combination of the methods enabled to develop timeseries human settlement maps using Landsat data, single ones of which could be affected by cloud contaminations. We applied the algorithm to Landsat data of 838 WRS2 tiles for cities with more than one million people worldwide for 1990, 2000, 2005, and 2010. MCD12Q1, a MODIS-based global land cover map with 500-m resolution, was used as training data. Visual assessment of the results suggested next steps for improvement of the method.