Employing Big Data To Monitor Global Land Change

The ‘Blue Marble’ photograph of Earth, taken by the Apollo 17 crew in 1972, depicted the illuminated Earth in the black void of space, inspiring many to recognize our shared global environment.  In the same year, the first earth observation satellite, Landsat 1, was launched.  A series of Landsat missions have followed.  However, despite an over 40 year history, it was only recently that conditions were met to map global land change using Landsat data.  Cloud computing, advanced algorithms and progressive data policies together have made possible the efficient processing and characterization of the Landsat data record.  Concerning data policies, it was only in 1999 that Landsat began to employ a global strategy for imaging the entire land surface.  However, this record was not fully exploitable until 2008, when the USGS made the entire Landsat archive accessible, free of charge.  Instantly, over 1,000,000 images were available for analysis, when previous studies had typically employed a single Landsat image.  In lieu of an open Landsat archive, we had previously processed coarse spatial resolution data sets to map the global land surface at spatial resolutions of 100km, 8km, 1km, 0.5km and 0.25km.  To perform these analyses, we employed data mining algorithms, specifically decision trees, which were highly suited to the complex relationships between land cover and the multi-temporal, multi-spectral feature space of satellite imagery.  Landsat data, with a spatial resolution of 30m, represents a significant advance in quantifying land change, as 30m is a scale finer than most human-induced land dynamics.  Given data and methods, the primary limitation to global land change mapping was processing capability.  The advent of cloud computing, in our case the development of the Google Earth Engine, has allowed us to process and characterize the Landsat archive.  The Earth Engine contains the entire Landsat record and enables image processing operations to be performed in parallel across a large number of computers.  Our prototype exercise at the global scale focused on mapping forest cover extent and change.  A total of 20 terapixels of data were processed using one million CPU-core hours on 10,000 computers in order to characterize year 2000 percent tree cover and subsequent tree cover loss and gain through 2012.  This effort is the first of many forthcoming analyses that will quantify important land changes, including trends in urbanization, cropped area, glacier retreat, and other important dynamics.  After more than 40 years since the taking of the Blue Marble image and the onset of space-based earth observation, the operational monitoring of global land change has been realized.