Probabilistic methods for land cover change detection

Abstract

One of the most powerful tools we possess for global scale monitoring of the surface of the earth are spacebourne remote sensing satellites. Every day they capture vast amounts of data at different spatial, spectral and temporal resolutions which must be processed in order to generate meaningful insights. This thesis focuses on the particular problem of detecting if and when a particular region of land cover experiences a change from one type to another. This problem is made difficult by the fact that the majority of land cover on earth, especially vegetated, undergoes natural variations on both annual and inter-annual time scales driven by changes in season and climate. In this thesis we argue that the key to detecting unnatural changes as accurately and rapidly as possible is to do so with respect to a probabilistic model of the natural variations estimated for the particular region of interest. A method each for change detection, change point estimation and online change monitoring are proposed that follow this strategy. These methods are evaluated on reflectance time series from the Moderate Resolution Imaging Spectroradiometer (MODIS) for two change detection problems, detecting unplanned settlement expansion in South Africa and detecting deforestation of protected areas in Australia. In each case the proposed methods are shown to be effective and require little human supervision suggesting that this approach has potential for use in production systems.