Mapping a Brazilian deforestation frontier using multi-temporal TerraSAR-X data and supervised machine learning

Abstract

Satellite remote sensing enables a repeated survey of the earth’s surface. With machine learning it is possible to recognize complex patterns from extensive data sets. Using methods from machine learning, remote sensing images are utilized to derive large scale land use and land cover (LULC) maps, carrying discrete information on the human management of land and intact primary forests, as well as change processes. Such information is particularly relevant in little developed regions, and areas which are undergoing transformation. Therefore, satellite remote sensing is generally the preferred method for generating LULC products within tropical regions, and particularly useful to assist tracking of change processes with regard to deforestation or land management. The Amazon is the largest area of continuous tropical forest in the world, and of substantial importance with regard to biodiversity, its influence on global climate, as well as providing living space for a large number of indigenous tribes. As tropical region, the Amazon is particularly affected by cloudy conditions, which pose a serious challenge to many remote sensing efforts. Utilization of Synthetic Aperture Radar (SAR) hence is promoted, as this warrants data availability at fixed intervals. Performing land cover mapping at the deforestation frontier in the Brazilian states of Pará and Mato Grosso, the aim of this thesis is to evaluate latest concepts from machine learning and SAR remote sensing in the light of real world applicability. As a cumulative effort, this thesis provides a scalable method based on Markov Random Fields, to increase classification performance. This method is especially useful to enhance the outcome of SAR classifications, as it directly addresses inherent SAR properties such as multi-temporality and speckle. Furthermore, ALOS-2, RADARSAT-2, and TerraSAR-X, which are current SAR sensors fulfilling different properties with regard to ground resolution and wavelength, are being investigated concerning their synergetic potentials for the mapping of vegetated LULC classes of the Brazilian Amazon. Here, the additional value of combining multiple frequencies is evaluated using reliable validation techniques based on area adjustment. Additionally, single performance of the three sensors is evaluated and their potentials concerning the task of tropical mapping are estimated. Lastly, different potentials of TanDEM-X for the purpose of tropical mapping are investigated. TanDEM-X is the first continuous spaceborne missionvi to offer a bi-static acquisition of data, enabling the generation of height models and the collection of coherence layers via a single pass.