Automatic classification of Google Earth images for a larger scale monitoring of bush encroachment in South Africa

Abstract

Bush encroachment of savannas and grasslands is a common form of land degradation in the rangelands of South Africa. To assess the carrying capacity of the land and to understand underlaying processes of bush encroachment, continuous monitoring of this phenomenon is needed. The aim of this study is to provide training sites for satellite-based monitoring of bush encroachment in South Africa on a medium spatial resolution satellite sensor (e.g. MODIS or Landsat) scale. Since field surveys are time consuming and of limited spatial extent, the satellite based creation of training sites using Google Earth images is intended. Training pixels for woody vegetation and non-woody land cover were manually digitized from 50 sample Google Earth images. A Random Forests model was trained to delineate woody from non-woody pixels. The results indicate a high performance of the model (AUC=0.97). The model was applied to a further 500 Google Earth images with a spatial extent of 250m×250m. The classified images form the database of training sites which can be used for larger scale monitoring of bush encroachment in South Africa.