Machine learning and multi-sensor based modelling of woody vegetation in the Molopo Area, South Africa

Abstract

Bush encroachment is a highly relevant environmental issue in South African savannas, influencing ecological processes as well as the grazing capacity of the land. However the drivers of bush encroachment are not yet fully revealed which can partly be attributed to the problem that large-scale data of woody vegetation cover are missing. Using a multi-scale and a multi-sensor approach, this study aimed at providing the status of woody vegetation cover for the Molopo Area in South Africa. Training data for woody vegetation was derived from unsupervised classification of high-resolution aerial image tiles. To derive spatially continuous estimates of fractional woody cover for the entire Molopo Area, Sentinel-1 and Sentinel-2 data were applied in a machine learning based modelling approach. Therefore, a database of training samples was generated by aggregating the classified aerial image tiles to the geometry of the Sentinel data. A Random Forest algorithm with a forward feature variable selection was then trained to relate the spectral and radar information to fractional woody cover. The model was applied to make spatial predictions of fractional woody cover at 10 m resolution for the entire Molopo Area for the year 2016. Spatial cross-validation revealed a prediction error in fractional cover of 12%. The derived model and cover data show the potential for upcoming time series analysis of Sentinel-based woody cover estimates which can serve as a basis to bring new insights into the drivers of bush encroachment.