Bush encroachment

mapping and monitoring in 

South Africa


This research, mainly funded by an MMU PhD Scholarship, investigates woody vegetation encroachment in the Limpopo Province of South Africa.

Increasing attention is being directed at mapping the fractional woody cover of savannahs using Earth-observation data. In this study, we test the utility of Landsat TM/ ETM-based spectral-temporal variability metrics for mapping regional-scale woody cover in the Limpopo Province of South Africa, for 2010 (Higginbottom et al. 2018). We employ a machine learning framework to compare the accuracies of Random Forest models derived using metrics calculated from different seasons. We compare these results to those from fused Landsat-PALSAR data to establish if seasonal metrics can compensate for structural information from the PALSAR signal. Furthermore, we test the applicability of a statistical variable selection method, the recursive feature elimination (RFE), in the automation of the model building process in order to reduce model complexity and processing time. All of our tests were repeated at four scales (30, 60, 90, and 120 m-pixels) to investigate the role of spatial resolution on modelled accuracies.

The work on woody vegetation mapping and monitoring is ongoing. We are currently using Deep Learning (Convolutional Neural Networks) to map multi-temporal woody vegetation encroachment in the neighbouring Northwest Province with the aim to identify degrading areas where mitigation measures are required.


The Limpopo Province
within South Africa & the Landsat scenes that correspond to its area

To create training data the six aerial imagery subsets were classified into woody/non woody masks. We opted for aerial image classification to enable methods to be transferable to other locations, due to the generally satisfactory availability of aerial imagery at appropriate scales (Staben et al., 2016). Firstly, a principal components analysis (PCA) was applied to the three RGB layers and the first two components were extracted. Secondly, we calculated the visible vegetation index (Joseph and Devadas, 2015Ludwig et al., 2016) which uses visible light spectra to estimate photosynthetic activity. A Random Forest classifier was used to create the binary woody-non woody layers from the original RGB layers, principle components, and VVI. Individual modelswere generated for each image using 400 manually selected points per image (75/25% training-validation split). The mean classification accuracy was 85%. Full accuracy statistics are given in the Appendix A. An example classified mask is shown in Fig. 4.


of the RGB woody classification

(a) Raw RGB image; (b) classified woody cover shown in red, and (c) 30 m grid for fractional cover sampling.

fractional woody cover results

for the Limpopo Province based on the Recursive Feature Elimination model at the 120 m pixel scale. Black squares A and B are the locations of the subsets in the figure below

Spatial patterns of woody cover
for subsets A and B of the previous figure, at 30 and 120 m pixel scales. 

Five model predictions and the respective reference aerial imagery from the NGI are shown. Aerial imagery acquisition dates: A: 19 April 2009, B: 30 April 2009.