Mapping land degradation using remote sensing data and an unsupervised clustering algorithm in the eThekwini Metropolitan Area.

Masters Degree. University of KwaZulu-Natal, Pietermaritzburg.Land degradation is a major environmental problem facing South Africa and many other countries around the world. For proper management and adoption of best rehabilitation strategies, a compendious regional-scale assessment approach is needed to attain the full extent of the impairment. The aim of this study was to assess the spatial extent of land degradation with the use of GIS and remote sensing techniques in the eThekwini Metropolitan Area (EMA), KwaZulu-Natal, South Africa. The first objective was to review the status of land degradation in South Africa, as well as tracking of emerging trends in remote sensing and Geographic Information Systems research. Historically, in South Africa, land degradation has been associated with poverty and rurality. While conducting studies was also a challenge, demanding high human and economic resources. Although these studies were accurate and invaluable, most of them were too localized and highly difficult to replicate. The introduction of remote sensing has bought a new dimension with a timely spatial mapping of land degradation at regional scales. As a result, there thus been a sharp increase in remote sensing-based land degradation studies, this is also accompanied by the recent improvements in capabilities of remote sensors and associated GIS platforms. However, there is still a challenge of accessibility, especially for financial constricted regions such as the sub-Sahara of Africa. Most of the cutting-edge remote sensing data such as the hyperspectral and high spatial resolution imagery are highly expensive and therefore inaccessible to those not affording. However, the use of new-age medium resolution sensors is a potential solution. The second objection of this study was to detect and map the spatial distribution of land degradation in the EMA through use of Sentinel-2 derived vegetation indices (VIs) in conjunction with a hierarchical clustering algorithm. Data from Sentinel-2 was used to derive VIs used in this study, these are namely; NDVI, RVI, SAVI; and SARVI. The framework using Ward’s hierarchical clustering performed relatively good to produce 6 clusters that achieved an overall classification accuracy (OA) of 88.81% when mapping land-cover including land degradation. In this regard, land degradation achieved the highest classification accuracy of up to 100%, while water achieved the lowest at 63.33%. Although there was quite a significant difference in accuracies between different land-cover classes, overall, the results were still reasonably good with an error rate of 0.14 and Kappa Coefficient of 0.86. The results from this study, therefore, suggest that Ward’s unsupervised clustering approach is a suitable tool for mapping of complex land-cover classes, particularly land degradation