Deriving indices of landscape function from spectral reflectance of grassland and savanna on gold mines in South Africa

The aim of this study is to develop hyperspectral (HS) models using partial least squares regression (PLSR) for predicting indices of grassland and savanna ecological condition on deep-level gold mines in a semi-arid region. Landscape Function Analysis (LFA) indices (surface stability, infiltration and nutrient cycling) were derived from four, increasingly complex, vegetation types on each end of a disturbance continuum in the dry season (winter) and wet season (summer). PLSR models for one of the most structurally simple vegetation types (non-rocky grassland on the lowest rainfall mine in summer) produced the strongest validation CoD for indices predicting stability and nutrient cycling (R² = 0.70, P < 0.0001 and R² = 0.71, P < 0.0001 respectively), whereas the infiltration index had the strongest CoD for validation with HS data from non-rocky grassland on the highest rainfall mine in summer (R² = 0.63, P < 0.0001). Increasingly complex vegetation structure (rocky grassland and dolomite sinkhole woodland) had weaker validation CoDs for LFA indices. Combining all vegetation categories or mining regions in a model also weakened CoD.