Estimating canopy nitrogen concentration across C3 and C4 grasslands using WorldView-2 multispectral data and the random forest algorithm

This paper assesses the potential of WorldView-2 (WV2) multispectral data to estimate and map the variability in canopy concentration of nitrogen (N), across C3 and C4 grasslands. The WV2 satellite image was acquired for the Cathedral Peak region of the Drakensberg Mountain range, South Africa. Random forest (RF) regression algorithm was used to develop a relationship between two-band vegetation indices (NDVIs) computed from the WV2 image and N concentration. The RF-based variable importance scores calculated using the training dataset (n = 150) showed that the NDVI computed involving the costal-blue and yellow bands is the most important, when predicting canopy concentration of N in the area. Using the validation dataset (n = 64), the RF explained 71% of the variation, with a Nash-Sutcliffe efficiency (NSE) = 0.68, in predicting N across the C3 grass, Festuca costata, and C4 grasses, Themeda triandra and Rendlia altera grasslands. Overall, results from this study suggest that new multispectral data with unique band setting, such as WV2, are capable of estimating or mapping N concentration.