An investigation into the potential application of multi- and hyperspectral remote sensing for the spectral characterisation of maize and related weeds in the Free State Province of South Africa.

MSc. (Geography)Growing concerns with regards to the environmental and economic impacts related to the application of herbicides to control the spread and abundance of weeds in agricultural crops have created a need for the development of novel agricultural management systems that are less dependent on herbicide usage and tillage. Such concerns have given rise to the need for the variable spatial treatment of croplands aimed at the minimization of requirements for the application of herbicides and the subsequent minimization of excess materials released into the surrounding environment. Remote sensing provides an opportunity for the fast and cost-effective delineation of weed patches in croplands over large areas where traditional scouting techniques would be impractical. The differences in spectral reflectance from different plants at certain wavelengths due to species specific variations in biochemical and physical characteristics is what lays the basis for the distinction of vegetation species within remotely sensed images and ultimately the potential detection of weed-species in croplands. This study investigates the potential spectral characterisation of maize and commonly occurring weed-species by (1) making use of reflectance spectra collected at leaf-level to identify statistically significant differences in reflectance between individual species throughout the visible (VIS), Near-Infrared (NIR) and Shortwave-Infrared (SWIR) regions of the electromagnetic spectrum, determining the potential of the Red-Edge Position (REP) and slope for this particular application and testing the accuracy at which reflectance spectra may be classified according to vegetation species based on spectral reflectance at specific wavebands and REP as input predictor variables, (2) testing the potential effect of mixed spectral responses and soil-background interference through the analysis of reflectance spectra collected at canopy-level, and (3) determining the potential effect of the spectral generalisation associated with multispectral reflectance through the analysis of spectral responses resampled to the spectral band designations of representative high spatial resolution multispectral sensors. The results showed that maize may be spectrally distinguished from all of the weed-species included in the analysis based on leaf-level hyperspectral reflectance throughout the Visible-to-Near-Infrared (V-NIR) and SWIR-regions of the electromagnetic spectrum, however, the unique characterisation of weed-species is not possible for all species and where it is possible, it is highly wavelength-specific and would require high spectral resolution hyperspectral data. The wavelengths most suitable for the spectral characterisation of maize-crops and weed species in the study area were identified as: 432.1nm, 528.2nm, 700.7nm, 719.4nm, 1335.1nm, 1508.1nm, 2075.8nm, 2164.5nm and 2342.2nm. The output predictor model was able to classify reflectance spectra associated with maize crops and weeds in the study area at an overall accuracy of 89.7 per cent and it was shown that the inclusion of the REP as predictor variable did not improve the overall accuracy of the classification, however, may be used to improve the classification accuracies of certain species..