Evaluating the performance of Pixel-based vs. Object-based classifiers for Extracting High Resolution Land cover product from SPOT 6 imagery

Remote sensing measurements provide an accurate and timeous record of the landscape components. This enables the extraction of important information from images pertaining to spatial distribution of land cover captured by various earth observation satellites. Hence, up-to-date land cover information is of critical importance to planners, scientists, resource managers, decision-makers and various applications. The choice of an optimum sensor for land cover mapping is influenced by its spatial, temporal resolution and spectral resolution. Therefore the availability of new earth observation data, such as SPOT 6, which has improved spatial and spectral resolutions from its predecessor, revolutionize the utility of contemporary methods for land cover mapping. Recognizing the opportunity to map land cover at high resolution from SPOT 6, this evaluated the conventional pixel-based classification methods (namely; supervised Maximum Likelihood Classifier and unsupervised ISODATA classifier) and object-based (K-Nearest Neighbour) classification methods for their performance in mapping land cover in a semi-arid environment. Semi-arid environments are usually comprised of discrete land cover units, as a result, sophisticate land cover mapping activities since the resultant classes are largely generalised especially when using high spatial resolution data. This was identified as a gap; and therefore, was addressed. The study found that object-based K-NN classifier performed better than pixel-based ISODATA and Maximum Likelihood classifiers, with overall accuracies of 82%, 50.7% and 36.7% respectively. We concluded that SPOT 6 data, coupled with object-based classification methods have potential to provide rapid land cover classifications for South Africa