Landcover classification in a heterogenous savanna environment : investigating the performance of an artificial neural network and the effect of image resolution.
The aim of this study was to investigate the role of spatial and categorical resolution of satellite images in landcover classification. Three images namely, SPOT 5, Landsat TM, and MODIS were used, each of varying spatial resolution. Landcover classes were chosen for each of the classifications, were placed into groups of 11, and then merged to 8. This was to evaluate the effect that the categorical resolution plays on the final classification algorithm. Three traditional classifiers were used to create landcover maps. It was found that the higher resolution imagery produced higher accuracies at the 11 class level and these accuracies were improved by reducing the number of classes to 8. The coarser resolution imagery was able to classify larger features more accurately than the smaller features. This allowed the conclusion to be drawn that, before classifications are to be done, the size of the features to be detected should be considered when deciding which imagery to use. To improve upon the accuracy of the maximum likelihood classifier, an Artificial Neural Network was trained using ancillary data and the SPOT 5 image. Results showed an increase of over 30% in the classification accuracy of the ANN. Specific classes were easily identified, showing the ability of the ANN to classify imagery from a complex savanna environment. Experiments with various parameters of the neural network confirmed that there are no general guidelines that can be applied to a neural network to obtain high classification accuracy.