Improving the potential of pixel-based supervised classification in the absence of quality ground truth data

Abstract

The accuracy of classified results is often measured in comparison withreference or ``ground truth{''} information. However, in inaccessible orremote natural areas, sufficient ground truth data may not becost-effectively acquirable. In such cases investigative measurestowards the optimisation of the classification process may be required.The goal of this paper was to describe the impact of various parameterswhen applying a supervised Maximum Likelihood Classifier (MLC) to SPOT 5image analysis in a remote savanna biome. Pair separation indicators andprobability thresholds were used to analyse the effect of training areasize and heterogeneity as well as band combinations and the use ofvegetation indices. It was found that adding probability thresholds tothe classification may provide a measure of suitability regardingtraining area characteristics and band combinations. The analysisillustrated that finding a balance between training area size andheterogeneity may be fundamental to achieving an optimum classifiedresult. Furthermore, results indicated that the addition of vegetationindex values introduced as additional image bands could potentiallyimprove classified products and that threshold outcomes could be used toillustrate confidence levels when mapping classified results.