Isoa, Mary Itohan2018-09-102018-09-102018Isoa, Mary Itohan, (2018) Exploring issues of balanced versus imbalanced samples in mapping grass community in the telperion reserve using high resolution images and selected machine learning algorithm, University of the Witwatersrand, Johannesburg, https://hdl.handle.net/10539/25639.https://hdl.handle.net/10539/25639Accurate vegetation mapping is essential for a number of reasons, one of which is for conservation purposes. The main objective of this research was to map different grass communities in the game reserve using RapidEye and Sentinel-2 MSI images and machine learning classifiers [support vector machine (SVM) and Random forest (RF)] to test the impacts of balanced and imbalance training data on the performance and the accuracy of Support Vector Machine and Random forest in mapping the grass communities and test the sensitivities of pixel resolution to balanced and imbalance training data in image classification. The imbalanced and balanced data sets were obtained through field data collection. The results show RF and SVM are producing a high overall accuracy for Sentinel-2 imagery for both the balanced and imbalanced data set. The RF classifier has yielded an overall accuracy of 79.45% and kappa of 74.38% and an overall accuracy of 76.19% and kappa of 73.21% using imbalanced and balanced training data respectively. The SVM classifier yielded an overall accuracy of 82.54% and kappa of 80.36% and an overall accuracy of 82.21% and a kappa of 78.33% using imbalanced and balanced training data respectively. For the RapidEye imagery, RF and SVM algorithm produced overall accuracy affected by a balanced data set leading to reduced accuracy. The RF algorithm had an overall accuracy that dropped by 6% (from 63.24% to 57.94%) while the SVM dropped by 7% (from 57.31% to 50.79%). The results thereby show that the imbalanced data set is a better option when looking at the image classification of vegetation species than the balanced data set. The study recommends the implementation of ways of handling misclassification among the different grass species to improve classification for future research. Further research can be carried out on other types of high resolution multispectral imagery using different advanced algorithms on different training size samples.Online resource (79 leaves)enVegetation and climate--Remote sensingVegetation mappingThesis