Taona, Mazarire Theresa2019-09-092019-09-092019https://hdl.handle.net/10539/28057A research report submitted in partial fulfillment of the requirements for the degree in Master of Science in GIS and Remote Sensing Faculty of Science University of the Witwatersrand 2019 Johannesburg, South AfricaDetailed and accurate studies designed to map crop types play a strategic role in giving information about proper management strategies for precision agriculture, crop monitoring, crop yield estimation, and food security management. The availability of moderate to high resolution remote sensing imageries such as Sentinel-2 in the last few years with less or no cost has enormously increased their use for research and other applications. Together with the latest Landsat mission, Landsat-8, these sensors provide reliable, inexpensive and timely data coupled with a wide range of spatial, temporal and spectral characteristics suitable for mapping the dynamics in agriculture landscape worth exploring. Endeavors to improve crop type maps accuracy have seen the progression of machine learning algorithms to further advance the image classification techniques. Regardless of these efforts, crop classifications generated from these techniques are still deemed as inadequate for agriculture applications, due to the slight inconsistency between the derived classified maps and the information on the ground. In this regard, this research pursued two main objectives: firstly, to compare and explore the utility of Sentinel-2 and Landsat-8 imageries for mapping crop types, secondly, to test the two machine learning algorithms- Random Forest (RF) and Support Vector Machine (SVM) performance and efficiency in classifying different crop types and discriminating them from the co-existing land use classes in a heterogeneous agriculture landscape in Marble Hall, Limpopo, South Africa. Multi-temporal images from both sensors (from June to September 2017) were processed in order to map the crop types and co-existing land use classes. Fourteen classes in total were classified using both RF and SVM. The optimum time based on the study was July-August as this is the peak time for crop development. The utility and contribution of different bands in each sensor for classification were evaluated using RF mean decrease Gini for variable importance. Vegetation red edge, SWIR, and blue bands contributed the most in classifying Sentinel-2 data whilst the nearinfrared, panchromatic and aerosol bands were mostly utilized in Landsat-8 data. To measure the accuracy of the generated thematic maps, accuracy assessment was undertaken with respective independent validation data sets. The best performance was obtained from August 2017 Sentinel2 data classification achieving Overall Accuracy (OA) of 95.4% and a kappa value of 0.94%. The lowest accuracy was obtained from September 2017 Landsat-8 data classification which obtained an overall accuracy of 91.19 % and 87 % with a kappa coefficient of 0.9 and 0.85 using SVM and RF respectively. The present study found that RF and SVM classifiers performed similarly though SVM outperformed RF by 1.5%. Overall, it was concluded that Sentinel-2 data performs better than Landsat-8 data in discriminating crop type classes and co-existing land use classes in a heterogeneous landscape because of its high spectral and spatial resolution. SVM was the most efficient and best classifier due to its capability to process a few training samples using limited parameters. Consequently, this research revealed the ability of medium resolution multispectral data in crop type classification using the propagated image classification techniques. Conclusively, the need to employ more advanced image classification techniques and utilize diffusion of radar and optical data for crop type mapping on complex landscapes remain paramount.enThesis