Mapping Prosopis glandulosa (mesquite) invasion in the arid environment of South African using remote sensing techniques

A dissertation submitted to the School of Geography, Archaeology and Environmental Studies, Faculty of Science, University of the Witwatersrand, Johannesburg, in fulfilment of requirements for the degree of Master of Science in Environmental Sciences. Johannesburg, March 2016.Mapping Prosopis glandulosa (mesquite) invasion in the arid environment of South Africa using remote sensing techniques Mureriwa, Nyasha Abstract Decades after the first introduction of the Prosopis spp. (mesquite) to South Africa in the late 1800s for its benefits, the invasive nature of the species became apparent as its spread in regions of South Africa resulting in devastating effects to biodiversity, ecosystems and the socio-economic wellbeing of affected regions. Various control and management practices that include biological, physical, chemical and integrated methods have been tested with minimal success as compared to the rapid spread of the species. From previous studies, it has been noted that one of the reasons for the low success rates in mesquite control and management is a lack of sufficient information on the species invasion dynamic in relation to its very similar co-existing species. In order to bridge this gap in knowledge, vegetation species mapping techniques that use remote sensing methods need to be tested for the monitoring, detection and mapping of the species spread. Unlike traditional field survey methods, remote sensing techniques are better at monitoring vegetation as they can cover very large areas and are time-effective and cost-effective. Thus, the aim of this research was to examine the possibility of mapping and spectrally discriminating Prosopis glandulosa from its native co-existing species in semi-arid parts of South Africa using remote sensing methods. The specific objectives of the study were to investigate the spectral separability between Prosopis glandulosa and its co-existing species using field spectral data as well as to upscale the results to different satellites resolutions. Two machine learning algorithms (Random Forest (RF) and Support Vector Machines (SVM)) were also tested in the mapping processes. The first chapter of the study evaluated the spectral discrimination of Prosopis glandulosa from three other species (Acacia karoo, Acacia mellifera and Ziziphus mucronata) in the study area using in-situ spectroscopy in conjunction with the newly developed guided regularized random forest (GRRF) algorithm in identifying key wavelengths for multiclass classification. The GRRF algorithm was used as a method of reducing the problem of high dimensionality associated with hyperspectral data. Results showed that there was an increase in the accuracy of discrimination between the four species when the full set of 1825 wavelengths was used in classification (79.19%) as compared to the classification used by the 11 key wavelengths identified by GRRF (88.59%). Results obtained from the second chapter showed that it is possible to spatially discriminate mesquite from its co-existing acacia species and other general land-cover types at a 2 m resolution with overall accuracies of 86.59% for RF classification and 85.98% for SVM classification. The last part of the study tested the use of the more cost effective SPOT-6 imagery and the RF and SVM algorithms in mapping Prosopis glandulosa invasion and its co-existing indigenous species. The 6 m resolution analysis obtained accuracies of 78.46% for RF and 77.62% for SVM. Overall it was concluded that spatial and spectral discrimination of Prosopis glandulosa from its native co-existing species in semi-arid South Africa was possible with high accuracies through the use of (i) two high resolution, new generation sensors namely, WorldView-2 and SPOT-6; (ii) two robust classification algorithms specifically, RF and SVM and (iii) the newly developed GRRF algorithm for variable selection and reducing the high dimensionality problem associated with hyperspectral data. Some recommendations for future studies include the replication of this study on a larger scale in different invaded areas across the country as well as testing the robustness of the RF and SVM classifiers by making use of other machine learning algorithms and classification methods in species discrimination. Keywords: Prosopis glandulosa, field spectroscopy, cost effectiveness, Guided Regularised Random Forest, Support Vector Machines, Worldview-2, Spot-