Repository logo
 

The use of machine learning algorithms to assess the impacts of droughts on commercial forests in KwaZulu-Natal, South Africa.

Simple item page
dc.contributor.advisorLottering, Romano Trent.
dc.contributor.advisorHlatshwayo, Sizwe Thamsanqa.
dc.contributor.authorButhelezi, Mthokozisi Ndumiso Mzuzuwentokozo.
dc.date.accessioned2020-11-05T12:14:02Z
dc.date.available2020-11-05T12:14:02Z
dc.date.created2020
dc.date.issued2020
dc.descriptionMasters Degree. University of KwaZulu-Natal, Pietermaritzburg.en_US
dc.description.abstractDroughts are a non-selective natural disaster in that their occurrence can be in both high and low precipitation areas. However, this study acknowledged that droughts are more recurrent and a regular feature in arid and semi-arid climates such as that of Southern Africa. Some of these countries rely strongly on commercial forests for their gross domestic product (GDP), especially South Africa and Mozambique which means droughts pose a significant threat to their economy and the society that depends on this economy. The risks associated with droughts have consequently created an increased demand for an efficient method of analysing and investigating droughts and the impacts they impose on forest vegetation. Therefore, this study aimed to examine the effects of droughts on all commercial forests within the province of KwaZulu-Natal (KZN) at a catchment and provincial scale by employing Kernel Support Vector Machine (Kernel –SVM), Rotation Forests (RTF) and Extreme Gradient Boosting (XGBoost) algorithms. These were based on Landsat and MODIS derived vegetation and conditional drought indices. The main aim of this study was achieved by the following objectives: (i) to improve methods for classifying droughts; (ii) to achieve medium spatial resolution drought analysis using Landsat sensors; (iii) to determine the accuracy of machine learning algorithms (MLAs) when employed on remote sensing data and (iv) to improve the usability of conditional drought indices and vegetation indices. The results obtained there-after demonstrated that the objectives of this study were met. With the MLAs performing better when using conditional drought indices compared to vegetation indices, therefore, highlighting drawbacks already associated with vegetation indices. Where at the catchment scale, Kernel – support vector machine (SVM) produced an overall accuracy (OA) of 94.44% when based on conditional drought indices compared to 81.48% when based on vegetation indices. On the same scale, Rotation forests (RTF) produced 96.30% and 81.84% when using conditional drought indices and vegetation indices, respectively. At a provincial scale, RTF produced an OA of 76.6% and 70.7% when using conditional drought indices and vegetation indices respectively. This was compared to extreme gradient boosting (XGBoost) which produced an OA of 81.9% and 69.3% when using conditional drought indices and vegetation indices respectively. These results also indicate that it is possible to analyse droughts at provincial and catchment scale. Although the results presented in this study were promising, more research is still required to improve the applicability of MLAs in drought analysis.en_US
dc.description.notesDedication is listed on page iii.en_US
dc.identifier.urihttps://researchspace.ukzn.ac.za/handle/10413/18805
dc.language.isoenen_US
dc.subject.otherMachine learning algorithms.en_US
dc.subject.otherDrought.en_US
dc.subject.otherLandsat.en_US
dc.subject.otherMODIS.en_US
dc.subject.otherCommercial forests.en_US
dc.subject.otherSupport vector machine.en_US
dc.subject.otherRotation forests.en_US
dc.titleThe use of machine learning algorithms to assess the impacts of droughts on commercial forests in KwaZulu-Natal, South Africa.en_US
dc.typeThesisen_US

Files

Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
Buthelezi_Mthokozisi_Ndumiso_Mzuzuwentokozo_2020.pdf
Size:
5.02 MB
Format:
Adobe Portable Document Format
Description:
Download
License bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
license.txt
Size:
1.64 KB
Format:
Item-specific license agreed upon to submission
Description:
Download

Collections

Masters Degrees (Environmental Science)