Electronic Theses and Dissertations (Masters)

Permanent URI for this collectionhttps://hdl.handle.net/10539/38009

Browse

Now showing 1 - 7 of 7

Assessing and comparing the performance of different machine learning regression algorithms in predicting Chlorophyll-a concentration in the Vaal Dam, Gauteng
(University of the Witwatersrand, Johannesburg, 2024-03) Mahamuza, Phemelo Hope; Adam, Elhadi
The state of Vaal Dam is influenced by various land uses surrounding the Dam, including agricultural activities, mining operations, industrial enterprises, urban settlements, and nature reserves. Mining activities, farming practices, and sewage outflows from nearby villages led to access contamination within the Dam, increasing algal bloom levels. Sentinel-2 MSI data were utilized to forecast and comprehend the spatial pattern of Chlorophyll-a concentration, indicating algal bloom occurrence in the Vaal Dam. Targeting Sentinel-2 Level-1C, the image was preprocessed on the Google Earth Engine (GEE) with acquisition dates from 25 – 26 October 30, 2016, corresponding to the on-site data collection between October 26 and October 28, 2016. Due to limited resources, up-to-date data on the Vaal Dam could not be collected. However, since this study focuses on applying various machine learning regression models to predict chlorophyll-a levels in waterbodies, the dataset is used to test the models rather than reflect the current state of the Vaal Dam. The dataset, comprising 23 samples, was divided into 70% training and 30% test sets, allowing for comprehensive model evaluation. Band ratio reflectance values were extracted from the satellite image and correlated with in-field Chlorophyll-a values. The highest correlation coefficient values were utilized to train five machine-learning models employed in this study: Random Forest (RF), Support Vector Regression (SVR), Least Absolute Shrinkage and Selection Operator (LASSO), Ridge Regression, and Multilinear Regression (MLR). Each model underwent training with ten iterations each; the best learning iteration was then used to generate the final Chlorophyll-a predictive model. The predictive models were validated using the Sentinel-2 MSI satellite data and in-situ measurements using R2, RMSE, and MAPE. Among the five machine learning algorithms trained, RF performed the best, with an R2 of 0.86 and 0.95, an RMSE of 1.38 and 0.8, and MAPE of 15.09% and 10.92% for the training and testing sets, respectively, indicating its ability to handle small, non-linear datasets. SVR also demonstrated a fair performance, particularly in handling multicollinearity in the data points with an R2 of 0.68 and 0.87, an RMSE of 2.37 and 1.56, and MAPE of 18.13% and 19.28% for the training and testing sets, respectively. The spatial pattern of Chlorophyll-a concentrations, mapped from the RF model, indicated that high concentrations of Chlorophyll-a are along the Dam shorelines, suggesting a significant impact of land use activities on pollution levels. This study emphasizes the importance of selecting suitable machine learning algorithms tailored to the dataset's characteristics. RF and SVR demonstrated proficiency in handling nonlinearity, with RF displaying enhanced generalization and resistance to overfitting. Limited field data evenly distributed across the Dam and satellite overpass dates may affect result accuracy. Future research should align satellite pass dates with fieldwork dates and ensure an even distribution of in-field samples across the Dam to represent all land uses and concentration levels.
Detecting Disease in Citrus Trees using Multispectral UAV Data and Deep Learning Algorithm
(University of the Witwatersrand, Johannesburg, 2024-06) Woolfson, Logan Stefan; Adam, Elhadi
There is a high prevalence, in South Africa, of fruit tree related diseases infesting lemon trees, subsequently affecting overall crop yield and quality. Ultimately, the income for the farmers is significantly diminished and limits the supply of nutritional food crops for the South African population, who already suffer from a high incidence of malnutrition. Currently, there are various methods utilized to detect diseases in fruit trees, however they pose limitations in terms of efficiency and accuracy. By employing the use of drones and machine learning methods, fruit tree diseases could be detected at an earlier stage of development and with a much higher level of accuracy. Consequently, the chances of remedying the trees before the disease spreads is greatly improved, and the supply of nutritious fruit within South Africa is increased. This research report’s aim is to investigate the effectiveness of a deep learning algorithm for detecting and classifying diseases in lemon orchards using multispectral drone imagery. This entails assessing the performance of a pretrained ResNet-101 model, fine-tuned with additional sample images, in accurately identifying and classifying diseased lemon trees, specifically those affected by Phytophthora root rot. The methodology involves the utilization of a pretrained ResNet-101 model, a deep learning architecture, and the retraining of its layers with an augmented dataset from multispectral aerial drone images of a lemon orchard. The model is fine-tuned to enhance its ability to discern subtle spectral variations indicative of disease presence. The selection of ResNet-101 is grounded in its proven success in image recognition tasks and transfer learning capabilities. The results obtained demonstrated an impressive accuracy of 80%. The deep learning algorithm exhibited notable performance in distinguishing root rot-affected lemon trees from their healthy counterparts. The findings indicate the promise of utilizing advanced deep learning methods for timely and effective disease detection in agricultural farmlands, facilitating orchard management.
Estimating rooftop solar energy potential using spatial radiation models and thermal remote sensing: The case of Witwatersrand University
(University of the Witwatersrand, Johannesburg, 2023) Ndemera, Rudo Hilda; Adem, Ali K.; Adam, Elhadi
The main purpose of this research was to estimate the University of Witwatersrand building’s rooftop solar energy potential using the GIS-based solar Area Solar Radiation (ASR) analyst upward hemispherical view shed algorithm. The two major datasets used in this research for rooftop solar energy potential modelling are building footprint data and the Digital Surface Model. Building footprint data, specifically rooftop area was extracted using machine learning CNTK unified toolkit and deep neural networks. The data was presented as individual polygon shape files for each building. The high-resolution Digital Surface Model imagery was sourced from the Advanced Land Observation Satellite. Pre-processing of the imagery was done for atmospheric correction. The DSM was then used in the Area Solar Radiation model to create an upward view shed for every point on the study area which is essential for computing solar radiation maps. The efficiency of using this algorithm is that it considers the shading effects caused by surrounding topography and surrounding man-made features, alterations in the azimuth angle and the position of the sun. Apart from the incoming solar radiation reaching the rooftops, the elevation and orientation of the rooftop cells limit the solar panel tilt angle and intensity of the incoming solar radiation, respectively. These factors were used in setting the suitability criteria together with solar radiation for the identification of suitable rooftop cells in this research. The relationship between land surface temperature and solar radiation values was assessed to determine if it can be used as an indicator for solar panel efficiency. Results from this research indicate that the University of Witwatersrand receives high levels of incoming solar radiation and has a high solar energy rooftop generation capacity that can meet the energy demand on campus. To improve accuracy of the research results, a drone could have been used to measure insolation across the study area to improve the spatial resolution. However, this was not possible due to various restrictions.
Evaluating the spatiotemporal changes of urban wetlands in Klip River wetland, South Africa
(University of the Witwatersrand, Johannesburg, 2023-09) Nxumalo, Nolwazi; Knight, Jasper; Adam, Elhadi
This study assesses the impacts of land use / land cover (LULC) change in an urban wetland over the past 30 years utilizing machine learning and satellite-based techniques. This study looked at LULC distributions in the Klip River wetland in Gauteng, South Africa. The aims and methods used in this study were: (1) to conduct a comprehensive analysis to map and evaluate the effects of LULC changes in the Klip River wetland spanning from 1990 to 2020, employing Landsat datasets at intervals of 10 years, and to quantify both spatial and temporal alterations in urban wetland area. (2) To predict the change in urban wetland area due to specific LULC changes for 2030 and 2040 using the MOLUSCE plugin in QGIS. This model is based on observed LULC including bare soil, built-up area, water, wetland, and other vegetation in the quaternary catchment C22A of the Klip River wetland, using multispectral satellite images obtained from Landsat 5 (1990), Landsat 7 (2000 and 2010) and Landsat 8 OLI (2020). (3) For the results of this study, thematic maps were classified using the Random Forest algorithm in Google Earth Engine. Change maps were produced using QGIS to determine the spatiotemporal changes within the study area. To simulate future LULC for 2030 and 2040, the MOLUSCE plugin in QGIS v2.8.18 was used. The overall accuracies achieved for the classified maps for 1990, 2000, 2010, and 2020 were 85.19%, 89.80%, 84.09%, and 88.12%, respectively. The results indicated a significant decrease in wetland area from 14.82% (6949.39 ha) in 1990 to 5.54% (2759.2 ha) in 2020. The major causes of these changes were the build-up area, which increased from 0.17% (80.36 ha) in 1990 to 45.96% (22 901 ha) in 2020—the projected years 2030 and 2040 achieved a kappa value of 0.71 and 0.61, respectively. The results indicate that built-up areas continue to increase annually, while wetlands will decrease. These LULC transformations posed a severe threat to the wetlands. Hence, proper management of wetland ecosystems is required, and if not implemented soon, the wetland ecosystem will be lost.
Investigating the impact of the land reform policy on land use and land cover changes, in Ngaka Modiri Molema district of the North West province
(University of the Witwatersrand, Johannesburg, 2024) Mmangoedi, Molebogeng Precious; Adam, Elhadi
The purpose of this study was to assess how land reform policies affected changes in land use and land cover in the province of North West's Ngaka Modiri Molema district municipality. The study employed remote sensing technologies to analyse changes in land use and land cover (LULC) resulting from the implementation of land reform programs between 1985 and 2015. The primary objective of the research was to systematically map Land Use and Land Cover types across five-year intervals spanning from 1985 to 2024, leveraging Landsat earth observation data in conjunction with a random forest classifier. These methodologies were employed to facilitate the identification of spatial patterns and trends associated with the implementation of land reform policies within the study area. Furthermore, the study utilized Landsat data and advanced change detection algorithms to quantitatively assess LULC changes over the specified timeframes. Through the application of spatial analysis techniques, the research aimed to elucidate the relationship between the implementation of land reform measures and corresponding shifts in LULC patterns across the research study area. The findings of the investigation indicated a noticeable expansion in built-up areas between the years 1985 and 2024 which was approximately 10.86%. This expansion was primarily attributed to the growth experienced by the municipality during this period. Additionally, more opportunities might have risen from the agricultural farming activities and also from the land reform policy being implemented. However, as the ownership changed due to land redistribution and more land was being acquired by black people through the land reform policy, agricultural farming decreased slightly throughout the years. The reduction was due to the factors that arose from inefficient policy implementation. The study also recommends that remote sensing techniques should be utilised to carry out studies to determine LULC changes that derive from land policies aiming at dealing with socio-economic factors and urbanisation. An incorporated agrarian reform sustainable programme has vast potential in cultivating the production of the projects, particularly if it involves packages in rural infrastructure, support services, and co-operatives. The major role of such an approach should be in the trainings conducted for the farmers, obtaining, and distributing agricultural resources and equipment to agrarian reform or beneficiaries of the land reform projects. Additionally, there should be an allowance for special grants which will be useful in supporting the government’s efforts.
Mapping and monitoring land transformation of Boane district, Mozambique (1980 – 2020), using remote sensing
(University of the Witwatersrand, Johannesburg, 2023) Dengo, Claudio Antonio; Atif, Iqra; Adam, Elhadi
Although natural and environmental factors play a significant role in land transformation, human actions dominate. Therefore, to better understand the present land uses and predict the future, accurate information describing the nature and extent of changes over time is necessary and critical, especially for developing countries. It is estimated that these countries will account for 50% of the world's population growth in the next few years. Hence, this research was an attempt to assess and monitor land cover changes in Boane, Mozambique, over the past 40 years and predict what to expect in the next 30 years. This district has been challenged by a fast-growing population and land use dynamic, with quantitative information, driving forces and impacts remaining unknown. Through a supervised process in a cloud base Google Earth Engine platform, a set of five Landsat images at ten-year intervals were classified using a random forest algorithm. Seven land classes, i.e., agriculture, forest, built-up, barren, rock, wetland and water bodies, were extracted and compared through a pixel-by-pixel process as one of the most precise and accurate methods in remote sensing and geographic information system applications. The results indicate an active alternate between all land classes, with significant changes observed within agriculture, forest and build-up classes. As it is, while agriculture (-26.1%) and forest (-21.4%) showed a continuously decreasing pattern, build-up class (45.8%) increased tremendously. Consequently, over 69% of the forest area and 59% of the agricultural area shifted into build-up, i.e., was degraded or destroyed. Similarly, the conversion of barren land area (57.2%) and rock area (47.3%) into build-up indicates that those areas were cleaned. The overall classification accuracy averaged 90% and a kappa coefficient of 0.8779 were obtained. The CA-Markov model, used to assess future land uses, indicates that build-up will continue to increase significantly, covering 60% of the total area. From this finding, the land cover situation in the next 30 years will be critical if no action is taken to stop this uncontrolled urban sprawl. An adequate land use plan must be drawn, clearly indicating the locations for different activities and actions for implementation.
Spatiotemporal characteristics of surface water in Sua Pan, Botswana, using Earth Observation data: 1992–2022
(University of the Witwatersrand, Johannesburg, 2024-10) Peplouw, Muchelene Tiara; Adam, Elhadi; Grab, Stefan
Surface water is a critical resource for sustaining both human and ecological health. However, climate change and human actions threaten its availability in semi-arid regions like Botswana. In addition, current research on monitoring and understanding surface water dynamics in Botswana lacks the application of remote sensing and machine learning. This highlights a crucial gap in knowledge that this study aims to address. This study investigates the spatiotemporal dynamics of land use/land cover (LULC) and surface water extent changes in Sua Pan, Botswana, from 1992 to 2022. Employing remote sensing, machine learning, and statistical techniques, the research offers valuable insights into the intricate relationships between land cover modifications, surface water variations, and climatic variables. Google Earth Engine (GEE) facilitated efficient analysis of Landsat imagery for LULC mapping. Random Forest (RF) effectively classified several land cover types within Sua Pan. To address the challenges of saline environments, a novel water index, the Saline Water Index (SWI), was developed specifically for Sua Pan. The McNemar statistical test compared the performance of SWI to established indices like the Modified Normalised Difference Water Index (MNDWI) and the Normalised Difference Salinity Index (NDSI). Surface water variations were analysed using homogeneity tests and the Mann-Kendall trend test. The relationships between hydro climatic data (rainfall, evapotranspiration, land surface temperature) retrieved from GEE and surface water area for both wet and dry seasons were evaluated using Pearson correlation coefficients and visualised by line and area graphs. Additionally, the influence of the El Niño Southern Oscillation (ENSO) on rainfall and surface water area was assessed using Analysis of Variance (ANOVA) to identify the specific ENSO phases that exert an influence. The findings demonstrate the effectiveness of GEE for LULC mapping with the RF algorithm, achieving moderate to high classification accuracy (65.2% - 90.69%) and Kappa coefficients (0.54 - 0.85). Surface water and bare area exhibited increasing trends (coefficients: 13.017 and 9.0609, respectively), whereas vegetation and salt hard pan showed decreasing trends (-16.786 and -5.3081, respectively). The newly developed SWI outperformed MNDWI and NDSI in detecting surface water, achieving the highest overall accuracy (94%) compared to MNDWI (64%) and NDSI (59%). The McNemar test confirmed no significant statistical difference between the SWI map and the validation dataset (p = 0.2673), while both MNDWI and NDSI maps showed significant differences (p < 0.0001). Utilising SWI, the study revealed that surface water was most prevalent in central and northeastern regions, with an average coverage of 33%. Seasonal homogeneity tests indicated a non-homogenous distribution of surface water area in wet seasons, with abrupt changes in 1994 and 2003. Conversely, dry seasons exhibited a homogenous distribution. The Mann-Kendall trend test identified a statistically significant (p-value = 0.01) but weak positive trend (tau = 0.329) for surface water areas in wet seasons. In contrast, the dry seasons displayed a non-significant (p-value = 0.734) and a very weak positive trend (tau = 0.043). Surface water area, rainfall, evapotranspiration, and temperature consistently increase during the wet seasons compared to the dry seasons. Notably, increased evapotranspiration significantly impacted surface water presence. ENSO exhibited no significant influence on either rainfall or surface water extent (p-value > 0.05 for both). These findings highlight the potential of earth observation data for real-time surface water monitoring in salt pans. The developed techniques offer valuable insights for policy decisions regarding environmental management and conservation efforts in Sua Pan. In addition, the study emphasises the importance of cost-effective approaches for water change assessment, particularly appropriate for under-resourced regions.