School Mathematics, Statistics and Computer Science

Permanent URI for this communityhttps://hdl.handle.net/10413/6526

Browse

Now showing 1 - 19 of 19

Analysis of discrete time competing risks data with missing failure causes and cured subjects.
(2023) Ndlovu, Bonginkosi Duncan.; Zewotir, Temesgen Tenaw.; Melesse, Sileshi Fanta.
This thesis is motivated by the limitations of the existing discrete time competing risks models vis-a-vis the treatment of data that comes with missing failure causes or a sizableproportions of cured subjects. The discrete time models that have been suggested to date (Davis and Lawrance, 1989; Tutz and Schmid, 2016; Ambrogi et al., 2009; Lee et al., 2018) are cause-specific-hazard denominated. Clearly, this fact summarily disqualifies these models from consideration if data comes with missing failure causes. It is also a well documented fact that naive application of the cause-specific-hazards to data that has a sizable proportion of cured subjects may produce downward biased estimates for these quantities. The existing models can be considered within the multiple imputation framework (Rubin, 1987) for handling missing failure causes, but the prospects of scaling them up for handling cured subjects are minimal, if not nil. In this thesis we address these issues concerning the treatment of missing failure causes and cured subjects in discrete time settings. Towards that end, we focus on the mixture model (Larson and Dinse, 1985) and the vertical model (Nicolaie et al., 2010) because these models possess certain properties which dovetail with the objectives of this thesis. The mixture model has been upgraded into a model that can handle cured subjects. Nicolaie et al. (2015) have demonstrated that the vertical model can also handle missing failure causes as is. Nicolaie et al. (2018) have also extended the vertical model to deal with cured subjects. Our strategy in this thesis is to exploit both the mixture model and the vertical model as a launching pad to advance discrete time models for handling data that comes with missing failure causes or cured subjects.
Application of ELECTRE algorithms in ontology selection.
(2022) Sooklall, Ameeth.; Fonou-Dombeu, Jean Vincent.
The field of artificial intelligence (AI) is expanding at a rapid pace. Ontology and the field of ontological engineering is an invaluable component of AI, as it provides AI the ability to capture and express complex knowledge and data in a form that encourages computation, inference, reasoning, and dissemination. Accordingly, the research and applications of ontology is becoming increasingly widespread in recent years. However, due to the complexity involved with ontological engineering, it is encouraged that users reuse existing ontologies as opposed to creating ontologies de novo. This in itself has a huge disadvantage as the task of selecting appropriate ontologies for reuse is complex as engineers and users may find it difficult to analyse and comprehend ontologies. It is therefore crucial that techniques and methods be developed in order to reduce the complexity of ontology selection for reuse. Essentially, ontology selection is a Multi-Criteria Decision-Making (MCDM) problem, as there are multiple ontologies to choose from whilst considering multiple criteria. However, there has been little usage of MCDM methods in solving the problem of selecting ontologies for reuse. Therefore, in order to tackle this problem, this study looks to a prominent branch of MCDM, known as the ELimination Et. Choix Traduisant la RÉalite (ELECTRE). ELECTRE is a family of decision-making algorithms that model and provide decision support for complex decisions comprising many alternatives with many characteristics or attributes. The ELECTRE algorithms are extremely powerful and they have been applied successfully in a myriad of domains, however, they have only been studied to a minimal degree with regards to ontology ranking and selection. In this study the ELECTRE algorithms were applied to aid in the selection of ontologies for reuse, particularly, three applications of ELECTRE were studied. The first application focused on ranking ontologies according to their complexity metrics. The ELECTRE I, II, III, and IV models were applied to rank a dataset of 200 ontologies from the BioPortal Repository, with 13 complexity metrics used as attributes. Secondly, the ELECTRE Tri model was applied to classify the 200 ontologies into three classes according to their complexity metrics. A preference-disaggregation approach was taken, and a genetic algorithm was designed to infer the thresholds and parameters for the ELECTRE Tri model. In the third application a novel ELECTRE model was developed, named ZPLTS-ELECTRE II, where the concept of Z-Probabilistic Linguistic Term Set (ZPLTS) was combined with the traditional ELECTRE II algorithm. The ZPLTS-ELECTRE II model enables multiple decision-makers to evaluate ontologies (group decision-making), as well as the ability to use natural language to provide their evaluations. The model was applied to rank 9 ontologies according to five complexity metrics and five qualitative usability metrics. The results of all three applications were analysed, compared, and contrasted, in order to understand the applicability and effectiveness of the ELECTRE algorithms for the task of selecting ontologies for reuse. These results constitute interesting perspectives and insights for the selection and reuse of ontologies.
Bayesian spatio-temporal and joint modelling of malaria and anaemia among Nigerian children aged under five years, including estimation of the effects of risk factors.
(2023) Ibeji, Jecinta Ugochukwu.; Mwambi, Henry Godwell.; Iddrisu, Abdul-Karim.
Childhood mortality and morbidity in Nigeria have been linked to malaria and anaemia. This thesis focused on exploring the risk factors and the complexity of the relationship between malaria and anaemia in under 5 Nigerian children. Data from the 2010 and 2015 Nigeria Malaria Indicator Survey conducted by Demographic Health Survey were used. In 2010, the prevalence of malaria and anaemia was 48% and 72%, respectively, while in 2015, 27% and 68% were the respective prevalences of malaria and anaemia diseases. Machine learning-based exploratory classification methods were used to explain the relationship and patterns between the independent variables and the two dependent variables, namely malaria and anaemia. Decisions made by the public health body are centered on the administrative units (i.e., states) within the country. Therefore, the development of disease mapping and a brief overview of limiting assumptions and ways of tackling them was explained. Consequently, malaria and anaemia spatial variation for 2010 and 2015 was analyzed with the inclusion of their respective risk factors. A separate multivariate hierarchical Bayesian logistic model for each disease was adopted to investigate the spatial pattern of malaria and anaemia and adjust for the risk factors associated with each disease. Furthermore, a multilevel model analysis was applied to independently investigate the spatio-temporal distribution of malaria and anaemia. A joint model was further adopted to check for the relationship between malaria and anaemia and their common risk factors and relax the nonlinearity assumption. In the 2010 data, type of place of residence, mother’s highest educational level, source of drinking water, type of toilet facility, child’s sex, main floor material, and households that have electricity, radio, television, and water were significantly associated with malaria and anaemia. While in the 2015 data, the type of place of residence, source of drinking water, type of toilet facility, households with radio, main roof material, wealth index, child’s sex, and mother’s highest educational level had a significant relationship with malaria and anaemia. The results from this study can guide policymakers to tailor-make effective interventions to reduce or prevent malaria and anaemia diseases. This will help adequately distribute limited state health system resources, such as personnel, funds and facilities within the country.
Deep learning for brain tumor segmentation and survival prediction.
(2024) Magadza, Tirivangani Batanai Hendrix Takura.; Viriri, Serestina.
A brain tumor is an abnormal growth of cells in the brain that multiplies uncontrolled. The death of people due to brain tumors has increased over the past few decades. Early diagnosis of brain tumors is essential in improving treatment possibilities and increasing the survival rate of patients. The life expectancy of patients with glioblastoma multiforme (GBM), the most malignant glioma, using the current standard of care is, on average, 14 months after diagnosis despite aggressive surgery, radiation, and chemotherapies. Despite considerable efforts in brain tumor segmentation research, patient diagnosis remains poor. Accurate segmentation of pathological regions may significantly impact treatment decisions, planning, and outcome monitoring. However, the large spatial and structural variability among brain tumors makes automatic segmentation a challenging problem, leaving brain tumor segmentation an open challenge that warrants further research endeavors. While several methods automatically segment brain tumors, deep learning methods are becoming widespread in medical imaging due to their resounding performance. However, the boost in performance comes at the cost of high computational complexity. Therefore, to improve the adoption rate of computer-assisted diagnosis in clinical setups, especially in developing countries, there is a need for more computational and memoryefficient models. In this research, using a few computational resources, we explore various techniques to develop deep learning models accurately for segmenting the different glioma sub-regions, namely the enhancing tumor, the tumor core, and the whole tumor. We quantitatively evaluate the performance of our proposed models against the state-of-the-art methods using magnetic resolution imaging (MRI) datasets provided by the Brain Tumor Segmentation (BraTS) Challenge. Lastly, we use segmentation labels produced by the segmentation task and MRI multimodal data to extract appropriate imaging/radiomic features to train a deep learning model for overall patient survival prediction.
Deep learning framework for speech emotion classification.
(2024) Akinpelu, Samson Adebisi.; Viriri, Serestina.
A robust deep learning-based approach for the recognition and classification of speech emotion is proposed in this research work. Emotion recognition and classification occupy a conspicuous position in human-computer interaction (HCI) and by extension, determine the reasons and justification for human action. Emotion plays a critical role in decision-making as well. Distinguishing among various emotions (angry, sad, happy, neutral, disgust, fear, and surprise) that exist from speech signals has however been a long-term challenge. There have been some limitations associated with existing deep learning techniques as a result of the complexity of features from human speech (sequential data) which consists of insufficient label datasets, Noise and Environmental Factors, Cross-cultural and Linguistic Differences, Speakers’ Variability and Temporal Dynamics. There is also a heavy reliance on huge parameter tunning, especially for millions of parameters before the model can learn the expected emotional features necessary for classification emotion, which often results in computational complexity, over-fitting, and poor generalization. This thesis presents an innovative deep learning framework-based approach for the recognition and classification of speech emotions. The deep learning techniques currently in use for speech-emotion classification are exhaustively and analytically reviewed in this thesis. This research models various approaches and architectures based on deep learning to build a framework that is dependable and efficient for classifying emotions from speech signals. This research proposes a deep transfer learning model that addresses the shortcomings of inadequate training datasets for the classification of speech emotions. The research also models advanced deep transfer learning in conjunction with a feature selection algorithm to obtain more accurate results regarding the classification of speech emotion. Speech emotion classification is further enhanced by combining the regularized feature selection (RFS) techniques and attention-based networks for the classification of speech emotion with a significant improvement in the emotion recognition results. The problem of misclassification of emotion is alleviated through the selection of salient features that are relevant to emotion classification from speech signals. By combining regularized feature selection with attention-based mechanisms, the model can better understand emotional complexities and outperform conventional ML model emotion detection algorithms. The proposed approach is very resilient to background noise and cultural differences, which makes it suitable for real-world applications. Having investigated the reasons behind the enormous computing resources required for many deep learning based methods, the research proposed a lightweight deep learning approach that can be deployed on low-memory devices for speech emotion classification. A redesigned VGGNet with an overall model size of 7.94MB is utilized, combined with the best-performing classifier (Random Forest). Extensive experiments and comparisons with other deep learning models (DenseNet, MobileNet, InceptionNet, and ResNet) over three publicly available speech emotion datasets show that the proposed lightweight model improves the performance of emotion classification with minimal parameter size. The research further devises a new method that minimizes computational complexity using a vision transformer (ViT) network for speech emotion classification. The ViT model’s capabilities allow the mel-spectrogram input to be fed into the model, allowing for the capturing of spatial dependencies and high-level features from speech signals that are suitable indicators of emotional states. Finally, the research proposes a novel transformer model that is based on shift-window for efficient classification of speech emotion on bi-lingual datasets. Because this method promotes feature reuse, it needs fewer parameters and works well with smaller datasets. The proposed model was evaluated using over 3000 speech emotion samples from the publicly available TESS, EMODB, EMOVO, and bilingual TESS-EMOVO datasets. The results showed 98.0%, 98.7%, and 97.0% accuracy, F1-Score, and precision, respectively, across the 7 classes of emotion.
Diffuse radio emission in ACTPol clusters.
(2021) Sikhosana, Sinenhlanhla Precious.; Moodley, Kavilan.; Knowles, Kenda Leigh.; Hilton, Matthew James.
Low-frequency radio observations of galaxy clusters reveal cluster-scale diffuse emission that is not associated with individual galaxies. Studying the properties of these diffuse radio sources gives insight into astrophysical processes such as cosmic ray transportation in the intracluster medium (ICM). Observations have linked the formation of radio halos and relics with turbulence caused by cluster mergers and the formation of mini-halos to gas sloshing in cool-core clusters. Statistical studies of large galaxy cluster samples have been used to determine how the radio properties of diffuse emission scale with the mass and X-ray luminosity of the host clusters. Such studies are crucial for refining the formation theories of diffuse emission. New generation telescopes with wide bandwidths and high sensitivity such as the upgraded Giant Metrewave Radio Telescope (uGMRT) andMeerKAT are advantageous for the study of faint extended emission in large cluster samples. The main aim of this thesis was to do an in-depth study of the diffuse radio emission using a cluster sample that spans a wider mass and redshift range compared to the currently studied parameter space. We developed data reduction techniques for calibrating data from telescopes such as uGMRT and MeerKAT. The wide bandwidth of these telescopes introduces directional dependent effects (DDEs) that make the calibration process extremely complicated. However, such observations are excellent for studies of the faint diffuse emission and in-band spectral indices of this emission. In the first part of this thesis, we focused on the study of diffuse radio emission in a Sunyaev- Zeldovich (SZ) selected sample of clusters. These clusters were observed by the Atacama Cosmology Telescope’s Polarimetric extension (ACTPol). We used archival and new GMRT observations for the radio analysis of this sample. We reported newly detected diffuse emission in the following clusters: a radio halo and revived fossil plasma in ACT-CL J0137.4 0827, a radio relic in ACT-CL J2128.4+0135, and a candidate relic in ACT-CL J0022.2 0036. The radio analysis of the full sample revealed that the fraction of clusters in the sample hosting diffuse emission is 26.7% excluding candidate emission and 30% when it is included. The detection rate of the diffuse emission over all categories is lower than the detection rates reported in literature. We note that this may be because the sample comprised high redshift (z ¡ 0.5) and low mass clusters (M500c;SZ 5 1014 Md), though future more sensitive observations of these clusters could reveal fainter diffuse emission structures. We compared our results to the most recent radio halo and radio relic scaling relations. The radio halo P1:4GHz M500 scaling relation plot indicates that a few flatter spectrum radio halos are located in the region previously known to be populated by ultrasteep spectrum radio halos (USSRHs). Finally, we presented preliminary results of the uGMRT wideband backend (GWB) data reduction for ACT-CL J0034.4+0225, ACT-CL J0137.4 0827, and ACT-CL J2128.4+0135. We prioritised these clusters because the narrowband data revealed that they host diffuse emission. However, once the data reduction algorithm is improved, we will reduce the remaining clusters with non-detections. Comparing the GWB results to the narrowband GMRT data, we note that the radio halo observed in ACT-CL J0137.4 0827 is more extended in the GWB data. The diffuse emission is detected at a higher signal-to-noise ratio in the GWB images for the three clusters. We note that an improvement in the GWB reduction algorithm might reveal diffuse emission that was not detected in the narrowband data. In the second part of the thesis, we used MeerKAT observations to study diffuse emission in the Bullet Cluster (1E0657 56), RXCJ1314.4 2515, Abell 3562, and Abell 3558. We detected new extended features in the radio halos hosted by the Bullet cluster and Abell 3562. We assume that the decrement feature in the Bullet cluster might be an indication of a second wave of merger activity. The ridge feature in the peripheral region of the radio halo in Abell 3562 overlaps with the edge of the X-ray emission. Hence, we assume that the feature might be related to a shock region. We also reported the detection of a new mini-halo in Abell 3558. MeerKAT’s sensitivity and wide bandwidth enabled us to perform in-band spectral index studies and produce spectral index maps for the Bullet cluster, RXCJ1314.4 2515, and Abell 3562. The spectral index maps of the relics in the Bullet cluster and RXCJ1314.4 2515 indicate a spectral steepening towards the cluster center, while the spectral index map of the radio halo in the Bullet cluster indicates radial spectral steepening. The spectral index map of Abell 3562 indicates that the radio halo and ridge have similar spectral index variations, which suggests that the ridge feature is related to the radio halo.
Discrete time-to-event construction for multiple recurrent state transitions.
(2023) Batidzirai, Jesca Mercy.; Manda, Samuel.; Mwambi, Henry Godwell.
Recent developments in multi-state models have considered discrete time rather than continuous time in the modeling of transition intensities, whose major drawback lies in the possibility of resulting in biased parameter estimates that arise from issues of handling ties. Discrete-time models have included univariate multilevel models to account for possible dependence among specific pairwise recurrent transitions within the same subject. However, in most cases, there would be several specific pairwise transitions of interest. In such cases, there is a need to model the transitions with the aim of identifying those transitions that are correlated. This provides insight into how the transitions are related to each other. In order to investigate the interdependencies between transitions, the unique contribution of this thesis is to propose a multivariate discrete-time multi-state model with multiple state transitions. In this model, each specific recurrent transition is associated with a random effect to capture possible dependence in the transitions of the same type or different types. The random effects themselves were then modeled by a multivariate normal distribution and model parameters were estimated using maximum likelihood methods with Gaussian quadratures numerical integration. A simulation study was done to evaluate the performance of the proposed model. The model yielded satisfactory results for most fixed effects and random effects estimates. This is noticed by near-zero biases and mean square errors of the average estimates as well as high 95% coverage probabilities of the 95% confidence intervals from 1000 replications. The proposed methodology was applied to marriage formation and dissolution data from KwaZulu-Natal province, South Africa. Five transitions were considered, namely: Never Married to Married, Married to Separated, Married to Widowed, Separated to Married and Widowed to Married. The presence of very small unobserved subject-to subject heterogeneity for each transition and a weak positive correlation between transitions were produced. Statistically, the model produced smaller standard errors compared to those from univariate models, hence it is more precise on estimates. The multivariate modeling of discrete time-to-event models provides a better understanding of the evolution of all transitions simultaneously, thus in addition to covariate effects, giving an assessment of how one transition is associated with the other. Empirical results confirmed well known important socio-demographic predictors of entering and exiting a marriage. Age at sexual debut played a positive critical role in most of the transitions. More educated subjects were associated with a lower likelihood of entering a first marriage, experiencing a marital dissolution as well as remarrying after widowhood. Subjects who had a sexual debut at younger ages were more likely to experience a marital dissolution than those who started late. Age at first marriage had a negative association with marital dissolution. We may, therefore, postulate that existing programs that encourage delay in onset of sexual activity for HIV risk reduction for example, may also have a positive impact on lowering rates of marital dissolution, thus ultimately improving psychological and physical health.
Exploration of ear biometrics with deep learning.
(2024) Booysens, Aimee Anne.; Viriri, Serestina.
Biometrics is the recognition of a human using biometric characteristics for identification, which may be physiological or behavioural. Numerous models have been proposed to distinguish biometric traits used in multiple applications, such as forensic investigations and security systems. With the COVID-19 pandemic, facial recognition systems failed due to users wearing masks; however, human ear recognition proved more suitable as it is visible. This thesis explores efficient deep learning-based models for accurate ear biometrics recognition. The ears were extracted and identified from 2D profiles and facial images, focusing on both left and right ears. With the numerous datasets used, with particular mention of BEAR, EarVN1.0, IIT, ITWE and AWE databases. Many machine learning techniques were explored, such as Naïve Bayes, Decision Tree, K-Nearest Neighbor, and innovative deep learning techniques: Transformer Network Architecture, Lightweight Deep Learning with Model Compression and EfficientNet. The experimental results showed that the Transformer Network achieved a high accuracy of 92.60% and 92.56% with epochs of 50 and 90, respectively. The proposed ReducedFireNet Model reduces the input size and increases computation time, but it detects more robust ear features. The EfficientNet variant B8 achieved a classification accuracy of 98.45%. The results achieved are more significant than those of other works, with the highest achieved being 98.00%. The overall results showed that deep learning models can improve ear biometrics recognition when both ears are computed.
Financial modelling of cryptocurrency: a case study of Bitcoin, Ethereum, and Dogecoin in comparison with JSE stock returns.
(2022) Kaseke, Forbes.; Ramroop, Shaun.; Mwambi, Henry Godwell.
The emergency of cryptocurrency has caused a shift in the financial markets. Although it was created as a currency for exchange, cryptocurrency has been shown to be an asset, with investors seeking to profit from it rather than using it as a medium of exchange. Despite being a financial asset, cryptocurrency has distinct, stylised facts like any other asset. Studying these stylised facts allows the creation of better-suited models to assist investors in making better data-driven decisions. The data used in this thesis was of three leading cryptocurrencies: Bitcoin, Ethereum, and Dogecoin and the Johannesburg Stock Exchange (JSE) data as a guide for comparison. The sample period was from 18 September 2017 to 27 May 2021. The goal was to research the stylised facts of cryptocurrencies and then create models that capture these stylised facts. The study developed risk-quantifying models for cryptocurrencies. The main findings were that cryptocurrency exhibits stylised facts that are well-known in financial data. However, the magnitude and frequency of these stylised facts tend to differ. For example, cryptocurrency is more volatile than stock returns. The volatility also tends to be more persistent than in stocks. The study also finds that cryptocurrency has a reverse leverage effect as opposed to the normal one, where past negative returns increase volatility more than past positive returns. The study also developed a hybrid GARCH model using the extreme value theorem for quantifying cryptocurrency risk. The results showed that the GJR-GARCH with GDP innovations could be used as an alternative model to calculate the VaR. The volatile nature of cryptocurrency was also compared with that of the JSE while accounting for structural breaks and while not accounting for them. The results showed that the cryptocurrencies’ volatility patterns are similar but differ from those of the JSE. The cryptocurrency was also found to be an inefficient market. This finding means that some investors can take advantage of this inefficiency. The study also revealed that structural breaks affect volatility persistence. However, this persistence measure differs depending on the model used. Markov switching GARCH models were used to strengthen the structural break findings. The results showed that two-regime models outperform single-regime models. The VAR and DCC-GARCH models were also used to test the spillovers amongst the assets used. The results showed short-run spillovers from Bitcoin to Ethereum and long-run spillovers based on the DCC-GARCH. Lastly, factors affecting cryptocurrency adoption were discussed. The main reasons affecting mass adoption are the complexity that comes with the use of cryptocurrency and its high volatility. This study was critical as it gives investors an understanding of the nature and behaviour of cryptocurrency so that they know when and how to invest. It also helps policymakers and financial institutions decide how to treat or use cryptocurrency within the economy.
Flexible Bayesian hierarchical spatial modeling in disease mapping.
(2022) Ayalew, Kassahun Abere.; Manda, Samuel.
The Gaussian Intrinsic Conditional Autoregressive (ICAR) spatial model, which usually has two components, namely an ICAR for spatial smoothing and standard random effects for non-spatial heterogeneity, is used to estimate spatial distributions of disease risks. The normality assumption in this model may not always be correct and misspecification of the distribution of random effects could result in biased estimation of the spatial distribution of disease risk, which could lead to misleading conclusions and policy recommendations. Limited research studies have been done where the estimation of the spatial distributions of diseases under the ICAR-normal model were compared to those obtained from fitting ICAR-nonnormal model. The results from these studies indicated that the ICAR-nonnormal models performed better than the ICAR-normal in terms of accuracy, efficiency and predictive capacity. However, these efforts have not fully addressed the effect on the estimation of spatial distributions under flexible specification of ICAR models in disease mapping. The overall aim of this PhD thesis was to develop approaches that relax the normality assumption that is often used in modeling and fitting of ICAR models in the estimation of spatial patterns of diseases. In particular, the thesis considered the skewnormal and skew-Laplace distributions under the univariate, and skew-normal for the multivariate specifications to estimate the spatial distributions of either univariable or multivariable areal data. The thesis also considered non-parametric specification of the multivariate spatial effects in the ICAR model, which is a novel extension of an earlier work. The estimation of the models was done using Bayesian statistical approaches. The performances of our suggested alternatives to the ICAR-normal model were evaluated by simulating studies as well as with practical application to the estimation of district-level distribution of HIV prevalence and treatment coverage using health survey data in South Africa. Results from the simulation studies and analysis of real data demonstrated that our approaches performed better in the prediction of spatial distributions for univariable and multivariable areal data in disease mapping approaches. This PhD has shown the limitations of relying on the ICAR-normal model for the estimations of spatial distributions for all spatial analyses, even when the data could be asymmetric and non-normal. In such scenarios, skewed-ICAR and nonparametric ICAR approaches could provide better and unbiased estimation of the spatial pattern of diseases.
Forest image classification based on deep learning and ontologies.
(2024) Kwenda, Clopas.; Gwetu, Mandlenkosi Victor.; Fonou-Dombeu, Jean Vincent.
Forests contribute abundantly to nature’s natural resources and they significantly contribute to a wide range of environmental, socio-cultural, and economic benefits. Classifications of forest vegetation offer a practical method for categorising information about patterns of forest vegetation. This information is required to successfully plan for land use, map landscapes, and preserve natural habitats. Remote sensing technology has provided high spatio-temporal resolution images with many spectral bands that make conducting research in forestry easy. In that regard, artificial intelligence technologies assess forest damage. The field of remote sensing research is constantly adapting to leverage newly developed computational algorithms and increased computing power. Both the theory and the practice of remote sensing have significantly changed as a result of recent technological advancements, such as the creation of new sensors and improvements in data accessibility. Data-driven methods, including supervised classifiers (such as Random Forests) and deep learning classifiers, are gaining much importance in processing big earth observation data due to their accuracy in creating observable images. Though deep learning models produce satisfactory results, researchers find it difficult to understand how they make predictions because they are regarded as black-box in nature, owing to their complicated network structures. However, when inductive inference from data learning is taken into consideration, data-driven methods are less efficient in working with symbolic information. In data-driven techniques, the specialized knowledge that environmental scientists use to evaluate images obtained through remote sensing is typically disregarded. This limitation presents a significant obstacle for end users of Earth Observation applications who are accustomed to working with symbolic information, such as ecologists, agronomists, and other related professionals. This study advocates for the incorporation of ontologies in forest image classification owing to their ability in representing domain expert knowledge. The future of remote sensing science should be supported by knowledge representation techniques such as ontologies. The study presents a methodological framework that integrates deep learning techniques and ontologies with the aim of enhancing domain expert confidence as well as increasing the accuracy of forest image classification. In addressing this challenge, this study followed the following systematic steps (i) A critical review of existing methods for forest image classification (ii) A critical analysis of appropriate methods for forest image classification (iii) Development of the state-of-the-art model for forest image segmentation (iv) Design of a hybrid model of deep learning and machine learning model for forest image classification (v) A state-of-the-art ontological framework for forest image classification. The ontological framework was flexible to capture the expression of the domain expert knowledge. The ontological state-of-the-art model performed well as it achieved a classification accuracy of 96%, with a Root Mean Square Error of 0.532. The model can also be used in the fruit industry and supermarkets to classify fruits into their respective categories. It can also be potentially used to classify trees with respect to their species. As a way of enhancing confidence in deep learning models by domain experts, the study recommended the adoption of explainable artificial intelligence (XAI) methods because they unpack the process by which deep learning models reach their decision. The study also recommended the adoption of high-resolution networks (HRNets) as an alternative to traditional deep learning models, because they can convert low-resolution representation to high-resolution and have efficient block structures developed according to new standards and they are excellent at being used for feature extraction.
Hybrid genetic optimisation for quantum feature map design.
(2024) Pellow-Jarman, Rowan Martin.; Pillay, Anban Woolaganathan.; Ilya, Sinayskiy.; Petruccione, Francesco.
Good feature maps are crucial for machine learning kernel methods for effective mapping of non-linearly separable input data into a higher dimension feature space, thus allowing the data to be linearly separable in feature space. Recent works have proposed automating the task of quantum feature map circuit design with methods such as variational ansatz parameter optimization and genetic algorithms. A problem commonly faced by genetic algorithm methods is the high cost of computing the genetic cost function. To mitigate this, this work investigates the suitability of two metrics as alternatives to test set classification accuracy. Accuracy has been applied successfully as a genetic algorithm cost function for quantum feature map design in previous work. The first metric is kernel-target alignment, which has previously been used as a training metric in quantum feature map design by variational ansatz training. Kernel-target alignment is a faster metric to evaluate than test set accuracy and does not require any data points to be reserved from the training set for its evaluation. The second metric is an estimation of kernel-target alignment which further accelerates the genetic fitness evaluation by an adjustable constant factor. The second aim of this work is to address the issue of the limited gate parameter choice available to the genetic algorithm. This is done by training the parameters of the quantum feature map circuits output in the final generation of the genetic algorithm using COBYLA to improve either kernel-target alignment or root mean squared error. This hybrid approach is intended to complement the genetic algorithm structure optimization approach by improving the feature maps without increasing their size. Eight new approaches are compared to the accuracy optimization approach across nine varied binary classification problems from the UCI machine learning repository, demonstrating that kernel-target alignment and its approximation produce feature map circuits enabling comparable accuracy to the original approach, with larger margins on training data that improve further with variational training.
Mathematical modelling of the Ebola virus disease.
(2024) Abdalla, Suliman Jamiel Mohamed.; Govinder, Keshlan Sathasiva.; Chirove, Faraimunashe.
Despite the numerous modelling efforts to advise public health physicians to understand the dynamics of the Ebola virus disease (EVD) and control its spread, the disease continued to spread in Africa. In the current thesis, we systematically review previous EVD models. Further, we develop novel mathematical models to explore two important problems during the 2018-2020 Kivu outbreak: the impact of geographically targeted vaccinations (GTVs) and the interplay between the attacks on Ebola treatment centres (ETCs) and the spread of EVD. In our systematic review, we identify many limitations in the modelling literature and provide brief suggestions for future work. Our modelling findings underscore the importance of considering GTVs in areas with high infections. In particular, we find that implementing GTVs in regions with high infections so that the total vaccinations are increased by 60% decreases the cumulative cases by 15%. On the other hand, we need to increase the vaccinations to more than 1000% to achieve the 15% decrease in EVD cases if we implement GTVs in areas with low infections. On the impact of the attacks on ETCs, we find that due to the attacks on ETCs, the cumulative cases increased by more than 17% during the 2018-2020 Kivu outbreak. We also find that when 10% of the hospitalised individuals flee the attacks on ETCs after spending only three days under treatment, the cumulative cases increased by more than 30% even if these individuals all returned to the ETCs three days later. On the other hand, if only half of these individuals returned to ETCs for treatment, the cumulative cases increase by approximately 50%. Further, when these patients spend one more day in the community, after which they all return to ETCs, the cumulative cases rise by an additional 10%. Global sensitivity analysis also confirmed these findings. To conclude, our literature systematic review is used to identify many critical factors which were overlooked in previous EVD models. Our modelling findings show that the attacks on ETCs can be destructive to the efforts of EVD response teams. Hence, it is important for decision-makers to tackle the reasons for community distrust and address the roots of the hostility towards ETCs. We also find that GTVs can be used to contain the spread of EVD when ring vaccinations, contact tracing and antiviral treatments cannot successfully control the spread of EVD.
Pancreatic cancer survival prediction using Deep learning techniques.
(2023) Bakasa, Wilson.; Viriri, Serestina.
Abstract available in PDF.
Road obstacle detection Using YOLO algorithm based on attention mechanism.
(2024) Lekola , Bafokeng.; Viriri, Serestina.
Road obstacle detection is an important task in autonomous vehicles (AVs) and advanced driver assistance systems (ADAS) as they require real-time operation and high accuracy for safe operation. The mobile nature of the task means that it is carried out in a low-resourced environment where there is a need for an algorithm that achieves both high accuracy and high inference speed while meeting the requirement for lightweight. In this dissertation, an exploration of the effectiveness of the Attention-enhanced YOLO algorithm for the task of road obstacle detection is carried out. Several state-of-the-art attention modules that employ both channel and spatial attention are explored and fused with the YOLOv8 and YOLOv9 algorithms. These enhance feature maps of the network by suppressing non-distinctive features allowing the network to learn from highly distinctive features. The Attention-modified networks are trained and validated on the Kitti and BDD100k datasets which are publicly available. Comparisons are made between the models and the baseline. An improvement from the baseline is seen with the GAM attention achieving an accuracy rate of 93.3% on the Kitti dataset and 71.1% on the BDD100k dataset. The Attention modules generally achieved incremental improvements over the baseline.
Solar flare recurrence prediction & visual recognition.
(2024) Mngomezulu, Mangaliso Moses.; Gwetu, Mandlenkosi Victor.; Fonou-Dombeu, Jean Vincent.
Solar flares are intense outbursts of radiation observable in the photosphere. The radiation flux is measured in W/m2. Solar flares can kill astronauts, disrupt electrical power grids, and interrupt satellite-dependent technologies. They threaten human survival and the efficiency of technology. The reliability of solar flare prediction models is often undermined by the stochastic nature of solar flare occurrence as shown in previous studies. The Geostationary Operational Environmental Satellite (GOES) system classifies solar flares based on their radiation flux. This study investigated how Recurrent Neural Network (RNN) models compare to their ensembles when predicting flares that emit at least 10−6W/m2 of radiation flux, known as ≥C class flares. A Long-Short Term Memory (LSTM) and Simple RNN homogeneous ensemble achieved a similar performance with a tied True Skill Statistic (TSS) score of 70 ± 1.5%. Calibration curves showed that ensembles are more reliable. The balanced accuracies of the Simple RNN Ensemble and LSTM are both 85% with f1-scores of 79% and 77% respectively. Furthermore, this study proposed a framework that shows how objective function reparameterization can be used to improve binary (≥C or
Stable distributions with applications to South African financial data.
(2024) Naradh , Kimera.; Chinhamu, Knowledge.; Chifurira, Retius.
In recent times, researchers, analysts and statisticians have shown a keen interest in studying Extreme Value Theory (EVT), particularly with the application to mixture models in the medical and financial sectors. This study aims to validate the use of stable distributions in modelling three Johannesburg Stock Exchange (JSE) market indices, namely the All Share Index (ALSI), Banks Index and the Mining Index, as well as the United States of American Dollar (USD) to South African Rand (ZAR) exchange rate. This study leverages the unique properties of stable distributions when modelling heavy-tailed data. Nolan’s S0-parameterization stable distribution (SD) was fitted to the returns of the three FTSE/JSE indices and USD/ZAR exchange rate and a hybrid Generalized Autoregressive Conditional Heteroskedasticity (GARCH)-type model combined with stable distributions was fitted to each return series. The two-tailed mixture model of the Generalized Pareto Distribution (GPD), stable distribution, Generalized Pareto Distribution referred to as GSG, as well as the Stable-Normal-Stable (SNS) and Stable-KDE-Stable (SKS) was fitted to evaluate its relative performance in modelling financial data. Results show that the S0-parameterization SD fits the South African financial returns well. The hybrid GARCH (1,1)-SD model competes favourably with the GARCH-GPD model in estimating Value-at-Risk (VaR) for FTSE/JSE Banks Index, FTSE/JSE Mining Index and the USD/ZAR exchange rate returns. The hybrid EGARCH (1,1)-SD competes well against the GARCH-GPD model for the FTSE/JSE ALSI returns. Inconclusive results are observed for the short position of the fitted GKG models; however, in the long position, an appropriate fit of the GPD-KDE-GPD (GKG) model, where KDE is the kernel density estimator, is emphasised for all four return series. The proposed mixture models, GSG, SNS and SKS models, are found to be a good alternative in fitting South African financial data to the commonly used GPD-Normal-GPD (GNG) mixture model. The results of this study are important to financial practitioners, risk managers and researchers as the proposed mixture models add more value to the literature on the applications of extreme mixture models.
Statistical and machine learning methods of online behaviours analysis.
(2024) Soobramoney, Judah.; Chifurira, Retius.; Zewotir, Temesgen Tenaw.
The success of corporates is highly influenced by the effectiveness and appeal of each corporate’s website. This study was conducted on TEKmation, a South African corporate, whose board of directors lacked insight regarding the website’s usage. The study aimed to quantify the web-traffic flow, detect the underlying browsing patterns, and validate the web-design effectiveness. The website experienced 7,935 visits and 57,154 page views from 1 June 2021 to 30 June 2023 (data sourced by Google Analytics). Grubb’s test has identified outliers in visit frequency, the pageviews per visit, and the visit duration per visit. A small degree of missingness was observed on the mobile device branding (1.24%) and operating system (0.03%) features which were imputed using a Bayesian network model. To address a data-shift detected, an artificial neural network (ANN) was proposed to flag future data-shifts with important predictors being the period of year and volume of sessions. Prior to clustering, feature selection methods assessed the feature variability and feature association. Results indicated that low-incidence webpages and features with natural relationships should be omitted. The K-means, DBScan and hierarchical unsupervised machine learning methods were employed to identify the visit personas, labelled get-in-touch (12%), accidentals (11%), dropoffs (30%), engrossed (38%) and seekers (9%). It was evident that the premature drop-offs needed further exploration. The Cox proportional hazards survival model and the random survival forest (RSF) model have identified that the web browser, visit frequency, device category, distance, certain webpages, volume of hits, and organic searches proved to be drop-offs hazards. A tiered Markov chain model was developed to compute the transition probabilities of dropping-off. The contact (63%) and clients (50%) states recorded a high likelihood to drop-off early within the visit. In conclusion, using statistical methods, the study informed the board on of its audience, the flaws of the website and proposed recommendations to address concerns.
Statistical study on childhood malnutrition and anaemia in Angola, Malawi and Senegal.
(2023) Khulu, Mthobisi Christian.; Ramroop, Shaun.; Habyarimana, Faustin.
Malnutrition and anaemia continue to be a concern to the future of developing countries. This thesis aimed to examine the risk factors associated with malnutrition and anaemia among under five-year-old children in Angola, Malawi and Senegal. Statistical models and techniques have improved over the years to give more insight into malnutrition and anaemia, in terms of demographic, socio-economic, environmental, and geographic factors. This thesis also assessed the spatial epidemiological overlaps between childhood malnutrition and anaemia diseases which can lead to various advantages in intervention planning, monitoring, controlling and total elimination of such diseases, especially in high-risk regions. This is a secondary data analysis where national representative data from the three countries was used. The Demographic and Health Survey data from Angola, Malawi and Senegal were merged to create a pooled sample which was then used for all the analyses conducted in this study. The relationship between exploratory variables to malnutrition and anaemia was assessed to obtain variables that explain the two outcomes. Consequently, a generalized linear mixed model was used to investigate the significance of the child-level, community-level and household-level factors to malnutrition and anaemia separately. The relationship between the two diseases was further examined using the three joint modelling approaches: (1) a joint generalised linear mixed model; (2) a structural equation model, and (3) a bivariate copula geo-additive model. For each model employed, the significant factors of both malnutrition and anaemia were identified. The GLMM results on malnutrition revealed that children’s place of residence, age, gender, mother’s level of schooling, wealth status, birth interval and birth order significantly explain malnutrition at the 5% level of significance. Whereas, the GLMM results on anaemia revealed that children‘s age, gender, mother’s level of schooling, wealth status and nutritional status significantly explain anaemia at 5% level of significance. The findings of copula geo-additive modelling of malnutrition and anaemia indicated that there is an association between malnutrition and anaemia. There was a strong association observed between malnutrition and anaemia in the north-west districts of Angola when compared to other districts. The results imply that the policymakers of Angola, Senegal and Malawi can control anaemia through the intervention of malnutrition controlling. The overall findings of this study provide meaningful insight to the policymakers of Angola, Malawi and Senegal which will lead to the implementation of interventions that can assist in achieving the Sustainable Development Goal (SDG) of 25 deaths per 1 000 live births by 2030. To properly eradicate all the causes of malnutrition and anaemia, programs such as parental education, financial education, children's dietary focus programs and mobile health facilities could add a significant value. The results also highlighted the national priority areas related to child-related factors, household factors and environmental factors for childhood malnutrition and anaemia morbidity control. It also provided policy makers with valuable geographical information for developing and implementing effective intervention. There is a greater need for partnership and collaboration among the studied countries to achieve the SGD target.

Browse

Browsing School Mathematics, Statistics and Computer Science by SDG "SDG4"