Browse
Recent Submissions
Item Evaluating model compression techniques for efficient edge AI deployment(2025-11) Zungu, ML; Zwane, Skhumbuzo G; Adigun, MODeploying artificial intelligence models on edge devices presents challenges due to limited computational power, memory, and energy. Model compression is essential to reduce the size and computational requirements of AI models, enabling their deployment on smartphones, IoT sensors, and other edge devices. This paper evaluates three major compression techniques which are quantization, pruning, and tensor decomposition, applied to the YOLOv8n model for object detection. The study compares their impact on accuracy, model size, inference speed, and energy efficiency. Findings indicate that quantization achieves the best balance by reducing size and improving inference speed with minimal accuracy loss, pruning yields high energy savings but at the cost of accuracy, and tensor decomposition provides a balanced trade-off. The research underscores the importance of compression in enabling real-time Edge AI applications and highlights the quantization model compression technique as the most suitable technique for efficient edge deployment.Item Environmental impact analysis of traditional concrete road and ash-based concrete road construction(2025-10) Mashiyane, T; Mvelase, Gculisile MThis study compares the environmental performance of two pavement types: conventional concrete using cement, gravel, and crushed stone, and an ash-based mix incorporating legacy ash and ground granulated blast furnace slag (GGBFS). The ash-based option supports circular economy principles by repurposing waste and reducing dependence on virgin raw materials. Using life cycle assessment (LCA) in SimaPro, greenhouse gas emissions were evaluated across the pavements’ life cycles. Results show that conventional concrete generates about 66% more CO₂ emissions, mainly from cement production and aggregate processing. In contrast, the ash-based road achieved a 66% reduction in embodied carbon, highlighting its potential as a sustainable alternative. These findings demonstrate the value of circular economy strategies in road construction and confirm the role of secondary materials in lowering the sector’s carbon footprint.Item Benchmarking machine learning models for cloud resource forecasting: LSTM vs. SVR, Random Forest, XGBoost, and Prophet(2025-12) Mamushiane, Lusani; Hlophe, C; Kobo, Hlabishi I; Lysko, Albert A; Mwangama, JEffective cloud resource forecasting is essential for optimizing virtual machine (VM) provisioning in increasingly dynamic and resource-constrained environments, especially now as next-generation mobile systems (5G/6G) continue to emerge. This study examines a selection of learning-based forecasting approaches, including LSTM, SVM, Random Forest, XGBoost, and Prophet, for VM workload prediction using real-world data. The analysis emphasizes the balance that must be maintained between predictive performance and computational demand, two critical aspects for real-time and scalable deployment. Results show that while LSTM achieves superior accuracy, it incurs significant training overhead. Random Forest provides a more balanced trade-off, maintaining competitive accuracy with lower computational cost. These findings inform model selection in contexts like 5G/6G base station association, network slicing orchestration, multi-access edge computing, intelligent handover decisions, and dynamic service function chaining, where both efficiency and precision are vital for maintaining service-level guarantees under constrained resources.Item A framework for resource overprovisioning with machine learning to maximise revenue from 5G Core network slices(2025-12) Mamushiane, L; Mwangama, J; Lysko, Albert A; Kobo, Hlabishi INetwork slicing has emerged as a key enabler for delivering diverse services in 5G and beyond networks, where each slice must be provisioned with sufficient resources while maintaining operator profitability. Traditional admission control strategies are often conservative, leading to underutilization, while aggressive allocation risks violating service-level agreements (SLAs). This paper proposes a machine learning-based forecasting and admission control framework that leverages overprovisioning to balance utilization, revenue, and reliability. The framework combines CNN-LSTM forecasting for predicting slice resource demand with a policy that admits slice requests based on forecasted aggregate utilization scaled by an overprovisioning factor. To evaluate the approach, we use over 500 VM workload traces from the Materna dataset, grouping VMs into slices of ten. Results demonstrate that CNN-LSTM forecasting achieves consistent accuracy across diverse workloads, and that overprovisioning-based admission control improves net profit while reducing SLA violations compared to conservative baselines. These findings validate the feasibility of ML-driven overprovisioning for efficient and reliable slice admission in 5G networks.Item Benchmarking machine learning models for cloud resource forecasting: LSTM vs. SVR, Random Forest, XGBoost, and Prophet(2025-12) Mamushiane, Lusani; Hlophe, C; Kobo, Hlabishi I; Lysko, Albert A; Mwangama, JEffective cloud resource forecasting is essential for optimizing virtual machine (VM) provisioning in increasingly dynamic and resource-constrained environments, especially now as next-generation mobile systems (5G/6G) continue to emerge. This study examines a selection of learning-based forecasting approaches, including LSTM, SVM, Random Forest, XGBoost, and Prophet, for VM workload prediction using real-world data. The analysis emphasizes the balance that must be maintained between predictive performance and computational demand, two critical aspects for real-time and scalable deployment. Results show that while LSTM achieves superior accuracy, it incurs significant training overhead. Random Forest provides a more balanced trade-off, maintaining competitive accuracy with lower computational cost. These findings inform model selection in contexts like 5G/6G base station association, network slicing orchestration, multi-access edge computing, intelligent handover decisions, and dynamic service function chaining, where both efficiency and precision are vital for maintaining service-level guarantees under constrained resources.Item Cybersecurity in smart grids using machine learning: A systematic literature review(2025-12) Nelufule, Nthatheni; Mudau, Tshimangadzo C; Nkwe, Boitumelo CA; Chishiri, Samson; Mncwango, Lungisani S; Mutenwa, MuyowaThe Smart Grids due to their reliance to the connectedness of Internet of Things technologies, advanced metering infrastructure, and real-time communication are faced with significant cybersecurity and privacy risks. In this work a synthesized insights from recent literatures are presented which critically evaluate Machine Learning, Deep Learning, and privacy-preserving techniques for securing smart grids against cyber-attacks. The critical analysis of this review spans across several themes and topics such as ensemble ML models, federated learning, graph neural networks and blockchain-based protocols, scalability, and privacy-aware communication frameworks. However, there are several limitations such as reliance on simulated datasets, computational complexity, adversarial vulnerabilities, scalability issues, and lack of real-time data validation which limits their practical deployment in a real-world environment. This survey has also proposed some research solutions such as standardized real-world datasets, lightweight algorithms, adversarial training, explainable AI, and interoperable blockchain frameworks to enhance robustness, scalability, and trust. This study addresses key research questions on the effectiveness of learning based techniques, and the role of emerging technologies such as 5G technologies and explainable AI in securing a smart grids environment.Item Building a low-cost cloud Native 5G network slicing experimental testbed: Open-source solutions, lessons learned and future directions(2025-11) Otieno, HO; Mamushiane, Lusani; Maurine, C; Mukute, T; Mwangama, J; Malila, B; Modroiu, ER; Magedanz, TResearch on network slicing has gained some ground since its introduction to 5G networking. Network slicing has evolved from ideas, concepts, and simulations to real-world implementations. With this advancement, the realization and experimentation of network slicing remain a challenge due to the complexity involved. Research and experimentation testbeds have been built to explore and advance network slicing. However, most are quite expensive to set up or complicated for the novice researcher exploring 5G network slicing. In this work, we present a comprehensive and repeatable methodology that can be used to realize network slicing. Our approach focuses on building a low-cost, experimental cloud-native testbed using open-source solutions. The testbed leverages containerization and orchestration platforms to deploy the core network. We demonstrate end-to-end network slicing by extending the Open Air Interface’s reference setup with custom Helm charts. We validate the process of the user equipment requesting a slice on the network, to admission and management of the slice using slice-specific identifiers like Slice Service Type and Slice Differentiator. We highlight the lessons learned when automating this testbed setup to ensure the testbed is repeatable, consistent and proper resource isolation is achieved per slice as envisioned in network slicing. This work serves as a practical guide for researchers to experiment with and explore network slicing using readily available off-the-shelf hardware.Item Radar backscatter extraction from high resolution C-band airborne SAR measurements(2025-10) Tema, Thabo H; De Witt, Josias J; Blaauw, Ciara; Mosito, Katlego E; Nel, Willem AJ; Gaffar, YAIn this paper backscatter extraction and analysis from high-resolution, C-band, airborne Synthetic Aperture Radar (SAR) data acquired by the DSI CSIR SAR system are presented. First key aspects related to the radiometric calibration of SAR systems are described and related to the strategy used to calibrate the DSI CSIR SAR system. Backscatter statistics measured at slant- and cross-range resolutions of 0.5 m and 1 m, respectively are then presented for farmland and informal urban clutter types. The latter represents an important clutter type for urban sprawl monitoring for which little high-resolution C-band clutter measurements have been presented in literature. The measured mean backscatter results are shown to align well with published data, verifying the system's radiometric calibration. Amplitude distribution model fitting indicates that Rayleigh and Log-Normal distributions show good fits for farmland and informal urban clutter, respectively. Extracted amplitude model parameters for both types of clutter are presented.Item Proceedings of the 2025 29th International Conference Information Visualisation(2025-08) Banissi, E; Datia, N; Moura Pires, J; Ursyn, A; Nazemi, K; Secco, CA; Kovalerchuk, B; Andonie, R; Gavrilova, M; Mabakane, Mabule SIn the current information era, all aspects of society—from scientific research and economic planning to cultural practices and daily decision-making—are increasingly driven by data, information, and knowledge systems. The quality of our information infrastructure hinges not only on access to data but on the integrity, structure, and interpretability of that data, including its origin, completeness, classification, and transformation into higher-level knowledge. As such, the fields of Information Visualization, Analytics, and Machine/Deep Learning—underpinned by Artificial Intelligence (AI) advances—have become indispensable to how we interrogate, represent, and utilize data at scale. AI-enabled visualization represents a critical convergence of computational intelligence and human cognitive insight. These systems can discover latent structures, reveal non-obvious correlations, and enable intuitive interaction with complex, high-dimensional data. They now form the cornerstone of automated and semi-automated knowledge discovery pipelines, with increasing influence on strategic decision-making in science, policy, and industry. This transformation is not merely technological but epistemological, reshaping how knowledge is created, validated, and communicated. As data becomes increasingly heterogeneous, dynamic, and uncertain, a key research challenge is modelling uncertainty and risk throughout the data-to-knowledge lifecycle. Visualization—especially with AI—is vital in rendering uncertainty intelligible, making patterns and anomalies more interpretable, and supporting real-time exploration in complex domains. The evolution of this field thus reflects a broader shift toward explainable, interactive, and human-centric AI systems. The 29th International Conference on Information Visualization (IV2025), coupled with the 6th International Conference on AI & Visualization, highlights these developments and presents a curated selection of peer-reviewed papers exploring the frontiers of AI-driven visual analytics, interactive systems, human-computer interaction, and domain-specific applications. This year’s contributions represent over 100 institutions from 20+ countries, collectively showcasing contemporary visualization research's vitality and interdisciplinary nature. A key feature of IV2025 is the introduction of the Researchers Link initiative—an interdisciplinary design workshop focused on aligning AI and visualization research with the United Nations 2030 Sustainable Development Goals (SDGs). Facilitated in collaboration with Darmstadt University of Applied Sciences and coordinated by an international cohort of senior researchers, this initiative invites early-career and established academics to co-develop fundable research proposals during the conference. This model of embedded research generation reflects a growing emphasis on translational impact and cross-sector collaboration. Collectively, the contributions to this volume illustrate the dynamic state of the field, where data visualization, augmented by AI, continues to redefine both technical and societal boundaries. As we advance into the era of intelligent systems, the challenge is not only to innovate but to ensure that these innovations remain interpretable, ethical, and inclusive.Item A comparison of various machine learning algorithms on ISAR image classification of complex targets with varying levels of Gaussian noise(2025-08) Stewart-Burger, CD; Ludick, DJ; Potgieter, MoniqueThis paper investigates the ability to classify complex targets using inverse synthetic aperture radar (ISAR) images with varying noise levels. ISAR is an imaging technique used to generate high resolution two-dimensional images of radar targets. ISAR images contain more information about targets than one-dimensional data, such as high- resolution range (HRR) profiles. ISAR images therefore have useful applications in target classification.Item Climate resilience in urban transport planning: A case study of Cosmo City’s wetland dilemma(2025-07) Kgoa, LeratoUrban transport planning in climate-sensitive regions requires balancing accessibility needs with environmental sustainability, particularly in informal settlements. This paper examines the dilemma surrounding an operational informal taxi rank located in an ecologically sensitive area near wetland in Cosmo City, Johannesburg. Ideally situated for community access, the rank serves as a vital transit hub for workers, scholars, and traders. However, its location raises concerns about climate resilience and ecological impact, as the area is prone to flooding and holds significant biodiversity value. While formalising the rank would enhance safety and infrastructure, relocation to a non-wetland area risks underutilisation, as seen with a nearby formal facility that remains unused due to its inconvenient location. Using available environmental, land use and transport data, this paper explores the trade offs between environmental risks and community benefits in maintaining the taxi rank’s current location. By weighing climate resilience considerations against transport accessibility needs, the study offers insights into adaptive strategies for transport infrastructure in vulnerable urban areas. This case presents a broader challenge for Southern African cities, emphasising the need for context-sensitive solutions that address both climate resilience and societal demands. The paper invites debate on sustainable urban transport planning in high-density, climate-sensitive regions.Item Operational readiness for e-governance in local government: A case study of the City of Mbombela with a focus on e-participation(2025-11) Thulare, Tumiso; Maremi, Keneilwe JThis paper assesses the operational readiness of the City of Mbombela for implementing e Governance, focusing specifically on e-Participation to improve service delivery. Using a qualitative case study design approach, the research involves document analysis, secondary data, and key informant interviews to evaluate four readiness areas: digital infrastructure, institutional capacity, legal frameworks, and stakeholder engagement. Findings reveal that while Mbombela demonstrates policy alignment and partial ICT adoption, challenges persist in bridging the digital divide, strengthening interdepartmental coordination, and promoting inclusive citizen participation. Comparative insights from global examples like Seoul, Nairobi, Reykjavik, Cape Town, and Estonia suggest adaptable strategies such as mobile-first platforms, community access points, participatory budgeting, and open data dashboards. Building on these lessons, the paper proposes a phased roadmap for Mbombela, including broadband expansion, pilot digital engagement programs, digital literacy initiatives, and multi-stakeholder partnerships. The study contributes to e-Governance scholarship by contextualising operational readiness within the realities of South African municipalities, offering both a conceptual framework and practical strategies for institutionalising inclusive e-Participation. The paper concludes that sustainable e-Governance in Mbombela requires bridging technical, social, and institutional gaps while embedding citizen engagement into governance structures.Item Characterisation of CSIR’s Open-Jet 2-meter wind tunnel for IEC 61400 MEASNET-compliant calibration of cup anemometers(2025-09) Ragimana, Phumudzo; Dikgale, Moyahabo S; Mabeko, Philimon KAccurate measurement of wind speed is essential for effective assessment of wind resources and evaluation of turbine performance within the wind energy industry. This paper details the characterization of the Council for Scientific and Industrial Research (CSIR) open-jet 2-meter wind tunnel, which is designed to facilitate MEASNET-compliant calibration of cup anemometers in line with IEC 61400. The test section underwent a rigorous evaluation to determine flow uniformity, turbulence intensity, and both axial and vertical flow alignment, utilizing a grid of calibrated pitot tubes. Measurements were taken across various horizontal and vertical planes to evaluate the spatial velocity distribution and confirm compliance with international calibration standards. The open-jet design presents distinct aerodynamic challenges, especially in maintaining low turbulence and uniform velocity profiles in the wind tunnel test section. Through a process of iterative flow conditioning and systematic testing, the tunnel achieved velocity deviations within ±0.03 m/s, thereby complying with the rigorous criteria for high-precision anemometer calibration. Repeatability tests conducted with a reference anemometer, along with assessments of environmental stability, further substantiated the wind tunnel's performance under different ambient conditions. The findings affirm the tunnel's capability as a national calibration facility, providing traceable and internationally recognised wind speed measurement capabilities. This characterisation establishes a basis for future improvements, including broader calibration ranges, automated testing rigs, and integrated uncertainty modelling to meet the increasing demands of the wind energy sector. This characterisation lays the groundwork for future upgrades, like wider calibration ranges, automated calibration systems, and better uncertainty modelling, to keep up with the growing needs of accurate measurements within the wind energy industry.Item R.A.I.S.E - A novel framework for evaluating foundational AI models in medical deployment: Moving beyond traditional metrics to real-world deployability(2025-12) Adendorff, JAE; Lourens, Roger L; Delport, R; Marivate, V; Gichoya, JWThe shift from “narrow” traditional deep learning models to more generalist foundation models represents a paradigm shift for AI in medicine with the emergence of unimodal and multimodal systems such as MedGemma, Biomedclip, DINO models, and MedImageInsight. While these generalist models promise broad capabilities, they demand large datasets and high computational resources for training, and carry risks such as hallucinations, which can be hazardous in clinical use. In medicine, whether a model can be securely incorporated into actual clinical workflows is more important than whether it passes standard ized tests. Current assessment techniques for foundation models are fre quently based on multiple choice questions and do not account for real world deployment scenarios. At a two-day datathon (16-17 July 2025), we explored deploying MedGemma for chest X-ray reporting in South Africa. We proposed a gradual, radiologist-guided integration focused on controlled, automatable tasks rather than full diagnostic use. Our three pronged evaluation framework creates a uniform readiness score and al lows for continuous real-world monitoring by combining tailored deploy ment paths and hierarchical decision making with Go/No-Go thresholds.Item Scoping the approaches of frugal innovation, micro frontends, and microservices to provide recommendations for the design and development of a South African electronic health record system(2025-08) Lourens, Roger L; Maremi, Keneilwe J; Van der Westhuizen, MelanieThis study provides a set of recommendations informed by the socio-economic context and existing literature to guide the design and development of an Electronic Health Record (EHR) system in South Africa. A comprehensive scoping review was conducted to identify critical recommendations essential for ensuring the effectiveness and efficiency of the EHR system’s design. The review utilised academic databases, including SpringerLink, IEEE Xplore, ScienceDirect, and the CSIRIS library. Additionally, supplementary searches were conducted using the Google search engine to identify relevant literature aligned with the study’s objectives. The findings indicate that adopting a system architecture based on micro frontends and microservices enhances development efficiency, reduces download times, and improves overall system performance. Moreover, the study emphasises the importance of frugal innovation, advocating for the strategic use of minimal resources during the design and development phases of the EHR system. Consequently, this paper recommends the integration of micro frontends and microservices as a fundamental approach to the design and development of an EHR system tailored to the South African context.Item Misinformation detection in COVID-19 news for low-resource languages: A Sesotho case study(2025-05) Mokoena, L; Rananga, S; Mbooi, Mahlatse SAs digitization in public healthcare increases, so does the spread of misinformation in health-related news. This presents significant challenges, particularly in low-resource communities such as those speaking Sesotho. This study focuses on enhancing and evaluating machine-learning models for detecting misinformation in Sesotho. The dataset used consisted of COVID-19-related misinformation from websites and social media and user engagement data translated from English to Sesotho. Two separate translation methods were performed to address the language barrier: one using Helsinki-NLP’s Opus-MT model and the other using Google Translate. Both translated datasets were then used to train different machine-learning models: Support Vector Machine (SVM), Logistic Regression, Decision Trees, Gradient Boosting, Random Forest, and Naive Bayes. The performance of these models was evaluated by using metrics such as accuracy, F1 score, precision, and recall. The results indicate that the quality of data has a substantial impact on the model's performance, with Helsinki-NLP’s translations generally outperforming Google Translate. This research addresses the limited availability of studies on automated misinformation detection in Sesotho and the lack of translated datasets. It also aims to improve the tools for protecting public health in similar language communities.Item DFT-based evaluation of Li2MnO3 as a promising cathode coating material for lithium-ion batteries(2025-07) Kgasago, M; Phoshoko, Katlego W; Ngoepe, P; Ledwaba, RAn ideal coating material should combine chemical stability, mechanical strength, and suitable conductivity to enhance cathode durability. Li2MnO3 has previously been used as a coating material due to its stabilizing effect on the core, but other beneficial properties it may offer as a coating material are still underexplored. In this study, these ideal coating properties of Li2MnO3 were investigated using Density Functional Theory (DFT). To enhance accuracy, spin configurations were also considered, and calculations were performed using the GGA+U functional. The findings show that Li2MnO3 is thermodynamically stable, mechanically robust, and a semiconductor with a band gap of 2.17 eV. These results affirm Li2MnO3 as a promising cathode coating material, possessing the key attributes which are thermodynamic, electronic, and mechanical stability needed to enable durable, high-performance lithium-ion battery systems.Item A bio-optical modelling study of inland waterbodies of the Western Cape, South Africa(2025-05) Pillay, H; Smith, Marie E; Lain, Lisl; O’Shea, R; Guild, L; Torres-Perez, J; Sharp, S; Gitari, WM; Mudzielwana, R; Pindihama, GRationale: Globally, high-quality in situ bio-optical measurements of optically complex aquatic environments remain limited, yet they are crucial for the parameterization and calibration of optical models and datasets essential for remote sensing applications in these contexts. During November 2023, the BioSCape campaign in the Western Cape, South Africa, facilitated the acquisition of a comprehensive array of synchronized biogeochemical data, inherent optical properties (IOP), and hyperspectral radiometric measurements across four biologically and optically diverse inland water sites. Study Aim: Evaluate different Hydrolight radiative transfer models, using a variety of input combinations that include both pre-existing IOP models and in situ measurements, to determine the most effective approach and model combinations for characterizing and replicating the environmental conditions of our studysites.Item Coupling radiative transfer models and machine learning for crop trait retrieval in dryland ecosystems(2025-07) Masemola, Cecilia R; Bonnet, W; Cho, Moses ARadiative Transfer Models (RTMs) such as PROSAIL, which integrates leaf-level (PROSPECT) and canopy-level (SAIL) reflectance simulations, are increasingly employed to support biophysical trait retrieval in crop monitoring applications. In this study, we assess and compare the performance of three PROSAIL configurations—PROSPECT-5 + SAIL, PROSPECT-D + SAIL, and PROSPECT-PRO + SAIL—for estimating Leaf Area Index (LAI) and Canopy Chlorophyll Content (CCC) in dryland maize systems using synthetic Sentinel-2 reflectance data. Results from synthetic test datasets indicate that the PROSPECT-PRO + SAIL configuration achieved superior performance, with LAI retrieved at an R² of 0.88 and RMSE of 0.35 m²/m², and CCC estimated at an R² of 0.83 and RMSE of 4.1 µg/cm². These outcomes highlight the advantage of using the enhanced biochemical and structural parameterizations in PROSPECT-PRO, especially under semi-arid cropping conditions. Comparative analysis confirms that this configuration consistently yielded the lowest normalized RMSE (nRMSE) for both LAI (9.5%) and CCC (10.7%) across the variants tested. The findings substantiate the added value of improved leaf optical modeling for accurate trait estimation and suggest that PROSPECT-PRO + SAIL provides a robust forward modeling basis for data-driven crop monitoring frameworks.Item Initial airborne experiment of South Africa’s first dedicated bistatic SAR receiver for µSTAR(2025-10) Dass, Reevelen; Nel, Willem AJ; Tema, Thabo H; Magaoga, Mpereke E; Mosito, Katlego E; Blaauw, Ciara; Botha, Louis; Van Zyl, Casper WThe µSTAR concept proposes a constellation of satellite illuminators and ground-based receivers to form a space-to-ground Bistatic Synthetic Aperture Radar (SAR). This approach lowers costs and latency compared to traditional spaceborne SAR by separating the transmitter and receiver. The Council for Scientific and Industrial Research (CSIR) has developed a first generation µSTAR receiver and conducted initial airborne tests, demonstrating the potential of this technology to provide high-resolution imagery for various applications.