The Nelson Mandela African Institution of Science and Technology
Institutional Repository
Research Collection
Select to browse Research collections.
Recent Submissions
Development of deep learning model using satellite images to enhance forest-fire occurrence prediction linked with seasonality at Mount Kilimanjaro, Tanzania
(NM-AIST, 2026-06) Mambile, Cesilia
Forest fires represent a growing environmental and socio-economic challenge, particularly in ecologically sensitive regions such as Mount Kilimanjaro in Tanzania. Accurate prediction of forest fire outbreaks is critical for timely response, resource allocation, and environmental protection. Traditional models and statistical techniques have been in use for decades but often struggle with capturing spatial-temporal complexity and integrating important human activity data. While deep learning approaches show potential, they encounter challenges including data imbalance, noise, and limited validation against real-world events. This study aimed to develop a deep learning-based fire prediction model incorporating seasonality and human activities specific to the Mount Kilimanjaro area. A ConvLSTM model was trained using integrated data sources, including Sentinel-2 satellite imagery, meteorological patterns, vegetation indices (NDVI, NDWI, NBR), and anthropogenic features such as beekeeping and tourism activities. The validation process employed a three-tier framework: Statistical evaluation using classification metrics, historical validation using VIIRS fire records, and expert validation from fire management professionals. The model achieved high performance, with an accuracy of 97.65%, an AUROC of 0.9512, a precision of 98.55%, and a recall of 92.85%. Spatial and temporal analyses confirmed that the model effectively predicted fire-prone zones and seasonal trends, with a spatial alignment of 93.2% to observed fire events. Expert review further affirmed the model’s practical applicability: 87% of experts acknowledged its accuracy, and 82% observed alignment with historical patterns. However, concerns were raised regarding data resolution, the need for real-time deployment capabilities, and local capacity for model use. These concerns were beyond the scope of the current study and are identified as important areas for future exploration. The findings demonstrate the model’s strength in delivering interpretable, location-specific fire risk predictions while addressing common limitations in existing systems. By merging scientific modeling with local data and expert feedback, the study offers a scalable, AI-powered framework for fire early warning systems in climate-vulnerable landscapes. Future research should explore integrating real-time meteorological feeds, uncertainty quantification, and adaptive features for broader deployment across different ecological regions.
Agent-based modelling to predict water rationing and allocation enhanced by machine learning techniques: case study of the pangani basin
(NM-AIST, 2026-05) Lyuba, Matimbila
Water scarcity and inequitable distribution remain major challenges in many river basins, including the Pangani Basin in Tanzania. Previous studies have applied machine learning techniques to predict water consumption and Agent-Based Modelling (ABM) to support water allocation. However, the machine learning approaches used were often not validated for robustness or evaluated for generalization to unseen environments. Similarly, ABM approaches used relies on generated datasets, which introduce uncertainties when models are deployed in real-world settings. This study developed an ABM for the Pangani Basin to simulate water allocation, rationing, and reallocation in response to these challenges. Water user characteristics were identified through validated multi-model characterization using K-means, Hierarchical Agglomerative and Fuzzy C-means algorithms; and further enriched through association rule mining with minimum support ranging from 0.5% to 5.0% and confidence levels above 50%. The identified characteristics were then modelled as agents in NetLogo to simulate socioeconomic activities and predict their water abstraction through allocation, rationing, and reallocation. The dataset was analyzed using Interquartile Range (IQR) with capping, mean imputation, Principal Component Analysis (PCA), standardization, and Hopkins test. Cluster validation was conducted using the Davies-Bouldin Index, Silhouette Score, Calinski-Harabasz Index, and Dunn Index, while clustered data were evaluated using logistic regression model to determine prediction accuracy. The ABM was validated using both a paired t-test and expert evaluation through a developed web tool. The study identified four distinct clusters of water users with varying behaviors influencing water abstraction. The ABM simulation revealed that predicted mean abstraction values generally decreased as the number of users increased, with predicted values remaining below actual abstractions across the four clusters. Paired t-test results showed that differences between predicted and actual mean values were not statistically significant. Expert validation further indicated that the ABM predictions reflected field realities and could generalize to other basins with different geographical settings. The findings demonstrate improved allocation efficiency, reduced user conflict, and enhanced system resilience under climate variability and rising demand. Overall, the study shows that integrating ABM with machine learning provides a robust, adaptive, and data-driven framework for optimizing water rationing and allocation in the river basins.
Mathematical modelling of cryptosporidiosis transmission dynamics and control in humans and cattle
(NM-AIST, 2025-12) Luhanda, Faraja
Cryptosporidiosis is a globally significant zoonotic disease caused by Cryptosporidium, posing a substantial concern for public health and veterinary medicine. Despite the significant contribution of cattle and human immune status in shaping the transmission dynamics and control of cryptosporidiosis, there is limited evidence of mathematical models that concurrently incorporate this aspect. This study aimed to investigate the transmission dynamics and control of cryptosporidiosis in humans and cattle using mathematical modelling. Non-linear ordinary differential equations were used to formulate deterministic models, from which the assumptions for the corresponding continuous-time Markov chain (CTMC) stochastic models were derived. Using the next-generation matrix technique, the basic and effective reproduction numbers were computed. Latin Hypercube Sampling was applied to compute Partial Rank Correlation Coefficients, allowing for a global sensitivity analysis of the parameters influencing disease transmission dynamics. The existence of disease-free and endemic equilibria was well established, and their stability was examined using the Routh-Hurwitz criterion and the Lyapunov function method, respectively. The multitype branching process was employed to compute the likelihood of disease extinction or outbreak in the CTMC stochastic models. Lastly, the optimal control theory was applied to determine the most effective strategy for controlling and preventing cryptosporidiosis in humans and cattle. The results revealed that cattle significantly influence the dynamics of the disease. Immunocompromised humans contribute more substantially to the dynamics of the disease than immunocompetent humans. The persistence of cryptosporidiosis is influenced by increased disease transmission rates and higher shedding rates of Cryptosporidium oocysts into the environment by infectious cattle and humans. Cryptosporidiosis is more likely to extinct if it emerges from infected immunocompetent human compartments than from infected immunocompromised compartments. Contrarily, a disease major outbreak is probable if it originates from either infected cattle compartments or Cryptosporidium oocysts in the environment. The optimal control model findings suggest that the best strategy for controlling and preventing cryptosporidiosis is the simultaneous implementation of all six control efforts, grouped into preventions, treatments, and environmental decontamination. Thus, efforts to prevent and control cryptosporidiosis should concentrate on eliminating Cryptosporidium oocysts from the environment, treating infectious cattle and humans, maintaining proper sanitation practices and sound cattle farm management and hygiene practices.
Epidemiological Characteristics Of Imported Malaria In Pemba Island-Zanzibar: A Study From 2020 To 2022
(NM-AIST, 2026-01) Ali, Ali
Imported malaria remains a major obstacle to elimination efforts on Pemba Island. This study conducted a retrospective analysis of surveillance data from 2020–2022 to describe the spatial and temporal patterns, demographic characteristics, and origins of imported malaria cases. Entomological data from four sentinel sites were incorporated to assess area receptivity. Negative binomial regression was used to evaluate associations with geographic location, season, sex, and transmission intensity at travel destinations, while univariate and multivariate logistic regression assessed the odds of importation. A total of 2646 malaria cases were reported during the study period, of which 1411 (53.3%) were classified as imported. Imported infections were more common among males (57.3%) and adults aged ≥18 years (68%). Most cases originated from mainland Tanzania, predominantly Tanga (61%), followed by Dar es Salaam (12%) and the Coastal Region (10%), with only 2% linked to neighbouring countries. Spatial analysis showed reduced importation in Micheweni (IRR = 0.48, p = 0.001) and elevated risk in Mkoani (IRR = 2.45, p < 0.001). Imported cases were less frequent during the rainy season (IRR = 0.83, p = 0.049) and among individuals returning from low or moderate transmission settings. The likelihood of importation was significantly higher in 2021 (OR = 3.58) and 2022 (OR = 5.62) compared with 2020. No significant correlation was observed between imported case distribution and Anopheles biting density; however, Human Biting Rate (HBR) showed a strong positive association, identifying Mkoani as the most receptive district. These findings confirm imported malaria as a persistent barrier to elimination and highlight the need for strengthened cross-border surveillance, targeted interventions, and closer integration of entomological and epidemiological monitoring to inform elimination strategies.
A framework for designing web-based systems to disseminate crop production knowledge in Tanzania
(NM-AIST, 2026-07) Victor Ngessa
Smallholder farmers in Tanzania face critical gaps in accessing timely, relevant crop production knowledge due to inefficiencies in traditional extension systems. While digital platforms offer potential solutions, existing web-based systems often fail to address farmers’ literacy levels, technological access and localised needs, limiting their adoption and impact. This dissertation introduces a framework designed to enhance the utilisation of web technologies for disseminating crop production knowledge to smallholder farmers in Tanzania; a demographic essential to national food security. Data were collected through questionnaires, interviews and observations, involving 21 organisations and 827 farmers across eleven regions. The literature review revealed that the system’s purpose, farmer engagement, multimedia resource utilisation and ease of access are essential for effective knowledge dissemination. Additionally, the importance of high-quality, relevant and engaging content tailored specifically to smallholders was underscored. The study found that government institutions accounted for most website-based dissemination efforts (57.2%), while traditional methods remained prevalent. Websites typically contained content spanning all crop production stages, which 47% of participants regarded they were highly important to smallholder farmers compared to 81% to organisational staff and 71% to donors. Although 54% of farmers owned internet-enabled devices, only 40% demonstrated proficiency in internet usage. However, 91% of the farmers expressed willingness to engage with online learning. The national ICT policy further supports web-based agricultural knowledge dissemination. A participatory approach was used to develop a web-based system for knowledge dissemination. Its evaluation by smallholder farmers yielded a System Usability Scale (SUS) score of 87.5. The framework was constructed using the design science research methodology. The proposed framework promotes high-quality, well-organised, multi-sourced knowledge, presented in multimedia formats to enhance learning. It emphasises fast-loading systems and cross-device compatibility. Eight experts validated the framework, confirming its core components, identifying refinement needs, and refining its final structure. This research contributes substantially to the field of knowledge dissemination and serves as a reference for policymakers, researchers, and practitioners in Tanzania and similar contexts. Stakeholders in e-extension services are encouraged to adopt the framework to improve the design of web based systems for smallholder farmers.