MACHINE LEARNING

Fisheries and aquaculture are vital to global food security but face critical challenges, including security threats from unauthorized access, the limitations of reactive manual monitoring, and uncertainty in long-term production forecasting. Existing management tools often lack a unified approach to integrate real-time operational surveillance with predictive decision-support analytics. This study aims to address these gaps by designing and developing a hybrid methodological framework that integrates computer vision for security and advanced machine learning and time-series forecasting for production analysis and stock management. The study employed a dual-component methodology. For security, an OpenCV-based Histogram of Oriented Gradients (HOG) descriptor combined with a Support Vector Machine (SVM) was implemented to achieve automated, real-time human intrusion detection and alarm triggering. For predictive analytics, a hybrid approach was utilized, applying Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX) for long-term production forecasting and various ensemble classifiers (Random Forest, Gradient Boosting, K-Nearest Neighbors) to analyze fish survivability and production trends. Results demonstrate that this integrated framework is highly effective across operational and strategic levels. The computer vision system successfully provided continuous, automated surveillance with verifiable digital evidence. The ensemble classifiers achieved near-perfect performance in survivability prediction (e.g., Random Forest Accuracy = 0.98, F1-score = 0.99, AUC ≈ 0.999). Furthermore, the SARIMAX model effectively projected a steady increase in global fish production, forecasting growth from 232.19 million tonnes in 2024 to 272.12 million tonnes by 2043—a 17.19% increase—capturing long-term temporal dynamics that static regressors fail to model. Based on these findings, it is recommended that fisheries stakeholders prioritize the deployment of computer vision-based surveillance to reduce security risks and labor costs. Additionally, management authorities should integrate ensemble classifiers into stock-management systems to optimize survivability and adopt SARIMAX-based forecasting for robust, evidence-based policy planning. The implementation of these data-driven platforms will significantly enhance sustainable fisheries governance and support proactive resource allocation.

Machine learning has become an important technology used in many areas such as healthcare, banking, education, and mobile applications. However, traditional machine learning methods usually require large amounts of data to be collected and stored in one central location, which can create privacy and security concerns. This project focuses on the design and implementation of a Federated Learning system for privacy-preserving machine learning. Federated Learning is a modern approach that allows multiple users or organizations to train a machine learning model together without sharing their raw data. Instead of sending personal or sensitive information to a central server, each participant trains the model on their local device and only shares model updates. This helps to protect privacy while still allowing the model to learn effectively. In this project, a federated learning prototype was developed using Python and machine learning tools. The system was designed with features such as secure aggregation and differential privacy to improve data protection and reduce the risk of information leakage. The performance of the system was tested using distributed datasets to simulate real-world situations where data is stored in different locations. The results showed that the proposed system was able to maintain a good level of model accuracy while protecting user privacy. It also demonstrated that federated learning can be a practical solution for organizations that need to use machine learning without exposing sensitive data. The study concludes that federated learning has great potential for institutions such as the University of Benin and other organizations in Nigeria that require secure and privacy-aware machine learning solutions.

The increasing demand for flexible, affordable, and collaborative learning has created opportunities for digital platforms that support peer-to-peer skill exchange. This study focuses on the development of an online skill barter-based system for personalized competency matching using machine learning. The system enables users to offer and learn skills through direct exchange rather than monetary transactions. A K-Nearest Neighbor (KNN) algorithm was integrated with TF-IDF and Natural Language Processing (NLP) techniques to intelligently match users based on skill similarity, learning preferences, and availability. The platform was developed using Next.js and TailwindCSS for the frontend, with Supabase as the backend for authentication and data management. The system promotes inclusive and accessible learning by removing financial barriers while fostering collaboration and community growth. This research demonstrates that integrating machine learning into skill exchange platforms significantly enhances user experience, match accuracy, and trust in non-monetary, peer-driven learning environments.

Flooding remains one of the most pressing environmental challenges in the Niger Delta region of Nigeria, where rapid urbanization, land degradation, and climate variability interact to intensify hydrological extremes. This study evaluates the combined impact of land use/land cover (LULC) changes and climate variability on flooding in the Lower Niger Region, focusing on Edo, Delta, Bayelsa, and Rivers States. An integrated methodological framework was developed using remote sensing (RS) and machine learning (ML) techniques implemented on the Google Earth Engine (GEE) platform. Rainfall and temperature datasets were obtained from multiple satellite products, CHIRPS, PERSIANN-CDR, PERSIANN-CCS, ERA5, and CPC, and validated against observations from the Nigerian Meteorological Agency (NiMET). LULC classification was conducted using Random Forest (RF) and Support Vector Machine (SVM) algorithms applied to Sentinel-2 and Landsat-8 imagery, generating five major land cover classes: water bodies, forest, barren land, vegetation, and built-up areas. Flood dynamics were assessed using multiple spectral indices, including AWEI, FWI, NDWI, MNDWI, WRI, and NDVI, evaluated across different temporal scales to capture hydrological variations. The results reveal significant interannual and seasonal rainfall variability, with the Standardized Precipitation Index (SPI) effectively identifying alternating wet and dry cycles, although long-term annual rainfall trends were largely non-significant. There is, however, evidence of increasing frequency of extreme rainfall events, particularly in Delta and Rivers States. Temperature analysis (1971–2023) indicates a statistically significant warming trend, with minimum temperatures rising faster than maximum temperatures—most notably during the dry season (DJF). LULC assessment confirmed the superior performance of the RF classifier over SVM and showed a consistent pattern of urban expansion at the expense of vegetation and wetlands. Flood analysis revealed a persistent hotspot occurring along river corridors, low-lying floodplains, and southeastern basins. Overall, the findings demonstrate that while climate variability influences hydrological extremes, LULC transformations, particularly urban encroachment into natural flood buffers, have been a more decisive factor driving flood vulnerability in the Lower Niger Region.

Electronic mail (email) remains a primary communication channel globally, making it a prime target for malicious actors. The continuous rise in spam emails poses severe threats, including network congestion, financial losses, and security vulnerabilities like phishing and malware distribution. Traditional rule-based spam filters struggle to keep pace with the evolving tactics of pammers, creating a critical need for more adaptive and robust solutions. This project presents the Design and Implementation of a Spam Email etection System Using Artificial Intelligence.The system leverages Natural anguage Processing (NLP) and Machine Learning techniques to accurately classify mails as "Spam" or "Ham" (legitimate). The methodology involves data collection using a standardized public dataset, followed by intensive text preprocessing including tokenization, stop-word removal, and stemming. Text data is converted into numerical vectors using Term Frequency-Inverse Document requency (TF-IDF). Multiple AI algorithms, including Naïve Bayes, Support Vector Machines (SVM), and Random Forest, are implemented and evaluated to determine the most effective classifier.The system was developed using Python, utilizing libraries such as Scikit-Learn and NLTK, and deployed with a user-friendly graphical interface. Experimental results demonstrate that the AI-powered system achieves high performance, with the top-performing model attaining an accuracy rate exceeding [Insert Accuracy, e.g., 98.2%] and a low false-positive rate. This research concludes that integrating artificial intelligence significantly enhances email security by providing dynamic, real-time protection against malicious messages, thereby ensuring a safer digital communication ecosystem.

Phishing attacks, which use phoney emails and misleading websites to target people and organisations, have emerged as one of the most common cybersecurity threats. These attacks cause serious security breaches by tricking users into divulging private information, including login credentials and financial information. This project offers a machine learning-based phishing detection tool for emails and web browsers that uses logistic regression to efficiently detect and stop phishing threats. To improve the accuracy of phishing detection, the study investigates different feature extraction methods, dataset preprocessing, and model training approaches. Python was used to develop the system, and the Scikit-learn library was used to implement the model. Both authentic and phishing URLs and emails made up the dataset, which was preprocessed using techniques like feature scaling and outlier removal. Key performance metrics, such as accuracy, precision, recall, F1-score, and a confusion matrix, were used to train and assess the model. The outcomes showed that the suggested model was successful in differentiating between phishing and authentic entities, as evidenced by its high classification accuracy. Notwithstanding its encouraging results, the system has certain drawbacks, including the difficulty of balancing false positives and false negatives and vulnerability to new phishing tactics. To further enhance detection capabilities, future studies could use ensemble approaches or sophisticated deep learning models. By offering users and organisations a reliable and effective detection mechanism, this project supports the continuous efforts to strengthen cybersecurity defences against phishing attacks.

This study develops a machine learning model to predict abnormalities in commercial airplanes using real-world Automatic Dependent Surveillance-Broadcast (ADS-B) data, focusing on altitude changes exceeding 100 feet in 10 seconds. Following the methodology established by Passarella et al. (2024), this research implements and compares 25 different machine learning algorithms, ultimately selecting Quadratic Discriminant Analysis (QDA) as the optimal approach. The dataset comprises 167,844 records, including 84,074 normal and 83,770 abnormal instances, with features such as altitude, velocity, heading, latitude, and longitude. The theoretical foundation covers the comprehensive taxonomy of machine learning methods, from supervised learning algorithms like Support Vector Machines and Decision Trees to unsupervised approaches such as K-Means clustering. The QDA model achieves superior performance with 93-97% accuracy, 0.96-0.97 ROC-AUC, validated through stratified 5-fold cross-validation. Visualizations, including altitude plots and ROC curves, enhance interpretability for aviation professionals. This research demonstrates that QDA's ability to model non-linear decision boundaries with class-specific covariance matrices makes it particularly suitable for complex aviation data patterns, supporting enhanced flight safety and operational efficiency.

Manual pavement inspection methods are slow, subjective, and often inconsistent, leading to delayed maintenance and increased road deterioration. This study was carried out to develop an automated, image-based system capable of detecting and classifying visible defects in flexible pavements using machine learning. The objectives of the study were to review existing pavement inspection techniques, collect and preprocess pavement image data, and design and train a model capable of identifying pavement failures accurately. The study was with the aim of improving the speed, objectivity, and reliability of pavement condition assessments. A dataset of pavement images was obtained from the Edo State Ministry of Works, field surveys, and public sources. The images were annotated in YOLO format and augmented by flipping, rotation, cropping, and brightness adjustment. The YOLOv8 object detection model, implemented in Python using TensorFlow, PyTorch, and OpenCV, was trained on Google Colab with an NVIDIA T4 GPU. Training was performed at varying epochs (50, 100, and 200) and hyperparameters to optimize detection performance. The model’s accuracy was evaluated using mean Average Precision (mAP) and recall metrics to assess its ability to detect cracks, potholes, and rutting in flexible pavements. Results showed that the model achieved a mean Average Precision (mAP₅₀) of 0.68 and recall above 0.80 for visible defects such as potholes and alligator cracking, at a confidence level of 0.5. The model was less effective in detecting faint, low-contrast linear cracks. This study concluded that YOLOv8-based models can effectively automate pavement distress detection, providing a faster and more reliable alternative to manual inspection. It is recommended that future work expand the dataset and explore enhanced training strategies to improve the detection of subtle linear cracks.

This project work provides an overview on the application of linear algebra to artificial intelligence including natural language processing and machine learning. We discuss how linear algebra operations such as matrices, linear transformations, eigen values and eigen vectors, are used to optimize AI models, analyze complex data structures and enable efficient computation. Beginning with an overview of fundamental concepts in linear algebra, such as vectors, matrices, and linear transformations, the study delves into specific applications of these concepts in AI. One key area of focus is machine learning, where linear algebra forms the backbone of algorithms for tasks such as regression analysis, and principal component analysis for dimensionality reduction. This work also showcases the versatility of linear algebra by delving deep into the various reaches of linear algebra into many other fields and areas of study such as economics, physics and engineering.

Malaria remains a global health crisis, particularly in low-resource regions, where traditional diagnostic methods face challenges such as human error, resource constraints, and delayed detection. This project addresses these limitations by leveraging supervised machine learning (ML) to enhance malaria diagnosis and outbreak prediction. The motivation stems from the urgent need for scalable, accurate, and cost-effective solutions to reduce the disease’s burden, which claims over 600,000 lives annually. The objective is to develop robust ML models capable of automating malaria diagnosis using blood smear images and patient metadata while improving outbreak forecasting through environmental and epidemiological data analysis. Methodologically, the study employs supervised learning algorithms, including convolutional neural networks (CNNs) for image- based detection and random forests for tabular data. Datasets were preprocessed to handle class imbalance and missing values, followed by hyperparameter tuning and cross-validation to optimize performance. Results demonstrated that CNNs achieved 96% accuracy in classifying infected blood cells, outperforming traditional methods like microscopy. Random Forest models yielded 92% recall and 89% precision in predicting malaria risk from clinical data, highlighting their utility in early diagnosis. Additionally, stratified k-fold cross-validation ensured model generalizability across diverse datasets. This work underscores the transformative potential of supervised ML in malaria control, offering tools that enhance diagnostic speed, accuracy, and accessibility. By bridging technological innovation with public health needs, the project contributes to global efforts toward malaria eradication, particularly in endemic regions