DEEP LEARNING

Flooding along the River Niger remains a persistent hydrological and humanitarian challenge, exacerbated by data scarcity, climate intensification, and rapid land-cover transformation. This study advances a sustainable framework for flood prediction and risk mapping by integrating geospatial technologies with deep learning in data-limited riparian environments. The study focused on the Agenebode–Idah corridor, delineated by a 20 km buffer on either side of the River Niger. The research aimed to design, train, and validate a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN– LSTM) model capable of representing both the spatial complexity and temporal evolution of flood dynamics. The overarching purpose was to enhance anticipatory flood governance and inform climate-resilient development strategies across vulnerable communities of the lower Niger Basin. Methodologically, the study employed a multi-source geospatial architecture integrating CHIRPS rainfall, ERA5 soil moisture and temperature, SRTM topography, Sentinel-1 SAR inundation data, and Landsat-derived land-use/land-cover information. The CNN component extracted geomorphological and hydrological features, including elevation, slope, vegetation, and proximity to rivers, while the LSTM component learned the sequential dependencies of hydroclimatic variables over time. Model calibration and validation employed rigorous statistical and spatial metrics, including ROC–AUC, F1- score, RMSE, and confusion matrix analysis, alongside geospatial overlays between predicted and observed flood masks. The framework simulated both historical flood dynamics (2000–2024) and flood probabilities (2026–2035), allowing for a comprehensive spatiotemporal risk assessment. Results revealed that the hybrid CNN–LSTM achieved high predictive fidelity (ROC– AUC = 0.92; F1 = 0.86; RMSE = 0.081), outperforming baseline machine-learning and hydrodynamic models. Spatial analyses identified chronic flood hotspots in Agenebode, Anegbete, Ifeku Island, Ogurugu, and Idah. Forward projections indicated a 34% spatial expansion of high-risk zones by 2035, primarily across cropland and peri-urban settlements. The study concludes that coupling AI-driven deep learning with geospatial intelligence enhances flood forecasting capacity in data-scarce regions, providing a replicable, operational blueprint for early warning systems, spatial planning, and adaptive water governance in the face of accelerating climate variability.

This study presents the design and implementation of a smart, lightweight, federated deep learning system that integrates solar-powered cameras for automated attendance, unauthorized entry prevention and real-time cyber threat detection in academic environments. Using TensorFlow Lite, Python, and a Flask-based web interface, the model achieved high accuracy in facial recognition while maintaining low computational and energy costs. A structured SQLite3 database supported efficient local data handling, while solar energy integration enabled autonomous and sustainable operation. This project validates the potential of combining renewable energy, artificial intelligence, and federated learning to enhance classroom management and IoT security in low-resource settings.

Color detection is a task that humans perform effortlessly; however, enabling computers to accurately identify colors remains a challenging problem. In many industries, traditional color recognition systems rely heavily on manual processes and paid labor for color-coding items or datasets, which are often time-consuming, repetitive, and proneto human error. To address these limitations, this project presents a deep learning–based color detection program capable of recognizing multiple colors in real time. The system is implemented using Python, a high-level general-purpose programming language, in conjunction with the Open Source Computer Vision Library (OpenCV). By leveraging deep learning techniques, the proposed solution enhances accuracy and efficiency in automated color recognition tasks. The developed system enables computer devices to detect and classify multiple colors in real time, making it suitable for applications across various industries, including pharmaceutical manufacturing, autonomous vehicle development, and robotics. The adoption of this system can significantly reduce production time, minimize reliance on manual labor, and lower operational costs while improving overall productivity

This project implements a deep learning algorithm for image recognition, focusing on traffic sign classification. Traditional machine learning methods struggle with manual feature extraction and dataset diversity. To address these limitations, a robust convolutional neural network (CNN) with residual blocks, dropout layers, and global average pooling is utilized. Preprocessing techniques like normalization and data augmentation enhance accuracy and
generalization. Using TensorFlow and Keras, experiments were conducted on the German Traffic Sign Recognition Benchmark (GTSRB) and the Chinese Traffic Sign Dataset. The model achieved 99.54% validation accuracy and 94.95% test accuracy on GTSRB, but overfitting led to 60.38% accuracy on the smaller Chinese dataset. The study highlights CNN effectiveness in pattern recognition, with strengths in GPU acceleration and modular architecture. Challenges like overfitting and computational constraints persist. Future research should explore transfer learning, ensemble methods, and real-time optimization to enhance performance. This study advances deep learning-based
image recognition for applications in autonomous driving and traffic management. This project implements a deep learning algorithm for image recognition, focusing on traffic sign classification. Traditional machine learning methods struggle with manual feature extraction and dataset diversity. To address these limitations, a robust convolutional neural network (CNN) with residual blocks, dropout layers, and global average pooling is utilized. Preprocessing techniques like normalization and data augmentation enhance accuracy and generalization. Using TensorFlow and Keras, experiments were conducted on the German Traffic Sign Recognition Benchmark (GTSRB) and the Chinese Traffic Sign Dataset. The model achieved 99.54% validation accuracy and 94.95% test accuracy on GTSRB, but overfitting led to 60.38% accuracy on the smaller Chinese dataset. The study highlights CNN effectiveness in pattern recognition, with strengths in GPU acceleration and modular architecture. Challenges like overfitting and computational
constraints persist. Future research should explore transfer learning, ensemble methods, and real-time optimization to enhance performance. This study advances deep learning-based image recognition for applications in autonomous driving and traffic management.