joseph-bozimo/AI-Geofencing-Agadama

Official code repository for MSc Dissertation in Environmental Resource Management, Department of Geography, Nasarawa State University Keffi. Candidate: Bozimo Edisemi Joseph (AMS23ERM0036). # Management of Farmers-Herders Conflicts Using AI-Powered Geofencing in Agadama, Nigeria An enterprise-grade, data-driven framework and predictive deep learning pipeline engineered to mitigate pastoralist-agrarian resource friction in the Agadama community, Ughelli North, Delta State. This project moves beyond conceptual designs to establish a functional, high-sensitivity AI early-warning perimeter. ## 🚀 Core Architecture & Technological Stack * **Geospatial Processing Engine:** Integration of WGS 84 coordinate reference grids to map micro-level localized agricultural zones. * **Automated Data Harvesting Pipeline:** Programmatic API ingestion pipelines querying dynamic agro-climatological points (NASA POWER API), demographic footprints (WorldPop Project), and spatial surface categorization (Copernicus Land Monitoring Service). * **Algorithmic Optimisation (SMOTE):** Addresses severe real-world data sparsity and minority-class event imbalances by synthetically over-sampling active friction records from 28 to 1,725 vectors. * **Modern Deep Learning Predictor:** A regularized 3-layer feedforward Deep Neural Network (DNN) built with TensorFlow and Keras, integrating Batch Normalization and Dropout layers (0.30) to eliminate majority-class classification bias. ## 📊 Empirical Breakthroughs & Performance Evaluation Transitioning the framework from traditional shallow machine learning models to a balanced deep learning network successfully unlocked complex non-linear climate-demographic relationships: * **Shallow Baseline (Random Forest):** Exhibited an overall accuracy of 97.95% but recorded an absolute failure mode of **0.00% Recall and Precision**, proving completely blind to threat vectors due to class imbalance (AUC = 0.46). * **Advanced Deep Learning Pipeline (SMOTE-DNN):** A fundamental breakout from the zero-capacity trap, achieving an active threat **Recall rate of 57.14%** and a functional **0.61 AUC score** on untouched baseline evaluation sets. ## 📂 Reposit …