Iyinoluwa Oyelade is a Lecturer II in the Department of Information Technology at the Federal University of Technology, Akure (FUTA), Nigeria. She holds a Ph.D. in Computer Science from FUTA, with research focused on developing an Internet-of-Things-based farmland intrusion detection model. Her research interests include Artificial Intelligence, Machine Learning, Deep Learning, Computer Vision, Internet of Things (IoT), and intelligent systems. Her research applies these technologies to real-world challenges in agriculture, healthcare, and climate-related domains. She has published several research articles and conference papers on topics including livestock disease detection, farmland intrusion detection, facial emotion recognition, healthcare systems, and IoT applications. Dr. Oyelade is also actively involved in teaching, research, STEM education, and professional development. She has received recognition including the 3rd Best Poster Award at the 7th NAS Scientific Conference and the Female Lecturer of the Year award at FUTA.

  Modeling Agricultural Drought Severity using Explainable Machine Learning Models Enhanced with Bayesian OptimizationAbstractAgricultural drought poses an escalating threat to food security across the semi-arid Sudano-Sahelian belt of Northern Nigeria, yet operational early-warning systems in the region remain constrained by the trade-off between predictive accuracy, calibrated uncertainty, and interpretability. This study develops and evaluates a Bayesian-optimized, explainable machine learning framework for forecasting moderate agricultural drought (6-month Standardized Precipitation-Evapotranspiration Index, SPEI-6, ≤ −1.0) one month ahead across approximately 3,600 grid cells (0.05° resolution) spanning the Sudano-Sahelian states of Northern Nigeria. A gradient-boosted XGBoost classifier and regressor were tuned via Bayesian hyperparameter optimization (20 and 12 search trials respectively) over 144 engineered features derived from CHIRPS precipitation, ERA5 temperature and potential evapotranspiration, TAMSAT rainfall, MODIS NDVI, and SMAP soil moisture, and benchmarked against a Logistic Regression baseline, a probability-averaged ensemble model of the two, and a naive Persistence baseline, under chronological (temporal) holdout evaluation across two seasonal windows — the operationally relevant growing season and the full annual cycle. The Bayesian-optimized, sigmoid-calibrated XGBoost classifier achieved the strongest discrimination (AUC-ROC = 0.928, PR-AUC = 0.382 for the growing-season window), while the ensemble traded precision for markedly higher recall and balanced accuracy (recall = 0.693, balanced accuracy = 0.802), a property directly relevant to early-warning systems that prioritise not missing drought onset. The companion XGBoost regressor explained 51.7% of the variance in continuous SPEI-6 values (RMSE = 0.670, Spearman ρ = 0.776) but exhibited a well-defined compression bias, systematically underestimating the most extreme wet and dry anomalies. SHAP (SHapley Additive exPlanations) analysis identified the current-month SPEI-6 anomaly as the dominant driver of predicted drought probability by a wide margin, followed by its one-month lag, water-balance anomaly, and TAMSAT rainfall, confirming that the model's reasoning is physically grounded in precipitation and water-balance deficits rather than spurious correlations. These findings demonstrate that Bayesian hyperparameter tuning, explicit probability calibration, and post-hoc explainability can jointly deliver a transparent, accurate, and operationally interpretable one-month-ahead drought early-warning tool.Keywords: Machine Learning, Bayesian Optimization, Agricultural Drought, XGBoost, Linear Regression, Ensemble Modeling.