Machine Learning Pipeline Engineering for Real-Time Agricultural Decision Support in Northern Nigeria

Agricultural decision support systems powered by machine learning offer transformative potential for food security in Northern Nigeria, yet engineering robust, scalable, and operationally sustainable ML pipelines for this context involves research challenges distinct from those addressed in dominant ML engineering literature developed for data-rich, infrastructure-stable environments. This dissertation develops a machine learning pipeline engineering framework for real-time agricultural decision support in Northern Nigeria, making original contributions to MLOps theory and agricultural AI systems engineering. A design science research methodology is employed across four experimental phases, spanning the North West (Kano, Kebbi), North Central (Benue, Niger), and North East (Gombe) zones. Phase one analyses existing agricultural decision support software deployments through 22 interviews with agritech developers and agricultural extension workers. Phase two designs the Northern Nigeria Agricultural ML Pipeline (NNAMLP) framework, addressing data ingestion from heterogeneous satellite, IoT sensor, and community-sourced data streams; feature engineering for Nigerian agro-climatic contexts; model selection and training on sparse, imbalanced Northern Nigerian agricultural datasets; and edge deployment for offline inference on low-specification farm devices. Phase three implements and evaluates NNAMLP across crop yield prediction, pest outbreak classification, and irrigation optimisation tasks on a combined dataset of 85,000 records. Phase four conducts a 12-month production deployment evaluation at 40 farms across two northern states, measuring system uptime, prediction accuracy under data drift, and farmer decision alignment with recommendations. The dissertation contributes NNAMLP as a validated engineering framework. Keywords: MLOps, agricultural AI, Northern Nigeria, machine learning pipeline, decision support