Sequence Modeling Neural Network Model for Predicting Post-Harvest Losses (PHL) in Fufu: A Case Study of Selected Hubs in Oyo State

Abstract Post-harvest losses (PHL) in cassava-based fufu production remain a critical challenge to food security and income generation in Nigeria. Traditional methods for assessing spoilage such as manual inspection and static quality checks are often inadequate in capturing the dynamic conditions that lead to deterioration. This study explores the application of Sequence Modeling Neural Networks (SMNNs), particularly Long Short-Term Memory (LSTM) models, for predicting spoilage based on simulated time-series data such as temperature and humidity. A dataset of 11,230 responses was collected from selected fufu processing hubs in Oyo State, Nigeria, encompassing demographic, operational, and perception variables. Descriptive, inferential, and machine learning analyses including Random Forests, regression models, clustering, and PCA were conducted to evaluate spoilage risk factors and predict outcomes. Results showed that static features had limited predictive value, with traditional models yielding low accuracy and R² scores. In contrast, LSTM models trained on environmental sequences significantly outperformed classical approaches in forecasting spoilage probability. Additionally, K-Means clustering revealed distinct behavioral groups within the value chain, enabling targeted intervention design. The study concludes that real-time, sequence-based AI systems offer a more effective framework for reducing post-harvest losses in fufu production. It recommends the integration of sensor technologies, predictive mobile dashboards, and stakeholder training to facilitate the adoption of intelligent PHL management systems in cassava-rich regions.