An Original Computational Framework for Multi-Modal Crop Yield Prediction in West African Smallholder Farming Systems

Crop yield prediction in West African smallholder farming systems involves complex interactions of soil variability, micro-climate conditions, farming practices, and socioeconomic factors that are inadequately captured by models designed for large-scale monoculture agriculture. This study develops an original computational framework for multi-modal crop yield prediction adapted to West African smallholder contexts, using Nigeria as the primary development case. The framework integrates three novel components: a spatial-temporal multi-modal fusion architecture combining satellite remote sensing data (Sentinel-2 NDVI time series), weather station and gridded climate data, and farmer-reported management practice data through a cross-modal attention mechanism; a heterogeneous graph neural network layer that models the spatial relationships between adjacent smallholder plots to capture yield spillover and local agro-ecological effects; and a Bayesian uncertainty quantification module providing prediction intervals for decision support. The framework was trained on a dataset compiled from three growing seasons across 1,800 smallholder maize and cassava plots in Benue, Ondo, and Kano States, collected through the AfricaRice regional food security monitoring programme. Cross-validated RMSE for maize yield prediction was 0.31 tonnes/ha, outperforming the single-modality remote sensing baseline by 24 percent. The uncertainty quantification module demonstrated reliable calibration. The study constitutes an original computational methodology contribution and provides a publicly available framework implementation for West African agricultural AI research.

Keywords: crop yield prediction, smallholder farming, multi-modal learning, graph neural networks, Nigeria