Smallholder farmers in Sub-Saharan Africa lack access to affordable tools for yield forecasting and early pest detection. This paper presents an end-to-end system that uses low-cost smartphone cameras combined with lightweight deep learning models to predict maize yield and detect fall armyworm infestation. We collected 18, 400 field images and 1, 200 plot-level yield measurements across Nigeria and Ghana over two growing seasons. A MobileNetV3-Small model for pest classification achieved 92. 1% F1-score on-device, while a multimodal CNN + tabular regression model predicted yield with RMSE = 0. 41 t/ha. We show that models trained on low-resolution images captured under variable field conditions generalize to unseen farms when augmented with weather and soil data. Our system runs at 18 FPS on a 80 Android phone, enabling real-time decision support without internet connectivity. Results demonstrate that low-cost mobile AI can provide actionable agronomic insights at scale for resource-constrained farmers.
Onyemaechi et al. (Wed,) studied this question.