A predictive precision farming platform with online and offline weather-based advice for Nigerian farmers

Smallholder farmers make more than 90% of Nigeria's food supply but are largely excluded from AI-powered precision farming tools because of the low penetration of smartphones (less than 22%) and poor rural internet connectivity (less than 25%). This paper presents AgroCast, a dual-channel precision farming advisory platform that delivers weather-sensitive, crop-specific recommendations to Nigerian smallholder farmers regardless of their device type or connectivity status. The platform combines a web dashboard for internet-connected users with an Unstructured Supplementary Service Data (USSD) interface for basic feature phones and provides recommendations through Short Message Service (SMS). A total of 41,298 records were used to train crop yield prediction models for maize, rice, and soybean, using five machine learning algorithms using climate variables such as temperature, precipitation, and relative humidity as model inputs. The highest coefficient of determination (R²) was obtained for XGBoost with 0.972, 0.963 and 0.960 for maize, rice and soybean, respectively, thus indicating that the climate-yield relationship in Nigerian agriculture is best captured by gradient-boosted ensemble methods. Google's Gemini Large Language Model (LLM) was used to generate contextualised farming advice, with a deterministic rule-based fallback mechanism to ensure uninterrupted output in the case of API unavailability. All the user journeys were evaluated, and both interfaces were able to provide weather-based recommendations. AgroCast proves that AI-powered yield forecasting, LLM-driven advisory content, and the dual-channel access can help overcome the digital divide in agricultural extension services and provide a scalable solution for AI-driven food security interventions in low-connectivity developing areas.