Smartphone-based Farm Yield Forecasting with AI in Smallholder Markets

Thesis Abstract

Smartphone-enabled predictive analytics for smallholder crop yield is increasingly essential amidst climate variability, rising input costs, and fragmented market access. This study addresses the gap in accessible, accurate yield forecasting tools that smallholder farmers can use via ubiquitous mobile devices to improve decision-making, risk management, and market participation. The aim is to develop and validate an AI-driven smartphone forecasting framework that integrates agronomic, weather, and market data to produce actionable yield projections at the farm and village levels. Specific objectives are (1) to design a mobile app architecture that ingests real-time weather, soil, and crop management data; (2) to develop machine learning models that forecast end-of-season yield with quantified uncertainty; (3) to assess the performance of models under diverse agro-ecological zones and smallholder contexts; (4) to evaluate user acceptance, usability, and decision-usefulness among farmers and extension agents; and (5) to analyze the potential impacts on input use efficiency, risk management, and market access. The study adopts an explanatory sequential mixed-methods design. The research population comprises smallholder maize and sorghum farmers across three agro-ecological zones within a representative sub-national region. A sample of 600 farmers will be recruited using stratified random sampling to ensure variation in farm size, access to extension services, and mobile connectivity. Data collection will combine quantitative and qualitative instruments a structured survey to capture agronomic practices, input use, yields, and socio-economic indicators; historical meteorological and soil data sourced from publicly available government and satellite datasets; and in-app usage logs recording user interactions and forecast utilization. The smartphone application will implement machine learning pipelines, including gradient boosting (XGBoost) and deep learning models (LSTM) for time-series yield forecasting, with Bayesian methods to quantify predictive uncertainty. Feature engineering will incorporate weather anomalies, phenological indicators, soil moisture, fertilizer timing, and crop insurance status. Model validation will employ cross-validation and out-of-sample testing, with performance metrics such as RMSE, MAE, R-squared, and calibration plots. Causal analysis will be conducted via generalized linear models to explore relationships between forecast accuracy, farmer decisions (input adjustments, planting dates), and yield outcomes. The theoretical lens integrates the Technology Acceptance Model (TAM) to interpret user adoption, and the Resource-Based View (RBV) to frame the capacity of smallholders to leverage digital forecasts for competitive advantage. The study will also draw on the Theory of Planned Behavior to contextualize behavioral responses to forecast information. Thematic analysis of interview data with 60 stakeholders (farmers, extension officers, market actors) will identify barriers and enablers to forecast use and market integration. Expected findings indicate that AI-enhanced forecasts delivered through the smartphone app achieve significantly lower RMSE and MAE compared with baseline climatological forecasts, with acceptable calibration and narrow prediction intervals under diverse conditions. Model performance is anticipated to be robust across zones with transfer learning capabilities, and forecast usage is expected to correlate with improved decision timing for fertilizer applications, irrigation, and harvest planning. The study is likely to reveal differential effects of forecast accuracy on farm-level outcomes, moderated by farmer literacy, trust in technology, and connectivity. The contribution to knowledge lies in empirically validating a scalable, low-cost ICT solution for yield forecasting in smallholder systems, integrating state-of-the-art AI methods with user-centered design, and advancing understanding of how digital forecasts influence farm management and market participation. The final section will provide practical recommendations for policymakers, development agencies, and technology providers, including data governance protocols, cost-effective deployment strategies, training curricula, and guidelines for integrating forecast insights into extension services and commodity markets. The conclusion asserts that smartphone-based AI yield forecasting can enhance productivity, reduce risk exposure, and strengthen smallholders’ bargaining power, with policy support focusing on digital literacy, data privacy, and rural connectivity improvements.