Assessing the Impact of Weather Variability on Crop Yield Predictions Using Machine Learning

Thesis Abstract

The increasing variability in weather patterns has emerged as a critical challenge to accurate crop yield prediction, directly impacting agricultural planning and food security in many regions. This study aims to assess the extent to which weather variability influences the accuracy of crop yield predictions through the application of machine learning techniques. Specifically, the research investigates the predictive performance of various algorithms—such as Random Forest, Support Vector Regression, and Artificial Neural Networks—in modeling crop yields under differing weather scenarios, with a focus on maize and wheat as key commodities. The study's primary objectives are to quantify the impact of fluctuating weather parameters (temperature, rainfall, humidity, and solar radiation), identify the most influential weather variables on crop yield outputs, and evaluate the robustness of machine learning models in accommodating weather variability. The research employs a quantitative, empirical design, utilizing historical weather and crop yield data collected over a ten-year period (2012-2021) from a stratified random sample of 30 agricultural zones within a temperate region known for significant climatic fluctuations. Data are sourced from national meteorological agencies and regional agricultural departments, with crop yield records obtained through official agricultural surveys. Instruments include climate data loggers and farm survey questionnaires, which ensure comprehensive coverage of weather parameters and crop management practices. To establish the validity and reliability of the data collection instruments, the study adopts test-retest procedures and cross-verification with secondary data sources. Analytical methods involve preprocessing data through normalization and handling missing values, followed by exploratory data analysis to identify key patterns. The core of the analysis employs supervised machine learning techniques—assessing model performance via metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²). The models are trained and tested using an 80-20 split, with hyperparameter tuning performed through grid search and cross-validation to optimize predictive accuracy. The study also applies the Theory of Adaptation and the Climate Variability Framework to interpret the influence of weather fluctuations on crop productivity, providing a theoretical basis for understanding how farmers and policymakers can leverage predictive models under climate uncertainty. Expected findings indicate that weather variability significantly influences the accuracy of crop yield predictions, with models incorporating real-time weather anomaly data outperforming those relying solely on historical averages. Variable importance analyses are anticipated to reveal that rainfall and temperature deviations are the most critical predictors across models. The study also expects to identify specific weather conditions under which model performance deteriorates, emphasizing the need for adaptive modeling approaches in climate-sensitive agriculture. This research contributes novel insights into the integration of machine learning for crop yield prediction in climates with high weather variability, extending existing literature by systematically evaluating model robustness and variable importance under dynamic weather conditions. It advances theoretical understanding by empirically validating the applicability of the Climate Variability Framework in agricultural prediction models, thereby bridging the gap between climate science and agronomic decision-making. In conclusion, the findings underscore the importance of incorporating real-time weather data into predictive models to enhance their accuracy, particularly in regions experiencing climate change-induced variability. Policy recommendations include promoting the deployment of adaptive machine learning-based decision support systems for farmers and agricultural planners, and developing region-specific weather forecasting models that can mitigate yield prediction errors. The study advocates for continued research into integrating remote sensing technologies and expanding the scope to include other crops and socio-economic factors influencing productivity, thus contributing to resilient agricultural systems amidst changing climatic conditions.