Abstract
Prediction of crop production is a challenging task, as it is influenced by various factors such as environmental, technological, and biological factors. In Afghanistan, wheat production is particularly more challenging due to high annual climate volatility like droughts and floods which directly impact yield. A reliable wheat production prediction can help address potential wheat demand accurately. The dataset used in this study includes annual wheat production data from the United States Department of Agriculture and annual climate data from the Climate Change Knowledge Portal for the years 1990 to 2023. After basic cleaning, a correlation and XGBoost-based feature selection was applied. SHAP analysis revealed that extreme climate indices such as maximum number of consecutive wet days (cwd) have a greater influence on production variability than annual mean temperature and precipitation. The selected variables were then used to train and compare Autoregressive Integrated Moving Average Model (ARIMA), Autoregressive Integrated Moving Average with Exogenous Variables (ARIMAX), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANN). The models are evaluated using Mean Absolute Error (MAE), Root Mean Squared Eerror (RMSE), and the coefficient of determination$R^{2}$. Our experimental results indicate that XGBoost performs the best, achieving the lowest RMSE, and MAE values, and a higher$R^{2}$score.