Accurate forecasting of corn yield is crucial in maintaining food security, facilitating farm planning, and maximizing resource allocation. Nevertheless, missing values in time series farm data are still a major challenge to model accuracy. This research introduces a comparison of the Gated Recurrent Unit (GRU) models for the corn yield forecasting based on the analysis of the effect of various data imputation techniques — K-Nearest Neighbors (KNN), Multiple Imputation by Chained Equations (MICE), and Expectation-Maximization (EM) — in combination with optimized activation functions, learning rates, and training times. Model performance was evaluated in terms of forecasting error metrics, namely Test Mean Absolute Error (MAE), Test Root Mean Squared Error (RMSE), and R-squared (R2), as well as convergence metrics like epoch of convergence and converged validation loss. Results showed that the MICE imputation technique with the ELU activation function had the lowest test error rates (MAE = 0.0007, RMSE = 0.0008, R2 = 0.9999) and demonstrated fast convergence within four epochs. The KNN-Leaky ReLU was tolerable but showed relatively slower convergence with higher error values. Although the EM-ELU model converged faster with less validation loss, the MICE-ELU configuration yielded the most accurate and generalizable predictions, making it the most suitable model for corn yield forecasting.