Performance Enhancement of GRU-Based Corn Yield Forecasting through MICE Imputation, PCA-Based Feature, and ELU Activation

Accurate corn yield forecasting is significant for farm agriculture to improve both farm production and productivity. In agriculture, forecasting is important to ensure food security. This study presents an improved Gated Recurrent Unit (GRU) forecasting approach that combines Multiple Imputation by Chained Equations (MICE), Principal Component Analysis (PCA), and Exponential Linear Unit (ELU) activation functions. This study is novel because it combines the PCA and ELU in a GRU architecture for agricultural forecasting. Previous study shows little to no work exploring this specific approach, or it is rarely seen in existing literature. MICE accurately imputes missing agronomic information, and PCA deals with multicollinearity and reduction of feature dimensions. This optimized input enhances gradient flow during training and mitigates the vanishing gradient issue common in deep recurrent models. Moreover, the application of the ELU activation stabilizes learning as it keeps small gradient values. The experiments showed that the model trained with dimensionality reduction with PCA, ELU activation, enhances performance with much higher accuracy than the baseline GRU models. The result produces fewer forecasting errors and consistent results with the actual yield values. These demonstrate that using the data imputation method combined with the ELU activation function enhances the performance of deep learning models in corn yield forecasting. This innovative solution gives farmers and agricultural planners managing a small farm or large operations the ability to make better decisions based on data.