Comparison of Deep Learning Models for Corn Yield Prediction: Predictive Performance and Sensitivity to Weather Extremes

Accurate prediction of crop yield is important for food security, resource planning, and climate adaptation. Traditional statistical models and early machine learning methods often struggle with complex spatial and temporal patterns in agricultural systems. Deep learning models such as Long Short Term Memory (LSTM) and Gated Recurrent Units (GRU) have improved yield prediction by learning nonlinear relationships from daily weather and soil data. Transformer models, which use self attention to capture long range dependencies, have recently been applied in some agricultural studies, but their performance for county level corn yield prediction and their behavior under extreme weather remain less well understood. This dissertation compares three deep learning models for county level corn yield prediction under strictly identical data and experimental settings: a bidirectional LSTM, a bidirectional GRU, and a transformer based PatchTST model. The models are trained on daily weather, vegetation, and soil data from 1980 to 2014 and evaluated on a test period from 2017 to 2020 for 943 counties in the United States Corn Belt. Yields are detrended with piecewise linear regression to separate technological trends from weather driven variability, and each model processes 183 day growing season sequences. The baseline comparison shows that the transformer provides the highest predictive accuracy, with an overall test R2 of 0.805 and RMSE of 12.79 bu/acre, compared to R2 = 0.773 and RMSE of 13.80 bu/acre for the LSTM and R2 = 0.758 and RMSE of 14.23 bu/acre for the GRU. A three step, percentile based perturbation framework is then used to test model responses to drought, heat, compound hot and dry scenarios, and a reconstructed 2019 wet event. Under the most severe heat and compound scenarios, the transformer model PatchTST predicts substantially larger yield losses than the two recurrent models, LSTM and GRU, while these recurrent models match the observed 2019 wet anomaly more closely. Taken together, these findings show that, for this county level yield prediction task, PatchTST provides the most accurate baseline forecasts, whereas LSTM and GRU produce more moderate responses to extreme perturbations and reproduce the 2019 wet anomaly more closely.