Corn Yield Prediction With Ensemble CNN-DNN.

Area of Science:

• Agricultural Science
• Machine Learning
• Data Science

Background:

• Accurate corn yield prediction is crucial for food security and agricultural economics.
• Existing machine learning models have limitations in capturing complex spatio-temporal yield variations.

Purpose of the Study:

• To develop and evaluate novel Convolutional Neural Network-Deep Neural Network (CNN-DNN) ensemble models for predicting county-level corn yields.
• To compare the performance of homogenous and heterogenous ensemble approaches against individual machine learning models.

Main Methods:

• A dataset combining management, environmental, and historical corn yield data (1980-2019) was utilized.
• Two ensemble scenarios were explored: homogenous (identical base models with bagging) and heterogenous (same architecture, different hyperparameters).
• Ensembles were created using Basic Ensemble Method (BEM), Generalized Ensemble Method (GEM), and stacked generalized ensembles.

Main Results:

• Both homogenous and heterogenous CNN-DNN ensembles outperformed individual machine learning models (linear regression, LASSO, random forest, XGBoost, LightGBM).
• Homogenous ensembles demonstrated superior accuracy in corn yield predictions across US Corn Belt states.
• The best model achieved a root mean square error of 866 kg/ha (8.5% relative RMSE), explaining 77% of yield variation.

Conclusions:

• Homogenous CNN-DNN ensemble models provide highly accurate and reliable corn yield predictions.
• These models can be instrumental in developing decision-support tools for agronomists.
• The findings highlight the potential of advanced machine learning for optimizing agricultural management and planning.