A hybrid deep learning-based approach for optimal genotype by environment selection.

Area of Science:

• Agricultural Science
• Computational Biology
• Climate Science

Background:

• Accurate prediction of crop yields under varying weather conditions is essential for developing climate-resilient crop cultivars.
• Genotype-environment interactions significantly influence crop responses but are challenging to integrate into breeding programs.
• Machine learning offers a data-driven solution to predict yields by accounting for genotype-environment interactions.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting soybean yields.
• To investigate the effectiveness of ensemble methods in improving yield prediction accuracy.
• To identify key factors influencing soybean yield predictions.

Main Methods:

• Developed two convolutional neural network (CNN) models: a CNN model and a CNN-LSTM model, using a large dataset of soybean hybrid yields.
• Applied the Generalized Ensemble Method (GEM) to combine and optimize the CNN-based models.
• Evaluated model performance on unseen genotype-location combinations and analyzed feature importance.

Main Results:

• The GEM ensemble approach significantly improved prediction accuracy (RMSE and MAE) compared to individual CNN-LSTM and CNN models.
• Location, genotype, and year were identified as the most crucial predictors of soybean yield.
• Integrating state-level soil data did not substantially enhance model predictive capabilities.

Conclusions:

• The proposed data-driven GEM model provides a valuable tool for genotype selection, especially in scenarios with limited testing years.
• Machine learning, particularly ensemble methods, can effectively model complex genotype-environment interactions for yield prediction.
• Key weather variables like maximum direct normal irradiance and average precipitation are critical for accurate yield forecasting.