Enhancing yield prediction from plot-level satellite imagery through genotype and environment feature...

Area of Science:

• Agricultural Science
• Computer Science
• Genetics

Background:

• Accurate yield prediction is crucial for crop management and resource allocation.
• Current methods rely on satellite vegetation indices (VIs) and machine learning (ML) models like PCA and AEs.
• Plot-scale yield prediction faces challenges in disentangling genotype and environment interactions.

Purpose of the Study:

• To enhance pre-harvest yield prediction at the plot-scale using Compositional Autoencoders (CAE).
• To improve the separation of genotype (G) and environment (E) features from high-resolution satellite imagery.
• To better incorporate genotype-by-environment (GxE) interactions for more accurate yield predictions.

Main Methods:

• Utilized a dataset of ~4,000 satellite images from replicated plots of 84 hybrid maize varieties across five U.S. Corn Belt locations.
• Applied a deep-learning approach, Compositional Autoencoders (CAE), to disentangle G and E features from plot-level satellite data.
• Evaluated CAE performance against traditional autoencoders (AEs) and vegetation indices (VIs) for yield prediction.

Main Results:

• CAE features improved early-stage yield predictions by up to 10% compared to traditional AEs.
• CAE outperformed VIs by 9% in yield prediction accuracy across various growth stages.
• CAE achieved a high silhouette score of 0.919 for environmental factor clustering, demonstrating effective separation.
• CAE showed superior performance in predicting yield for unseen environments and genotypes, indicating strong generalizability.

Conclusions:

• Compositional Autoencoders (CAE) offer a significant advancement in plot-scale yield prediction by effectively disentangling genotype and environment effects.
• The CAE model enables more accurate early-stage yield predictions and better modeling of GxE interactions.
• This approach supports informed decision-making in precision agriculture and accelerates plant breeding programs.