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

Accurately predicting yield during the growing season enables improved crop management and better resource allocation for both breeders and growers. Existing yield prediction models for an entire field or individual plots are based on satellite-derived vegetation indices (VIs) and widely used machine learning-based feature extraction models, including principal component analysis (PCA) and autoencoders (AE). Here, we significantly enhance pre-harvest yield prediction at plot-scale using Compositional Autoencoders (CAE) - a deep-learning-based feature extraction approach designed to disentangle genotype (G) and environment (E) features - on high-resolution, plot-level satellite imagery. Our approach uses a dataset of approximately 4,000 satellite images collected from replicated plots of 84 hybrid maize varieties grown at five distinct locations across the U.S. Corn Belt. By deploying the CAE model, we improve the separation of genotype and environment effects, enabling more accurate incorporation of genotype-by-environment (GxE) interactions for downstream prediction tasks. Results show that the CAE-based features improve early-stage yield predictions by up to 10% compared to traditional autoencoder-based features and outperform vegetation indices (VIs) by 9% across various growth stages. The CAE model also excels in separating environmental factors, achieving a high silhouette score of 0.919, indicating effective clustering of environmental features. Moreover, the CAE consistently outperforms standard models in unseen environments and unseen genotypes yield predictions, demonstrating strong generalizability. This study demonstrates the value of disentangling G and E effects for providing more accurate and early yield predictions that support informed decision-making in precision agriculture and plant breeding.