Prediction of Corn Variety Yield with Attribute-Missing Data Via Graph Neural Network.

The crop variety yield prediction is widely used to select new varieties and select suitable planting areas for them, but it still suffers from multiple grand challenges, including sparse data and complex interaction mechanisms between the variety and different environments. Moreover, accurately predicting the yield of corn varieties often faces the problem of missing trait data, making it even more difficult. This study reports a corn variety yield prediction model based on generative adversarial network and graph attention network (YPM-GAN-GAT) to fill in the missing trait attributes and achieve corn variety yield prediction by establishing spatial and temporal correlations for multi-environment testing samples. Specifically, we first construct a variety spatio-temporal graph to utilize the structural features of the graph to effectively and fully express information such as relationships between different corn varieties planting data from different locations. Then, we make a shared-latent space assumption between the structural features of the graph and the attribute features of each node in the graph. This helps in developing a distribution matching based on graph neural network (GNN) and adversarial strategy to generate new data to fill in the missing attributes. Finally, we predict the corn variety yield based on the filled node attributes and two different types of GNN models. The experimental results show that the proposed missing trait imputation method outperforms the other five common data imputation methods, and the proposed GNN-based corn yield prediction model has better performance as compared to the other five machine learning-based models and two deep learning-based models. This study demonstrates a novel solution for improving the accuracy of large-scale corn variety yield prediction.