Maize yield prediction with trait-missing data via bipartite graph neural network.

Area of Science:

• Agricultural Science
• Machine Learning
• Data Science

Background:

• Accurate maize (Zea mays L.) yield prediction is vital for food security and agricultural policy.
• Current machine learning and deep learning models face limitations in handling data correlations and missing/unbalanced samples.
• Existing methods often overlook inter-sample yield correlations and the combined influence of meteorological and maize traits.

Purpose of the Study:

• To propose an end-to-end bipartite graph neural network (BGNN) model for maize trait data imputation and yield prediction.
• To address limitations of existing models by incorporating correlations within planting data and feature interactions.
• To develop a robust model capable of handling missing data and unbalanced sample sizes without pre-processing.

Main Methods:

• Maize planting data was transformed into a bipartite graph structure.
• A BGNN model was developed to simultaneously impute missing trait data and predict maize yield.
• A gradient balancing mechanism was employed in the loss function to mitigate the impact of unbalanced sample sizes.

Main Results:

• The proposed BGNN model effectively mined correlations between samples, meteorological features, and traits.
• The model demonstrated superior performance in yield prediction compared to existing methods, even with pre-processing of missing data.
• The gradient balancing loss function successfully reduced the negative effects of data imbalance on prediction accuracy.

Conclusions:

• The end-to-end BGNN model offers a significant advancement in maize yield prediction by effectively handling data complexities.
• This approach provides a robust solution for accurate yield forecasting, crucial for agricultural planning and food security.
• The method's ability to perform imputation and prediction concurrently, without data pre-processing, highlights its practical utility.