Estimating Soybean Yields Using Causal Inference and Deep Learning Approaches With Satellite Remote Sensing Data

Abstract

Timely and accurate crop yield estimation is crucial for managing crops, trade, and food security. The combination of remote sensing technology with machine learning methods is increasingly popular for global yield prediction. However, traditional machine learning methods rely on data correlation rather than causality, leading to poor interpretability. To address this issue, we propose a novel approach that combines a structural causal model (SCM) with deep learning to develop a causal graph attention network model (SCM-GAT) based on causal relationships for soybean yield prediction at the county level of 10 major soybean-producing states in the United States. The SCM-GAT model considers not only conventional vegetation indices and weather variables but also the causal relationships between variables as inputs. Using independent validation and five-fold cross-validation strategies, our results show that the predictive performance of the SCM-GAT model is not only superior to traditional prediction models, such as selection operator regression (LASSO), random forest, but also superior to the common deep learning models based on correlation relationships, such as long short-term memory network, and transformer. It has also proved to be more robust than other models under extreme weather events. In addition, we identify branching to blooming and blooming to setting pods as the key growth phases for soybean yield, and adding these phases as inputs to the model further improves prediction accuracy. Our findings suggest that incorporating causal relationships in deep learning models can improve prediction accuracy. This study provides a novel approach for remote sensing prediction of crop yield.

Keywords

• Causal inference
• deep learning
• graph attention network (GAT)
• soybean yield prediction
• structural causal model (SCM)