Crop yield prediction via explainable AI and interpretable machine learning: Dangers of black box models for evaluating climate change impacts on crop yield

Authors: HU, T. - The Ohio State University; ZHAO, K. - The Ohio State University; ZHOU, Y. - Iowa State University; LIU, Y. - The Ohio State University; BOHRER, G. - The Ohio State University; MARTIN, J. - The Ohio State University; LI, Y. - The Ohio State University; Zhang, Xuesong

Submitted to: Agricultural and Forest Meteorology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/9/2023
Publication Date: 5/6/2023
Citation: Hu, T., Zhao, K., Zhou, Y., Liu, Y., Bohrer, G., Martin, J., Li, Y., Zhang, X. 2023. Crop yield prediction via explainable AI and interpretable machine learning: Dangers of black box models for evaluating climate change impacts on crop yield. Agricultural and Forest Meteorology. 336:109458. https://doi.org/10.1016/j.agrformet.2023.109458.
DOI: https://doi.org/10.1016/j.agrformet.2023.109458

Interpretive Summary: Understanding the relationships between climate and crop yield helps identify vulnerable regions and adopt appropriate technologies to increase resilience of agroecosystems in the face of future climate change. Here, we developed a Bayesian model (BM) framework to characterize yield responses to climate variables and applied the method in Ohio, US to understand historical responses of corn yield to climate factors (i.e., temperature and precipitation). We found that the new BM method can better explain the climate-yield relationships than two machine learning techniques (i.e., Neural Networks and Random Forests) and provide more accurate predictions than conventionally used multivariate linear regression methods. As such, the new BM method provides a new tool for studying climate change impacts on crop yield and informing future agricultural adaptation efforts to ensure food security.

Technical Abstract: Mitigating impacts of climate change on crop yield requires a model of how yield responds to weather conditions. Empirical crop models have been increasingly used for this purpose in recent years. But choosing an appropriate model or method remains challenging because we need a model to have both good predictive power in projecting yields changes and good explanatory power in capturing direct linkages between yield responses and climate variables so that we can make adaptations according to these responses. Here, we proposed a Bayesian model (BM) framework based on Bayesian learning for characterizing yield responses to climate variables. This framework can capture nonlinear relationships between crop yield and climate variables and also has high interpretability. We compared the predictive and explanatory power of BM with two machine learning models which are usually good at prediction but with low interpretability and an inherently interpretable regression-based model via two experiments. In the first experiment with synthetic data, our results show that BM outperforms the two machine learning methods--Neural Networks (NN) and Random Forests (RF)--in unveiling relationships between explanatory and response variables (i.e., interpretability). In the second experiment of decomposing the complex climate and crop yield relationships with historical maize yield and climate records, BM is comparable with NN and RF in prediction abilities, and all of the three models perform better than a multivariate linear regression (MLR) model. Overall, the proposed BM framework achieves a balance between predictive power and explanatory power, and could be a competitive alternative to both conventional regression methods and machine learning models in characterizing climate impacts on crop yield and informing adaptation strategies.