Abstract
Accurately predicting crop yields is a critical challenge in sustainable agriculture, food security, and farm management. Traditional process-based models rely on agronomic domain knowledge, crop physiology and statistical approaches, while purely data-driven approaches leverage machine learning or deep learning models using meteorological and spatial data. Unfortunately, these black-box models(Data-drive approaches) often lack interpretability and fail to incorporate well-established physical principles. This project explores a hybrid approach by implementing Physics Informed Neural Networks, mainly, physics-based recurrent neural networks (PI-RNNs) for time-series yield prediction. PINNs allow for the integration of scientific knowledge directly into the model by embedding physical laws as constraints in the loss function. This enables the network to learn from both data and domain-specific rules, improving generalizability and interpretability. In this study, we apply PINNs to predict winter wheat yield in the states of Kansas and Oklahoma. This is done by utilizing meteorological data, and remote sensing input taken from the Cybench dataset and AgERA-5 datasets. We attempt to mainly utilize Harvest Index, and Radiation use efficiency and constrain them in physical ranges obtained by previous domain knowledge. Furthermore, we penalize the machine learning model based on both a data loss and a physics based loss, which is calculated by the distance between the constrained variables and its set range decided based on domain knowledge. By combining these insights with machine learning, this research aims to apply a more reliable, interpretable, and scientifically grounded yield prediction framework. Future work in this area would involve trying to apply similar techniques to more complex models, including time-series transformers or applying established Differential Equations to ML algorithms and utilizing a similar loss scheme.
Keywords
Physics Informed Neural Networks, Crop Yield Prediction, Meteorological and Spatial Data, Explainability and Interpretability, RNNs
Date of this Version
2025
Recommended Citation
Prabhu, Aditya P.; Poudel, Pratishtha; and Krogmeier, James V., "Application of Physics-Informed Neural Networks on crop yield prediction at multiple scales" (2025). Discovery Undergraduate Interdisciplinary Research Internship. Paper 70.
https://docs.lib.purdue.edu/duri/70
Included in
Agricultural and Resource Economics Commons, Artificial Intelligence and Robotics Commons, Data Science Commons