Knowledge Guided Machine Learning for Wheat Yield Downscaling Based on Global Gridded Crop Models

Precise and reliable projections of crop yields are necessary to maintain human food security. Global crop yields have been estimated by a set of Global Gridded Crop Models (GGCMs), but these model outputs are usually at coarse spatial resolutions, typically in half degrees. It's appropriate for studies at global scales, but not for regional and local ones where finer yield predictions are needed for robust assessments on agricultural impacts and decision making. High requirements for data availability and computational capacity make it hard to conduct high-resolution and large-scale simulations by process-based crop models. Traditional statistical downscaling has been widely used to increase spatial granularity but cannot capture nonlinear relationships. Machine learning as well as deep learning methods have strong potential for complex, nonlinear input-output modeling can help to bridge this gap. Compared to the black-box use of pure machine learning, knowledge-guided machine learning (KGML) shows a higher level of interpretability by incorporating knowledge from multi-sources. In this study, we put forward a KGML framework for yield downscaling, exploring two approaches - Random Forests (RF) and Long Short-Term Memory (LSTM) to develop models for the prediction of crop model outputs at fine spatial resolutions. Both algorithms are trained on global-scale simulations of a GGCM and then applied at a finer spatial resolution to obtain high resolution yield estimations. Knowledge generated from empirical observations are incorporated as additional constraints. Results show high accuracy with R2>0.9 for predictions of wheat yield in both cases using RF and LSTM while LSTM performs better on capturing the annual variations. Downscaled products provided more detailed spatial information of places with low yield and high climatic vulnerability. This framework of wheat yield downscaling can also be applied to other crops and has the potential to guide the local adaptations of policy makers.