Mixing process-based and data-driven approaches in yield prediction

Yield prediction models can be divided between data-driven and process-based (crop growth models). The first category contains many different types of with parameters learned from the data themselves where domain knowledge is only used to select predictors engineer features. In second category, are based upon biophysical principles, whose structure derived primarily knowledge. Here we investigate if integration two approaches beneficial as it allows overcome limitations taken individually - lack sufficiently large, reliable orthogonal datasets for need inputs models. applications categories have been reviewed, paying special attention cases mixed. By analysing literature identified three major approaches: (1) using crop features expand space, (2) use estimate missing (3) produce meta-models reduce computation burden. Finally propose a methodology on metamodels transfer learning integrate approaches.