Pf-Sds: a Framework for Crop Yield Prediction Based on Deep Forest and Interpretability

Crop yield forecasting is central to maintaining the global food security system. It enables governments and relevant regulatory agencies to anticipate potential supply and demand fluctuations in advance and make countermeasures. Some researchers have used different methods to predict it, including traditional methods based on empirical formulas, statistical models, or expert systems, while many have adopted strategies based on deep learning. In the former research work, accuracy and adaptability are challenging to be guaranteed. The latter has the disadvantage of poor interpretation and opacity. In response to the above problems, We designed a crop yield forecasting framework with high accuracy and high interpretability of variable dependencies, dubbed PF-SDS. In the framework of PF-SDS, a proven and feasible Sequential Forward Selection (SFS) feature selection algorithm was introduced to enhance the model's generalization ability. In addition, we use SHAP values to conduct an in-depth explanatory analysis of the importance of model features, thus providing a scientific and easy to understand basis for crop planting management practices. Numerous experimental results show the excellent performance of the PF-SDS framework.