Machine learning for large-scale crop yield forecasting

Many studies have applied machine learning to crop yield prediction with a focus on specific case studies. The data and methods they used may not be transferable other crops locations. On the hand, operational large-scale systems, such as European Commission's MARS Crop Yield Forecasting System (MCYFS), do use learning. Machine is promising method especially when large amounts of are being collected published. We combined agronomic principles modeling build baseline for forecasting. workflow emphasizing correctness, modularity reusability. For we focused designing explainable predictors or features (in relation growth development) applying without information leakage. created using simulation outputs weather, remote sensing soil from MCYFS database. emphasized modular reusable support different countries small configuration changes. can run repeatable experiments (e.g. early season end predictions) standard input obtain reproducible results. results serve starting point further optimizations. In our studies, predicted at regional level five (soft wheat, spring barley, sunflower, sugar beet, potatoes) three (the Netherlands (NL), Germany (DE), France (FR)). compared performance simple no skill, which either linear trend average training set. also aggregated predictions national past forecasts. normalized RMSE (NRMSE) (30 days after planting) were comparable NL (all crops), DE except soft wheat) FR sunflower). example, NRMSE was 7.87 wheat (NL) (6.32 MCYFS) 8.21 beet (DE) (8.79 MCYFS). contrast, NRMSEs (FR) potatoes twice much MCYFS. still NL, but worse FR. improved by adding new sources, more predictive evaluating algorithms. will motivate in