Combining remote sensing and data-driven models for early season wheat forecasts in Morocco: How transferable and early could a yield predictive model be?

Strong interannual variability in precipitation amounts and distribution, as well as recurrent droughts, are cornerstones of African countries. These phenomena primarily impact rainfed crops, of which wheat is the most important and accounting for more than 80% of cultivated areas in Morocco. An early and consistent projection of pre-harvest grain production would help decision-makers anticipate management demands, detect yield gaps, and better understand wheat response to local climatic circumstances. How early a prediction is needed and the required depend on the nature of the stakeholder. In other words, early in-season forecasts are useful for producers so that they can adjust their inputs accordingly, whereas late-season forecasts are acceptable for other stakeholders, for example those interested in production monitoring.

In this work, we used satellite-derived phenology measures, climate, and soil data to generate in-season yield prediction models for rainfed and irrigated wheat in Morocco. The primary aims were to evaluate the predictive capabilities of the models as time progresses and the transferability of the models outside the area of their implementation. The findings demonstrated that the generated models' accuracy increases over time (i.e., when additional phenological measures are integrated into the models) and that Ensemble models and Random Forest models outperformed the conventional MLR models, including the regularised regression models (Lasso, Ridge, ElasticNet).