A comparison of empirical and mechanistic models for wheat yield prediction at field level in Moroccan rainfed areas

Abstract

Abstract In the context of climate change, in-season and longer-term yield predictions are needed to anticipate local and regional food crises and propose adaptations to farmers’ practices. Mechanistic models and machine learning are two modelling options to consider from this perspective. In this study, multiple regression (MR) and random forest (RF) models were calibrated for wheat yield prediction in Morocco, using data collected from 125 farmers’ wheat fields. Additionally, MR and RF models were calibrated both with or without remotely sensed leaf area index (LAI), while considering all farmers’ fields, or specifically to agroecological zoning in Morocco. The same farmers’ fields were simulated using a mechanistic model (APSIM-wheat). We compared the predictive performances of the empirical models and APSIM-wheat. Results showed that both MR and RF showed rather good predictive quality (normalized root mean square errors (NRMSEs) below 35 %), but were always outperformed by the APSIM model. Both RF and MR selected remotely sensed LAI at heading, climate variables (maximal temperatures at emergence and tillering), and fertilization practices (amount of nitrogen applied at heading) as major yield predictors. Integration of remotely sensed LAI in the calibration process reduced NRMSE by 4.5 % and 1.8 % on average for MR and RF models, respectively. Calibration of region-specific models did not significantly improve the predictive. These findings lead to the conclusion that mechanistic models are better at capturing the impacts of in-season climate variability and would be preferred to support short-term tactical adjustments to farmers’ practices, while machine learning models are easier to use in the perspective of mid-term regional prediction.