An assessment of empirical models, structure, predictor variables, and performances for wheat yield prediction at field level in Moroccan rainfed areas.

Relationship between the performances of yield prediction models, complexity of their structures and the number and type of required input was always a recognized problematic by researchers. In the present study, we compared extracted empirical models with a previous calibrated and evaluated mechanistic model (APSIM-wheat) in yield prediction at field scale, by highlighting empirical models structure, predictor variables, their sustainability and the timely scope of yield prediction, and assessing the impact of integrating satellite-based dataset on models accuracy. We conducted a modelling framework for wheat yield prediction in Moroccan rainfed areas basing on two methods: multiple regression (MR) and random forest (RF) algorithms, and using input parameters database combine soil, climate, remotely-sensed LAI and crop management variables that were collected over three successive crop seasons (2018-2021) from 130 farmers¿ wheat fields located in Moroccan rainfed areas. Results show the relevance of remotely-sensed LAI-Z50, nitrogen fertilization and climate variables as predictors of yield. Almost identical wheat yield estimation performances using both empirical methods with RMSE < 0.9 t.ha-1 in most cases, whereas, APSIM-wheat has the highest potential in predicting wheat yield at field scale. Also, clear progresses were observed in models robustness when integrating LAI satellite-based parameters during empirical models development.