Calibração do modelo de simulação de cultura DSSAT: uma abordagem através do otimizador bayesiano

Abstract

The present work sought to show that, through the application of the Bayesian Optimization (BO) technique, it was possible to efficiently and computationally calibrate the parameters of a crop simulation model such as DSSAT (Decision Support System for Agrotechnology Transfer), with BO being a statistical and Machine Learning technique that robustly identified DSSAT parameters through optimization. In this study, the calibration was focused on parameters of the phenological submodel CROPGRO-Drybean for the common bean BRS Esplendor cultivar, in order to evaluate and demonstrate the effectiveness of the technique in isolation, keeping all other parameters constant and applying BO only to the phenological parameters of DSSAT. The use of the BO technique was carried out in the R software through the ParBayesianOptimization (PBO) function, which sought to estimate a set of parameters whose objective was to minimize the Objective Function, which, in this study, was given by the MAPE (Mean Absolute Percentage Error). It was an appropriate technique when optimization was computationally expensive and, moreover, it stood out for its intelligent way of selecting candidate points that composed the optimization result, using an Acquisition Function, which was the subject of analysis in this work. The results therefore aimed to show the efficiency of BO-based calibration when compared with manual calibration performed by an agronomic expert and with the separate and simultaneous estimation approaches proposed in the Calibration Protocol for Soil-Crop Models. The results examined the boxplots of the outputs obtained by each methodology in order to identify the efficiency of the calibrations, and, in this way, Bayesian Optimization proved to be an adequate technique for estimating the phenological hyperparameters of the BRS Esplendor cultivar when compared with agronomic protocols and manual calibration, as the results were able to optimize the hyperparameter values by reducing the MAPE (Mean Absolute Percentage Error). This showed that the technique could be expanded to other groups beyond the phenology of common bean.