Assimilating leaf area index data into a sugarcane process-based crop model for improving yield estimation

Abstract

The ability to estimate sugarcane yield is an important factor to improving the planning capacity of public and private sectors, and so food and energy security. One way of achieving this is by employing process-based crop models (PBM), which can be coupled to data assimilation (DA) algorithms to correct predictions along the crop season. While the application of PBMs often need careful parameterization or genotype-specific parameters, few studies focus on understanding the impacts of crop parametrization with different crop genotypes with DA. Moreover, dimensioning the number and timing of observations is key to effectively improve predictions with DA. This study assess the performance of a new sugarcane PBM (DSSAT/SAMUCA) coupled to three DA methods, and when the genotype-specific parameters are available or not. Data from 22 field experiments is utilized to compare the performance of using the ensemble Kalman filter (EnKF), ensemble smoother (ES) and weighted mean (WM) for assimilating leaf area index (LAI) to improve yields estimates. We also quantify the impact of using one genotype-specific calibration (cv. RB867515) on yield predictions of four non-calibrated genotypes (cv. NCo376, SP832847, R570, RB72454). Simulations of DA methods had better performance than employing the PBM without DA, so called open-loop (OP). The ES method resulted in the best performance (R² = 0.498 and RMSE = 20.268 Mg ha−1) followed by EnKF and WM. Utilizing a genotype-specific calibration showed substantially smaller RMSE for the three DA methods (EnKF = 16.76, ES = 16.70 and WM = 15.36 Mg ha−1) compared to non-calibrated (EnKF = 21.44–26.23, ES = 21.50–26.27 and WM = 23.38–28.37 Mg ha−1). Nevertheless, we also verified a higher improvement of model performance when applying EnKF and ES method to experiments where the cultivar does not match the genotype-specific calibration employed. While the WM had the opposite results, with the calibrated cultivar showing a higher improvement of model performance. As the number of LAI data assimilation increases, the DA methods tend to outperform the OP, but observations at late crop phenological stage of development showed a higher positive influence on SFY predictions.