Using pattern recognition for estimating cultivar coefficients of a crop simulation model

The introduction of a new cultivar in a process-based crop simulation model requires the estimation of cultivar coefficients that define its growth and development characteristics. An accurate estimation of these coefficients requires replicated field experiments that, in many cases, are not available to crop model users. The objective of this study was to employ a pattern recognition approach to estimate cultivar coefficients from a minimum set of experimental data for use with a crop simulation model. The pattern recognition approach is based on similarity measures. Its main goal is to classify groups of data or patterns based on either a priori knowledge or on statistical information extracted from the patterns. Based on the similarity measure as the central calculation of the pattern recognition approach, the algorithm searches the space of features of other cultivars in the database to find the most similar cultivar as the best match to the target cultivar. The approach of this study was based on a few key characteristics of maize crop growth and development, including anthesis and harvest maturity dates, maximum leaf area index (LAImax), final above ground biomass, and grain yield, which were used as the features vector. To construct the feature database, 27,789 hypothetical cultivars were constructed by combining different values of the six cultivar coefficients of the Cropping System Model (CSM)-CERES-Maize. Experiments performed in Florida (FL) and Iowa (IA) USA, Spain, central Punjab, Pakistan, and in Piracicaba, SP, Brazil were selected and later modified to provide a full potential production environment. The crop model was run for potential production for all 27,789 hypothetical cultivars and the outputs of these simulations were used as the feature database. For evaluation of this approach, we used the features for 29 different maize cultivars as reported from field experiments that are available in DSSAT maize cultivar database and also for four additional cultivars of which two had not been used in any aspect of this study. The model was run for all 33 cultivars, using the best match cultivar coefficients, for the conditions of the three study sites and locations where the latter four cultivars have been grown. The simulated crop characteristics were compared with the same simulated crop characteristics based on the original coefficients used to run the simulation model. We found that the approach based on pattern recognition was able to estimate the cultivar coefficients with a reasonable accuracy. The coefficient of determination (r ²), root mean square difference (RMSD), and relative root mean square of difference (RMSDr) confirmed that this approach provided reliable estimates for the maize cultivar coefficients. The highest R ² (0.98) was obtained for anthesis in Florida and the lowest (0.57) was obtained for grain yield in Spain. The highest RMSD (8.8) was obtained for maturity in Spain, while the lowest RMSD (1.1) was obtained for aboveground biomass in Florida. Although the values for RMSD were different across the different sites, this approach provided a level of accuracy that might be acceptable, especially for users who only have one year of experimental data and demand the best possible initial guess for the coefficients of their specific cultivar. This approach has been implemented in a simple tool that can be easily applied by users of DSSAT and the CSM-CERES-Maize model.