A FASTER ALGORITHM FOR CROP MODEL PARAMETERIZATION BY INVERSE MODELING: SIMULATED ANNEALING WITH DATA REUSE

摘要

Some recent studies contend that a simple exhaustive (grid) search through the complete inverse modeling parameter space can be faster than a more sophisticated algorithm such as simulated annealing (SA). The objective of this study was to implement an SA algorithm guaranteed to run worst-case as fast as the grid search. This algorithm resulted from: (1) working on a discrete parameter space similarly to the grid search, and (2) accelerating the optimization by avoiding multiple crop model runs of the same parameter combinations. The latter was achieved by storing simulation results in computer memory and checking for those data before running the model for any given parameter combination. The algorithm was tested relative to a pure grid search in two case studies: a synthetic problem having numerous local minima, and a crop modeling exercise using the CERES-Maize model with two years of observed maize data from 13 locations in a field near Murray, Kentucky. The performance of the modified SA algorithm depended on its parameter values, but a conservative parameter combination was found that ran much faster than the grid search, its run time tending asymptotically to that of the grid search as the number of locations to calibrate grew, while converging to objective function values (and usually parameter values) practically identical to the global optimum, and explaining 97% of the observed yield in the two years used for parameterization. Adoption of this algorithm can reduce parameterization run time on the order of 25% to 75%, depending on the geometry of the simulation domain. Additionally, it can be used to calibrate coupled spatial crop models in which parameter values at one location can affect parameter values at other locations, a task not possible using a grid search. Finally, the proposed algorithm can generate good “first cut” solutions in only a few hundred iterations, an important consideration when designing practical crop simulation applications