In order to describe underlying biophysical mechanisms, process-based plant growth models often contain an excessive number of parameters when only considering the accuracy of the model outputs. Regarding their influence on the model output, parameters of complex nonlinear plant growth models interact in ways that cannot be easily predicted based upon their roles in the component submodels describing biophysical processes. Parameter estimation in plant growth models is often challenged by lack of a means to interpret the relative importance of parameters in complex nonlinear models. In multi-crop models such as for agroforestry, increased model complexity and lack of data for novel crop combinations in varying environments further exacerbate the difficulties of discerning which parameters are important for estimation. The approach is based upon foundational system identification theory applied to a class of deterministic process-based predictive growth models. Evaluating the Hessian of the quadratic cost function determines the relative importance of parameters to its curvature. Subject to a list of model requirements, the Hessian can be computed given the input-output data and an estimated location in the parameter space provided by research into underlying biophysical processes and expert knowledge. For this analysis, field data are not required, rather, reliable simulated climate data are used to drive the model, which itself provides a priori output data for analysis. The analysis method is presented as a procedure for determining a ranking of parameter importance that can be used by model developers to provide end users with guidance for parameter estimation given real data for novel crops and crop combinations. The goal of this procedure is to arrive at a reduced-order parameter space that can be estimated entirely from input-output data. Furthermore, the goal is for the reduced model to closely follow the outputs of the original system (with any feasible parameterization) when driven by any input in the input class. The procedure is demonstrated on the well-known Yield-SAFE predictive agroforestry growth model. The advantages of an input-output system identification approach may also carry over into field trial design or model structure revisions. Further, because model parameterization can be based only on readily accessible model outputs, relatively low-tech data collection strategies emphasizing on-farm participatory research become possible. Participatory approaches allow a broader range of useful data to be collected for evaluating complex crop combinations.