Ecological models offer a valuable tool for understanding complex ecological systems. However, inaccurate conclusions can be drawn from models with a large degree of uncertainty around multiple parameter estimates if uncertainty is ignored. We addressed this issue for a mechanism-based model of Populus fremontii (Fremont cottonwood) populations on the Sacramento River, California, USA. The population dynamics of this riparian species are driven by complex interactions between biological factors (e.g. physiological constraints) and the physical drivers of river channel morphology (e.g., channel meandering and cutoff, flow variability and sediment transport). These forces create essential habitat for cottonwood, but also cause mortality through scour and sediment deposition. Widespread flow regulation and floodplain conversion have altered natural flow regimes on the Sacramento River, resulting in changes in the amount and timing of water availability and ultimately in reduced populations of Fremont cottonwood. Management recommendations for cottonwood in this system involve the allocation of scarce and costly water resources. These issues are extremely controversial and there is a sense of urgency in developing accurate predictive models for large-scale decision making. However this task is challenging due to the scale and complexity of the system.
Results/Conclusions
To provide useful information on which to base management decisions, we used an iterative modeling approach that combines mechanistic modeling, global sensitivity analysis, research prioritization, model improvement and multi-scale model validation. This process allowed us to maximize the use of currently available data, efficiently improve the model and assess the spatial and temporal scales on which the model could make accurate predictions. Using a global sensitivity analysis (GSA), Random Forest and Classification and Regression Trees (CART), we found that uncertainty in estimates of three physical factors (capillary fringe height, floodplain accretion rate and the stage-discharge relationship) and interactions among these factors, accounted for a large degree of the variation in model predictions. Additional research on the spatial variability in these factors allowed us to improve the model by incorporating systematic longitudinal variation (i.e. distance from the mouth of the river) in the physical factors included in the model. To validate the improved model we took a multiscale approach that included validating specific components of the model, including the model’s mechanisms and decision rules. This is particularly important for mechanistic models and for models that will be used to make predictions under changing future conditions.
