Model improvement via data assimilation toward ecological forecasting.

The field of ecology has changed to a data-rich enterprise, largely due to rapid development of measurement sensors and long-term accumulation of data from research networks. With implementation of the National Ecological Observatory Network (NEON), a network with different kinds of sensors and measurements at many locations over the United States, large volumes of ecological data will be generated every day. Thus, there will be an unprecedented demand to convert massive amounts of raw data into ecologically meaningful products using data assimilation (DA) techniques. DA is a tool that combines observational data with ecological models to provide reliable ecological forecasting. It can help improve model parameterization, inform choices between alternative model structures, design better sensor networks and experiments for data collection, and analyze uncertainty of ecological forecasts. Forecasting generally means a capability to project future states of a system by modeling the evolution of the system as a function of its state at an initial time. DA is an essential tool to constrain parameters and the initial system state before a model is used to project the evolution of the system. Ecology as a research community has recently explored, examined, and developed a variety of DA techniques to analyze multi-scale ecological data in space and time. To prepare the community for major research challenges in a data-rich era, we organized eight papers in this Invited Feature to address several issues on DA and ecological forecasting. The paper by Luo et al. offers a perspective on DA, ecological forecasting, and their relationships. Ecological forecasting (or prediction or projection) has been traditionally made using process-oriented models, informed by data in largely ad hoc ways. Most of the simulation models are generally not adequate to quantify real-world dynamics so as to provide reliable forecasts. DA uses data to constrain initial conditions and model parameters to yield simulations that approximate reality as closely as possible. Models conditioned upon the best information from both process-oriented thinking and empirical knowledge should generate the best forecasts. However, forecasts cannot be improved by DA when ecological processes are not well understood or never observed. LaDeau et al. use four case studies (Severe Acute Respiratory Syndrome [SARS], Dengue Fever, Lyme, and West Nile virus) to demonstrate that advances in disease forecasting require better understanding of the zoonotic host and vector ecology that support pathogen amplification and disease spillover into humans. The authors identify cases where additional data, greater biological understanding, and coherent treatment of spatiotemporal variability could substantially improve forecasts of disease dynamics. To do so, DA is required in a hierarchical state-space framework to (1) integrate multiple data sources across spatial scales to estimate latent parameters, (2) partition uncertainty in process and observation models, and (3) explicitly build upon existing ecological and epidemiological understanding. The Ensemble Kalman Filter (EnKF) has been successfully used in weather forecasting to assimilate observations into models. Gao et al. examine how effectively EnKF can improve forecasts of carbon sequestration. Eight data sets from Duke Forest between 1996 and 2004 were assimilated into a terrestrial ecosystem model, which was then used to forecast changes in carbon pools from 2004 to 2012. Parameter uncertainties decreased as data were sequentially assimilated into the model. Uncertainties in forecast carbon sinks increased over time for the long-term carbon pools but remained stationary for the short-term pools. EnKF effectively assimilated multiple data sets to constrain parameters, forecast dynamics of state variables, and evaluate uncertainty. Hill et al. assimilated data from both tower flux and atmospheric profile measurements into a coupled atmosphere–biosphere model to constrain ecosystem processes. Variations in temperature,