Introduction
One ambitious goal of ecology as a science discipline is to forecast states
and services of ecological systems. Forecasting in ecology is not only
desirable for scientific advances in this discipline but also has practical
values to guide resource management and decision-making towards a sustainable
planet Earth. The practical need for ecological forecasting is particularly
urgent in this rapidly changing world, which is experiencing unprecedented
natural resource depletion, increasing food demand, serious biodiversity
crisis, accelerated climate changes, and widespread pollution in the air,
waters, and soils (Clark et al., 2001; Mouquet et al., 2015). As a result, a
growing number of studies have reported forecasting of, e.g., phenology (Diez
et al., 2012), carbon dynamics (Luo et al., 2016; Gao et al., 2011; Thomas et
al., 2017), species dynamics (Clark et al., 2003; Kearney et al., 2010),
pollinator performance (Corbet et al., 1995), epidemics (Ong et al., 2010),
fishery (Hare et al., 2010), algal bloom (Stumpf et al., 2009), crop yield
(Bastiaanssen and Ali, 2003), biodiversity (Botkin et al., 2007), plant
extinction risk (Fordham et al., 2012), and ecosystem service (Craft et al.,
2009) in the last several decades. Despite its broad applications, ecological
forecasting is still sporadically practiced and lags far behind demand due to
the lack of infrastructure that enables timely integration of models with
data. This paper introduces a fully interactive infrastructure, the
Ecological Platform for Assimilating Data
(EcoPAD, v1.0) into models, to inform near-time ecological forecasting with
iterative data–model integration.
Ecological forecasting relies on both models and data. However, currently
the ecology research community has not yet adequately integrated
observations with models to best inform forecasts. Forecasts generated from
scenario approaches are qualitative and scenarios are often not based on
ecological knowledge (Coreau et al., 2009, 2010).
Data-driven forecasts using statistical methods are generally limited for
extrapolation and sometimes contaminated by confounding factors
(Schindler and Hilborn, 2015). The recent emergent mechanism-free
nonparametric approach, which depends on the statistical pattern extracted
from data, is reported to be promising for short-term forecast (Ward et
al., 2014; Perretti et al., 2013; Sugihara et al., 2012) but has limited
capability in long-term prediction due to the lack of relevant ecological
mechanisms. Process-based models provide capacity in long-term
prediction and flexibility in capturing short-term dynamics on the basis
of a mechanistic understanding (Coreau et al., 2009; Purves et al., 2013).
Wide applications of process-based models are limited by their often
complicated numerical structure and sometimes unrealistic parameterization
(Moorcroft, 2006). The complex and uncertain nature of ecology
precludes the practice of incorporating as many processes as possible into
mechanistic models. Our current incomplete knowledge about ecological
systems and unrepresented processes under novel conditions is partly
reflected in model parameters associated with large uncertainties.
Good forecasting therefore requires effective communication between
process-based models and data to estimate realistic model parameters and
capture context-dependent ecological phenomena.
Data–model fusion, or data–model integration, is an important step to combine
models with data. But previous data–model integration activities have mostly
been done in an ad hoc manner instead of being interactive. For
example, data from a network of eddy covariance flux tower sites across
the United States and Canada were compared with gross primary productivity (GPP)
estimated from different models (Schaefer et al., 2012). Luo and
Reynolds (1999) used a model to examine ecosystem responses to gradual, as in
the real world, vs. step increases in CO2 concentration as in
elevated CO2 experiments. Parton et al. (2007) parameterized
CO2 impacts in an ecosystem model with data from a CO2
experiment in Colorado. Such model–experiment interactions encounter a few
issues: (1) models are not always calibrated for individual sites and
therefore not accurate; (2) it is not very effective because it is usually
a one-time practice without many iterative processes between experimenters and
modelers (Dietze et al., 2013; Lebauer et al., 2013); (3) it is usually
unidirectional as data are normally used to train models, while the guidance of
the model for efficient data collection is limited; and (4) it is not streamlined
and could not be disseminated with common practices among the research
community (Lebauer et al., 2013; Dietze et al., 2013; Walker et al., 2014).
A few research groups have developed data assimilation systems to facilitate
data–model integration in a systematic way. For example, data–model
integration systems, such as the Data Assimilation Research Testbed – DART
(Anderson et al., 2009) – and the Carbon Cycle Data Assimilation
Systems – CCDAS (Scholze et al., 2007; Peylin et al., 2016), combine various
data streams (e.g., FLUXNET data, satellite data, and inventory data) with
process-based models through data assimilation algorithms such as the Kalman
filter (Anderson et al., 2009) and variational methods (Peylin et al., 2016).
These data assimilation systems automate model parameterization and provide
an avenue to systematically improve models by combining as much data as
possible. Data-informed model improvements normally happen after the end
of a field experiment, and interactive data–model integration is limited
as feedbacks from models to ongoing experimental studies are not adequately
realized. In addition, wide applications of these data assimilation systems
in ecological forecasting are constrained by limited user interactions with
a steep learning curve to understand these systems, especially for
experimenters who have limited training in modeling.
Web-based technology facilitates interactions. Web-based modeling, which
provides user-friendly interfaces to run models in the background, is usually
supported by the scientific workflow, a sequence of processes through which
a piece of work passes from initiation to completion. For example,
TreeWatch.Net (https://treewatch.net, last access: January 2019)
has recently been developed to make use of high-precision
individual tree-monitoring data to parameterize process-based tree models in
real time and to assess instant tree hydraulics and carbon status with online
result visualization (Steppe et al., 2016). Although the web portal of
TreeWatch.Net is currently limited to the purpose of visualization, it
broadens the application of data–model integration and strengthens the
interaction between modeling research and the general public. The
Predictive Ecosystem Analyzer (PEcAn) is a scientific workflow that wraps
around different ecosystem models and manages the flows of information coming
in and out of the model (Lebauer et al., 2013). PEcAn enables web-based model
simulations. Such a workflow has advantages, for example making ecological
modeling and analysis convenient, transparent, reproducible, and adaptable to
new questions (Lebauer et al., 2013) and encouraging user–model
interactions. PEcAn uses the Bayesian meta-analysis to synthesize plant trait
data to estimate model parameters and associated uncertainties, i.e., the
prior information for process-based models. Parameter uncertainties are
propagated to model uncertainties and displayed as outputs. It is still not
fully interactive in that states are not updated iteratively
according to observations and the web-based data assimilation and
ecological forecasting have not yet been fully realized.
Iterative model–data integration provides an approach to constantly
improve ecological forecasting and is an important step, especially in
realizing near-real-time ecological forecasting. Instead of projecting
into the future by assimilating observations only once, iterative
forecasting constantly updates forecasting along with ongoing new data
streams and/or improved models. Forecasting is likely to be improved
unidirectionally so that either only models are updated through
observations, or only data collections and field experimentations are improved
according to theoretical and model information, but not both. Ecological
forecasting can also be bidirectionally improved so that both models and
field experimentations are optimized hand-in-hand over time. Although the
bidirectional case is rare in ecological forecasting, unidirectional
iterative forecasting has been reported. One excellent example of
forecasting by dynamically and repeatedly integrating data with models
is from infectious disease studies (Niu et al., 2014; Ong et al., 2010).
The dynamics of infectious diseases are traditionally captured by
susceptible–infected–removed (SIR) models. In the forecasting of the
Singapore H1N1-2009 infections, SIR model parameters and the number of
individuals in each state were updated daily, combining data renewed from
local clinical reports. The evolution of epidemic-related parameters and
states was captured by iteratively assimilating observations to inform
forecasting. As a result, the model correctly forecasted the timing of the
peak and decline of the infection ahead of time. Iterative forecasting
dynamically integrates data with models and makes the best use of both data and
theoretical understandings of ecological processes.
The aim of this paper is to present a fully interactive platform, the web-based
EcoPAD (v1.0), to best inform ecological forecasting. The
interactive feature of EcoPAD (v1.0) is reflected in iterative model
updating and forecasting by dynamically integrating models with new
observations, bidirectional feedbacks between experimenters and modelers,
and flexible user–model communication through web-based simulation, data
assimilation, and forecasting. Such an interactive platform provides the
infrastructure to effectively integrate available resources from models
and data, modelers and experimenters, and scientists and the general public to
improve scientific understanding of ecological processes, boost ecological
forecasting practice, and transform ecology towards quantitative forecasting.
In the following sections, we first describe the system design and major
components of EcoPAD (v1.0). We then use the Spruce and Peatland Responses
Under Climatic and Environmental change (SPRUCE) experiment
(Hanson et al., 2017) as a test bed to elaborate
the functionality and new opportunities brought by the platform. We finally
discuss the implications of EcoPAD (v1.0) for better ecological forecasting.
Discussion
The necessity of interactive infrastructure to realize ecological
forecasting
Interactions enable the exchange and extension of information to benefit
from collective knowledge. For example, manipulative studies will have a
much broader impact if the implications of their results can be extended
from the regression between environmental variables and ecosystem response,
such as being integrated into an ecosystem model through model–data
communication. Such an approach will allow us to gain information about the
processes responsible for ecosystem response, constraining models, and
making more reliable predictions. Going beyond the common practice of model–data
assimilation, in which model updating lags far behind observations, EcoPAD
(v1.0) enables iterative model updating and forecasting by dynamically
integrating models with new observations in near-real time. This near-real-time interactive capacity relies on its scientific workflow that
automates data management, model simulation, data simulation, and result
visualization. The system design encourages thorough interactions between
experimenters and modelers. Forecasting results from SPRUCE were
shared among research groups with different backgrounds through the web
interface. Expertise from different research groups was integrated to
improve a second round of forecasting. Again, thanks to the workflow, new
information or adjustments are incorporated into forecasting efficiently,
making the forecasting system fully interactive.
We also benefit from the interactive EcoPAD (v1.0) platform to broaden
user–model interactions and to broadcast forecasting results. Learning about
ecosystem models and data–model fusion techniques may undermine one's
productivity and even discourage researchers from learning modeling techniques because of
their complexity and long learning curve. Because EcoPAD (v1.0) can be
accessed from a web browser and does not require any coding from the user's
side, the time lag between learning the model structure and obtaining
model-based results for one's study is minimal, which opens the door for
non-modeler groups to “talk” with models. The online storage of one's
results lowers the risk of data loss. The results of each model run can be
easily tracked and shared with a unique ID and web address. In addition,
the web-based workflow also saves time for experts with automated model
running, data assimilation, forecasting, structured result access, and
instantaneous graphic outputs, bringing the possibility for a thorough
exploration of more essential parts of the system. The simplicity in use of
EcoPAD (v1.0) at the same time may limit access to the code and lower
flexibility. Flexibility for users with higher demands, for example
those who want to test alternative data assimilation methods, use a
different carbon cycle model, change the number of calibrated parameters,
or include observations for other variables, is provided through the GitHub
repository (https://github.com/ou-ecolab, last
access: January 2019). This GitHub repository contains code and instructions for
installing, configuring, and controlling the whole system; users can adapt the
workflow to wrap their own model based on their needs. On the one hand, the
open-source web-based system broadens the user community. On the other
hand, it increases the risk of misuse and misinterpretation. We encourage
users to be critical and consult system developers to avoid inappropriate
application of the system.
Implications for better ecological forecasting
Specific to the reliable forecasting of carbon dynamics, our initial
exploration from EcoPAD-SPRUCE indicates that realistic model structure,
correct parameterization, and accurate external environmental conditions are
essential. The model structure captures important mechanisms that regulate
ecosystem carbon dynamics. Adjustment in the model structure is critical to our
improvement of methane forecasting. Model parameters may vary between
observation sites and change with time or environmental conditions (Medlyn et
al., 1999; Luo et al., 2001). A static or wrong parameterization misses
important mechanisms (e.g., acclimation and adaptation) that regulate future
carbon dynamics. Parameters that are not well constrained, for example caused by a lack
of information from observational data, contribute to high forecasting
uncertainty and low reliability in forecasting results. Correct
parameterization is especially important for long-term carbon pool
predictions as parameter uncertainty resulted in high forecasting uncertainty
in our case study (Jiang et al., 2018). Parameter values derived under
ambient conditions were not applicable to the warming treatment in our methane
case due to acclimation. The external environmental condition is another
important factor in carbon predictions. The external environmental condition
includes both the external climatic forcing that is used to drive ecosystem
models and also the environmental condition that is simulated by ecosystem
models. As we showed that air warming may not proportionally transfer to soil
warming, realistic soil environmental information needs to be appropriately
represented to predict soil carbon dynamics (Huang et al., 2017). The impact
of external forcing is especially obvious in short-term carbon flux
predictions. Forcing uncertainty resulted in higher forecasting uncertainty
in carbon flux compared to that from parameter uncertainty (Jiang et al.,
2018). Mismatches in forecasted vs. realized forcing greatly increased
simulated GPP and the discrepancy diminished in the long run. A reliable
external environmental condition, to some extent, reduces the complexity in
diagnosing modeled carbon dynamics.
Pool-based vs. flux-based predictions are regulated differently by external
forcing and initial states, which indicates that differentiated efforts are
required to improve short- vs. long-term predictions. External forcing, which
has not been well emphasized in previous carbon studies, has a strong impact on
short-term forecasting. The large response of GPP to forecasted vs. realized
forcing, as well as the stronger forcing-caused uncertainty in GPP
predictions,
indicates that correct forcing information is a key step in short-term flux
predictions. In this study, we stochastically generated the climate forcing
based on local climatic conditions (1961–2014), which is not sufficient to
capture local short-term climate variability. As a result, updated GPP went
outside our ensemble forecasting. On the other hand, parameters and
historical information about pool status are more important in long-term pool
predictions. Therefore, improvement in long-term pool size predictions cannot
be achieved with accurate climatic information alone. Instead, it requires
accumulation in knowledge related to site history and processes that regulate
pool dynamics.
Furthermore, reliable forecasting requires an understanding of uncertainty sources
in addition to future mean states. Uncertainty and complexity are major
factors that lead to the belief in “computationally irreducible” and low
intrinsic predictability of ecological systems (Beckage et al.,
2011; Coreau et al., 2010; Schindler and Hilborn, 2015). Recent advances in
computational statistical methods offer a way to formally account for
various uncertainty sources in ecology (Clark et al., 2001; Cressie et
al., 2009). The Bayesian approach embedded in EcoPAD (v1.0) brings the
opportunity to understand and communicate forecasting uncertainty. Our case
study revealed that forcing uncertainty is more important in flux-based
predictions, while parameter uncertainty is more critical in pool-based
predictions. Actually, how forecasting uncertainty changes with time, what
the dominate contributors of forecasting uncertainty are (e.g., parameter,
initial condition, model structure, observation errors, forcing, etc.), how
uncertainty sources interact among different components, and to what extent
unconstrained parameters affect forecasting uncertainty are all valuable
questions that can be explored through EcoPAD (v1.0).
Applications of EcoPAD (v1.0) to manipulative experiments and
observation sites
Broadly speaking, data–model integration stands to increase the overall
precision and accuracy of model-based experimentation (Luo et al.,
2011b; Niu et al., 2014). Systems for which data have been collected in the
field and that are well represented by ecological models therefore have the
capacity to receive the highest benefits from EcoPAD (v1.0) to improve
forecasts. In a global change context, experimental manipulations, including
ecosystem responses to changes in precipitation regimes, carbon dioxide
concentrations, temperatures, season lengths, and species compositional
shifts, can now be assimilated into ecosystem models (Shi et al., 2016; Xu
et al., 2006; Gao et al., 2011; Lebauer et al., 2013). The impacts of these global
change factors on carbon cycling and ecosystem functioning can now be
measured in a scientifically transparent and verifiable manner. This leads
to ecosystem modeling of systems and processes that can obtain levels of
confidence that lend credibility (from the public point of view) to the forward
progress of science toward forecasting and predicting
(Clark et al., 2001). These are the
strengths of a widely available interface devoted to data–model integration
towards better forecasting.
The data–model integration framework of EcoPAD (v1.0) creates a smart
interactive model–experiment (ModEx) system. ModEx has the capacity to form a
feedback loop in which field experiments guide modeling and modeling
influences experimental focus (Luo et al., 2011a). We demonstrated how EcoPAD
(v1.0) works hand-in-hand between modelers and experimenters in the
life cycle of the SPRUCE project. The EcoPAD-SPRUCE system operates while
experimenters are making measurements or planning for future research.
Information is constantly fed back between modelers and experimenters, and
simultaneous efforts from both parties illustrate how communication between
models and data advance and shape our understanding towards better forecasts
during the life cycle of a scientific project. ModEx can be extended to other
experimental systems to do the following: (1) predict what an ecosystem's response might be to
treatments once the experimenter has selected a site and decided the experimental
plan; (2) assimilate the data experimenters are collecting during the experiment
to constrain model predictions; (3) project what an ecosystem's responses may
likely be in the rest of the experiment; (4) tell experimenters which
important datasets they may want to collect in order to
understand the system; (5) periodically update the projections; and
(6) improve the models, the data assimilation system, and field experiments
during the process.
In addition to manipulative experiments, the data assimilation system of
EcoPAD (v1.0) can be used for automated model calibration for FLUXNET sites
or other observation networks, such as NEON and LTER (Johnson et al.,
2010; Robertson et al., 2012). The application of EcoPAD (v1.0) at FLUXNET,
NEON, or LTER sites includes three steps in general. First, build the climate
forcing in the suitable formats of EcoPAD (v1.0) from the database of each
site. Second, collect the prior information (include observations of state
variables) in the data assimilation system from FLUXNET, NEON, or LTER sites.
Third, incorporate the forcing and prior information into EcoPAD (v1.0) and
then run EcoPAD (v1.0) with the dynamic data assimilation system.
Furthermore, the proposed continental-scale ecology study (Schimel,
2011), EcoPAD (v1.0) once properly applied could also help evaluate and
optimize the field deployment of environmental sensors and supporting
cyber-infrastructure, which will be necessary for larger, more complex
environmental observing systems being planned in the US and across different
continents.
Future developments
EcoPAD (v1.0) will expand as time goes on. The system is designed to
incorporate multiple process-based models, diverse data assimilation
techniques, and various ecological state variables for different ecosystems.
The case studies presented in earlier sections are based primarily on one model.
A multiple (or ensemble) model approach is helpful in tracking uncertainty
sources from our process understanding. With rapidly evolving ecological
knowledge, emerging models with different hypotheses, such as the
microbial–enzyme model (Wieder et al., 2013), enhance our
capacity in ecological prediction but can also benefit from rapid tests
against data if incorporated into EcoPAD (v1.0). In addition to MCMC
(Braswell et al., 2005; Xu et al., 2006), a variety of data assimilation
techniques have been recently applied to improve models for ecological
forecasting, such as the EnKF (Gao et al., 2011), genetic
algorithms (Zhou and Luo, 2008), and 4-D variational assimilation
(Peylin et al., 2016). Future development will incorporate
different optimization techniques to offer users the option to search for
the best model parameters by selecting and comparing the best
method for their specific studies. We focus mostly on carbon-related state
variables in the SPRUCE example, and the data assimilation system in EcoPAD
(v1.0) needs to include more observed variables for constraining model
parameters. For example, the NEON sites not only provide measured ecosystem
CO2 fluxes and soil carbon stocks, but also resource (e.g.,
GPP–transpiration for water and GPP–intercepted PAR for light) use
efficiency (Johnson et al., 2010).
Researchers interested in creating their own multiple model and/or multiple
assimilation scheme version of EcoPAD (v1.0) can start from the GitHub
repository (https://github.com/ou-ecolab, last
access: January 2019) where the source code of the EcoPAD (v1.0) workflow is
archived. To add a new variable that is not forecasted in the EcoPAD-SPRUCE
example requires modelers and experimenters to work together to
understand their process-based model, observations, and how messaging works in
the workflow of EcoPAD (v1.0) following the example of EcoPAD-SPRUCE. To add
a new model or a new data assimilation scheme for variables that are
forecasted in EcoPAD-SPRUCE, researchers need to create additional dockers
and mount them to the existing workflow with knowledge of how information
is passed within the workflow (see the Supplement for detailed information).
With these improvements, one goal of EcoPAD (v1.0) is to enable the
research community to understand and reduce forecasting uncertainties from
different sources and forecast various aspects of future biogeochemical and
ecological changes as data become available. EcoPAD (v1.0) acts as a tool
to link model and data, not as a substitution for either model or data.
Ecological forecasting through EcoPAD (v1.0) relies strongly on theoretical
(model) and empirical (data) ecological studies. Questions regarding the
major factors regulating temporal variability in methane emissions cannot be
directly answered by EcoPAD (v1.0). How to make use of EcoPAD (v1.0) to
inspire breakthroughs in both theoretical and empirical ecological studies
is worth future exploration.
The power of EcoPAD (v1.0) also lies in the potential service it can bring
to society. Forecasting with carefully quantified uncertainty is helpful
in providing support for natural resource managers and policy makers
(Clark et al., 2001). It is always
difficult to bring complex mathematical ecosystem models to the general
public, which creates a gap between current scientific advances and public
awareness. The web-based interface from EcoPAD (v1.0) makes modeling as
easy as possible without losing the connection to the mathematics behind the
models. It will greatly transform environmental education and encourage
citizen science (Miller-Rushing et al., 2012; Kobori et al., 2016) in
ecology and climate change with future outreach activities to broadcast the
EcoPAD (v1.0) platform.
Conclusions
The fully interactive web-based Ecological Platform for
Assimilating Data (EcoPAD, v1.0) into models aims to
promote data–model integration towards predictive ecology by making
complex ecosystem model and data assimilation techniques accessible to
different audience. It is supported by meta-databases of biogeochemical
variables, libraries of modules for process models, a toolbox of inversion
techniques, and a scalable scientific workflow. Through these components, it
automates data management, model simulation, data assimilation, ecological
forecasting, and result visualization, providing an open, convenient,
transparent, flexible, scalable, traceable, and readily portable platform to
systematically conduct data–model integration towards better ecological
forecasting.
We illustrated several of its functionalities through the Spruce and
Peatland Responses Under Climatic and Environmental change (SPRUCE)
experiment. The iterative forecasting approach from EcoPAD-SPRUCE through
the prediction–question–discussion–adjustment–prediction cycle and extensive
communication between model and data creates a new paradigm to best inform
forecasting. In addition to forecasting, EcoPAD (v1.0) enables an interactive
web-based approach to conduct model simulation, estimate model parameters or
state variables, quantify uncertainty in estimated parameters and projected
states of ecosystems, evaluate model structures, and assess sampling
strategies. Altogether, EcoPAD-SPRUCE creates a smart interactive
model–experiment (ModEx) system from which experimenters can know what an
ecosystem's response might be at the beginning of their experiments,
constrain models through collected measurements, predict an ecosystem's
response in the rest of the experiments, adjust measurements to better
understand their system, periodically update projections, and improve models,
the data assimilation system, and field experiments.
Specifically for forecasting carbon dynamics, EcoPAD-SPRUCE revealed that
better forecasting relies on improvements in model structure,
parameterization, and accurate external forcing. Accurate external forcing is
critical for short-term flux-based carbon predictions, while the right process
understanding, parameterization, and historical information are essential for
long-term pool-based predictions. In addition, EcoPAD (v1.0) provides an
avenue to disentangle different sources of uncertainties in carbon cycling
studies and to provide reliable forecasts with accountable uncertainties.