Steering operational synergies in terrestrial observation networks: opportunity for advancing Earth system dynamics modelling

Advancing our understanding of Earth system dynamics (ESD) depends on the development of models and other analytical tools that apply physical, biological, and chemical data. This ambition to increase understanding and develop models of ESD based on site observations was the stimulus for creating the networks of Long-Term Ecological Research (LTER), Critical Zone Observatories (CZOs), and others. We organized a survey, the results of which identified pressing gaps in data availability from these networks, in particular for the future development and evaluation of models that represent ESD processes, and provide insights for improvement in both data collection and model integration.

From this survey overview of data applications in the context of LTER and CZO research, we identified three challenges: (1) widen application of terrestrial observation network data in Earth system modelling, (2) develop integrated Earth system models that incorporate process representation and data of multiple disciplines, and (3) identify complementarity in measured variables and spatial extent, and promoting synergies in the existing observational networks. These challenges lead to perspectives and recommendations for an improved dialogue between the observation networks and the ESD modelling community, including co-location of sites in the existing networks and further formalizing these recommendations among these communities. Developing these synergies will enable cross-site and cross-network comparison and synthesis studies, which will help produce insights around organizing principles, classifications, and general rules of coupling processes with environmental conditions.

1 Introduction

Complex interactions among rock, soil, water, air, and living organisms regulate the natural habitat and determine the availability of life-sustaining resources for human well-being (MEA, 2005). In the light of accelerating global change (e.g. Camill, 2010; IPCC, 2014) and safeguarding Earth as a habitable space, scientific and societal demands require improved understanding of Earth system dynamics (ESD). Understanding and modelling of Earth system processes and interactions among Earth system compartments can be enhanced by accessing a wider range of both observational and experimental data (Aronova et al., 2010; Banwart et al., 2012; Reid et al., 2010). For these purposes, observation networks aimed at temporal and multidisciplinary coverage of continental- and global-scale ecosystem observations have been developed in recent decades. The Critical Zone Observatory (CZO) network (White et al., 2015), the International Long-Term Ecosystem Research network (Mirtl et al., 2018), the US National Ecological Observatory Network (Loescher et al., 2017; Schimel et al., 2011), the Chinese Ecosystem Research Network (Fu et al., 2010), and the Australian Terrestrial Ecosystem Research Network (Lindenmayer, 2017) are examples of networks that focus at the continental to global spatial extent and daily to decadal temporal scales.

One overarching goal of these research and observation networks is to use measurement data to improve the predictive capabilities of current models (Loescher et al., 2017). The growing availability of data (Hampton et al., 2013) and an improved representation and resolution of modelled processes drive the development of ESD models with ever-increasing sophistication (Wood et al., 2011) and novel validation and assimilation techniques (Penny and Hamill, 2017). Terrestrial Earth system models represent a large range of processes, from tracking the fluxes and storage of energy, water, sediments, carbon, and other elements (scalars) to distributions and functional roles of organisms, land use practices, climate, and humans (Mirtl et al., 2013). However, the majority of these models focus on one or a few processes. In contrast, we define integrated models as terrestrial Earth system models which include interactions and feedbacks among water, energy, and weathering cycles with biota, ecosystem functions, and services (Vereecken et al., 2016b). Integrated models include cross-scale and cross-disciplinary processes that are needed to fully predict ESD responses to perturbations from driving forces at local to global scales. No single ESD model can accomplish the full representation of driver and response functions. For this reason, developing integrated models dealing with different processes, such as land surface models, or coupling existing process models in suites (e.g. Duffy et al., 2014; Peckham et al., 2013) are options to expand our current modelling capability to incorporate cross-disciplinary processes for improved prediction of whole ESD system-level understanding, as well as for policy and management decisions. The application of hydrological, meteorological, biogeochemical, and biodiversity measurements from within and across sites into such integrated model systems (Fig. 1) is a key component to providing the multi-scale/multi-process understanding that is needed to advance predictions of ESD responses to land use and climate changes, and the ever-increasing demand for natural resources.

Figure 1 Flowchart of concepts, pathways, and processes of applying terrestrial observatory network data to Earth system dynamics models; identifying the three challenges of (I) data application, (II) model integration, and (III) steering synergies in observation networks.

For these purposes, the CZOs provide essential datasets and a coordinated community of researchers that integrate hydrologic, geochemical, and geomorphic processes from soil grain to watershed scales (Brantley et al., 2017). CZOs examine the interactions among the lithosphere, pedosphere, hydrosphere, biosphere, and atmosphere (White et al., 2015; Banwart et al., 2012; Brantley et al., 2007). CZOs examine how scalar mass and energy fluxes interact with life and lithology over geological timescales that see the transformation of bedrock into soils, and how the same coupled processes enact feedbacks with changing climate and changing land use (Brantley et al., 2016; Lin et al., 2015; Sullivan et al., 2016). The US Long-Term Ecological Research (LTER) network was created with the aim to provide the data and information needed for long-term, integrative, cross-site, nationwide research based on principal investigators (PIs) to advance ecological literacy and act upon solving grand societal challenges (Callahan, 1984). This US project quickly gained attention and sparked the foundation of other national and regional LTER networks (e.g. China, Europe, Australia). This led to the foundation of the global International LTER network in 1993 (ILTER) (Kim, 2006; Mirtl, 2010; Vanderbilt and Gaiser, 2017), which currently comprises 44 formal national LTER networks. ILTER provides the scientific expertise, a global-scale network, and long-term datasets necessary to document and analyse environmental change. In addition, Long-Term Socio-Ecological Research (LTSER) platforms are designed to support research on long-term human–environment interactions and transdisciplinary approaches. Recognizing that the value of long-term data extends beyond use at any individual site, the global ILTER network aims at making data collected by all (I)LTER sites broadly accessible to other investigators (e.g. Breda et al., 2006; Parr et al., 2002; Vihervaara et al., 2013), enhanced by the standardized documentation of measurements and sites (Haase et al., 2016) and a proposal to unify abiotically and biotically oriented concepts (Haase et al., 2018). Site metadata and data are increasingly accessible via the DEIMS-SDR web service (Mirtl et al., 2018).

As the above review illustrates, each of these networks was created in an effort to address recognized science questions and knowledge gaps in specific disciplines that led to their creation (e.g. ecology, geology). The collection of data within each network – the variables measured, the methods by which they were measured, and associated campaign activities leveraging these networks – have been designed to address these specific questions and knowledge gaps. At the same time, the science questions, observed variables, and associated measurement methods lead to opportunities for across-network synthesis and co-production of knowledge across these networks and disciplines. Moreover, there is great potential to advance the development of ESD models, especially integrated ones, by providing site-level observations for calibration, validation, and data assimilation. However, the question remains – to what degree do modelling efforts use observatory network data, and how do those efforts fit within the broader activities of the networks?

For this study, the International Soil Modelling Consortium (ISMC, https://soil-modeling.org/, last access: 19 May 2018) conducted a survey with participation from the ILTER and the global CZO network (Critical Zone Exploration Network) to identify knowledge and functional gaps in data availability and in our ability to integrate models and the data application. ISMC envisions integration of models from different disciplines of hydrology, biogeochemistry, and ecology to increase the understanding and awareness of ESD processes, especially when these processes underpin other processes (e.g. carbon cycling, biological activities, soil formation, global and regional climate) (Vereecken et al., 2016b). To this end, the survey brings quantitative information on the level of integration of modelling approaches that use data from the LTER and CZO sites. Based on the results of the survey, we describe challenges and implications for (1) usage of observatory data in integrated ESD models, (2) model integration in relation to specific disciplines, and (3) complementarity and possibilities for steering network synergies (Fig. 1).

2 Material and methods

The survey was addressed to the principal investigators (PIs) of individual CZOs, to the larger Critical Zone Exploration Network (approximately 1600 individuals), and to the PIs of the ILTER network (approximately 400 individuals) with the request to forward the survey to associated modelers. The first part of the survey collected information on the model used, the geographic region, purposes of the modelling activity, spatial and temporal scale, compartments, disciplines, and model structure. The second part of the survey identified the type of variables and data used, data application (model input or calibration and validation), and the source of data used in the model (measurements at sites, remotely sensed, database, modelled, or literature).

As the purpose and scientific origin of the LTER and CZO networks differ, the respondents were asked about a diverse set of variables and parameters. In total, 52 variables were included in the survey (Table A1) based on the common measurements in the LTER and CZO networks (Chorover et al., 2015; Brantley et al., 2016). Survey results were tabulated and analysed to address several questions. Examining both networks separately and analysing them together, we tested (1) the degree to which variables or model characteristics were associated with a specific network, (2) the relationship between model integration (range between 0 and 1) and number of variables (range: 0 to 52) used in the model, and (3) a correspondence analysis for the data application across the models. The model “level of integration” (an index ranging from 0 [low] to 1 [high]) was calculated by normalizing the model-wise number of disciplines and compartments indicated by the responses to a scale of 0 (none) to 1 (many) and averaging these two indices. The survey variables were combined in an ordination using detrended correspondence analysis (Hill and Gauch, 1980) implemented in R and Fortran (Oksanen and Minchin, 1997). The ordination was designed to identify common features among the models used.

3 Challenge I: observatory data application to ESD modelling

3.1 Current status

The survey revealed a wide variety of models in terms of disciplines and scales (Fig. A1). Out of 118 completed surveys, 70 provided full information on model characterization and variables. Nearly half of the respondents (47 %) reported on use of CZO observational data, two-thirds (66 %) used LTER data, and 12 % reported model applications using data from both networks (Fig. A1a). Geographically, the majority of model applications came from Europe (63 %), followed by North America (27 %), the whole globe (18 %), Asia (12 %), and Africa (5 %). Particularly in Europe, a large fraction of respondents were associated with LTER, while in North America the CZO community was the most responsive.

Figure 2 Heat map of model applications, with spatial and temporal dimensions of the surveyed models. Most models were denoted using several temporal and spatial scales. For visualization, we present some individual models at one exemplary instance. On the right, CZO and LTER model applications are presented separately.

The average model used 14 variables of the supplied list of 52, ∼ two-thirds of the variables for model input and ∼ one-third for calibration and validation (Table A1). The majority of variables used in the models are sourced from on-site measurements (55 % on average). The rest of the data (45 %) are derived from other sources, mostly remote sensing (e.g. MODIS) and to a lesser extent modelled (e.g. North American Land Data Assimilation System), external database (e.g. FLUXNET), or literature sources. The most common remotely sensed variables were of the biosphere, especially habitat mapping, leaf area index, vegetation structure and dynamics, and above-ground biomass, but also snowpack distribution and duration. Variables used most frequently in models applied by CZO and LTER communities were from the atmospheric compartment (precipitation, air temperature, incoming shortwave radiation, humidity, wind speed and direction, and eddy flux of evapotranspiration and CO₂), followed by soil characterization (structure, texture, water content), above-ground biomass, and vegetation structure and dynamics. This reflects the current most frequent model requirements and applications in terrestrial Earth system science for coupled hydrological–biogeophysical models. Model applications affiliated with the CZO were more focused on the lithosphere and cryosphere, while biodiversity was addressed only in models of the LTER community (Fig. A1e). While the CZO model applications use variables and data related to saprolite and bedrock mineralogy, data on biotic and biodiversity variables were used more frequently in models associated with LTER. Models associated with CZOs applied significantly (Fisher's exact test, p<0.05) more data based on eddy flux measurements (evapotranspiration and CO₂), root density, soil water content, soil temperature, bedrock, and soil texture and physics compared to models associated with LTER. Models using data related to habitat mapping and biotic and biodiversity elements were associated with the LTER community.

There is a large congruence in the spatial and temporal scales between both communities. Spatial scales of models were primarily site to catchment scales, with few models at macropore, lab, or global scales (Fig. 2). The high density in the centre of Fig. 2 shows the focus for sub-catchment-scale modelling, and timescales of days to years, potentially decades, which is in line with the aims and conceptual basis of both LTER and CZOs. This result stresses the relevance of both observation networks to ESD processes in terms of the spatial and temporal scales in CZO and LTER modelling activities. At the same time, Fig. 2 reveals a lack of modelling activities at a larger extent (continental and global) in both communities. The specific inset diagrams show that for LTER, the prevailing yearly, potentially decadal, time span is mostly covered at the site scale, whereas CZO-associated responses work predominantly at the catchment scale and incorporate daily resolution. Some CZO models seem to cover a larger range of spatial scales since the models indicated in the survey cover the full spatial range from macropore to continental and global scales. In terms of the modelled timescale, the long timescales (centuries to millions of years) are mostly covered by models employed in the CZO community (Fig. A1d). These scales of modelling are consistent with the focus on process and system understanding of both networks' model applications, while the management and prediction aspect is more strongly embraced by LTER model applications in the survey results (Fig. A1c).

Figure 3 LTER sites answer specific ecological questions, for which specific data are gathered, e.g. black poplar population persistence under climate change. Black poplar population strength along the French Loire River section (a) and model projections (b) under current, climate change, and adaptation management scenarios (Van Looy and Piffady, 2017).

3.2 Example of data application from an LTER site

Next to the common, cross-site measurements, CZO and LTER datasets generally include site-specific types of observations gathered to answer site-specific scientific questions on model development, ecosystem response to global change, and prediction. One example is the vegetation dynamics modelling in the French LTER zone Atelier Loire (Van Looy and Piffady, 2017), which uses predicted hydrological changes in river flow regimes and droughts to predict changes in land use and vegetation in the Loire floodplain (Fig. 3). It enables the construction of population dynamics models for characteristic tree species black poplar and white elm for the LTER site where count data of the species populations are present. The proposed adaptation management scenario of water retention and restoration of flow regime and floodplain inundation proved successful according to the model to mitigate predicted climate change impacts on population dynamics.

Another example of vegetation dynamics modelling using observation network data concerns forest dieback under climate change (Breshears et al., 2005). At an intensively studied site of the Drought-Induced Regional Ecosystem Response Network, after 15 months of depleted soil water content, >90 % of the dominant overstorey tree species died. This combination of detailed spatial–temporal observational data on tree condition, soil water content, precipitation, and atmospheric conditions (temperature) allowed for data-driven development and validation of a regional model on drought-induced vegetation changes.

3.3 Open issues and implications

The LTER–CZO network sites monitor a wide range of environmental variables with long-term or at least regularly repeated measurements, expected to provide more reliable and robust results than single measurements that produce “snapshot” information only. The application of long-term monitoring data to enhance predictive capacity provides a strong opportunity in the era of ESD modelling (Parr et al., 2002). Application of the rich data collected at LTER–CZO network sites should improve process understanding and enable the scientific community to address the challenges of validating Earth system models that integrate coupled processes. Although integrated models at LTER and CZO sites are used to raise understanding of the coupled processes, cases are mostly restricted to individual sites and too limited in number. As the survey results demonstrated, for some themes (e.g. habitat–vegetation–crop), remotely sensed or existing database information was preferably used in contrast to potential data from on-site field measurements (Table A1). As on-site measurements generally are more accurate than remotely sensed or modelled data, this suggests a strong need for on-site measurements for modelling site-specific processes. Plausible causes for the lack of on-site measurements relating to vegetation and biota are the time- and personnel-consuming requirements for data collection, and the absence of harmonized measurement protocols. The strong complementarity identified in data sources and model applications, in terms of biotic vs. abiotic and above- vs. below-ground in LTER and CZO networks (respectively), does suggest strong potential benefits and gains in process understanding if data from both observatory networks can be applied simultaneously to integrated models at regional, continental, and global scales.

4 Challenge II: model integration

4.1 Current status

On average, models had a rather high level of integration; CZO and LTER data model applications cover on average multiple disciplines (mean $=\mathrm{3.6}\pm \mathrm{2.1}$ SD) and ecosystem compartments (mean $=\mathrm{2.7}\pm \mathrm{1.6}$). Model “level of integration” was strikingly similar in CZO (mean $=\mathrm{0.37}\pm \mathrm{0.18}$) and LTER (mean $=\mathrm{0.34}\pm \mathrm{0.17}$). The richness in variables was positively related to the number of disciplines (R²=0.29), and to compartments (R²=0.4). However, unifying compartments and disciplines to the level of integration measure correlated most strongly to the number of variables (R²=0.47) (Fig. 4).

4.2 Example of integrated modelling

Plant–soil interactions are changing across the globe, whether by the encroachment of woody species into polar, alpine, and temperate grassland areas (Archer et al., 1995; Jackson et al., 2002), the increase in atmospheric CO₂ concentrations that potentially alter the depth penetrations of roots (Bond and Midgley, 2012; Van Auken, 2000), or changing land cover (agriculture, forest plantations; Van Minnen et al., 2009). Subsurface changes to the root system architecture (root function, density, and depth) alter the introduction (spatial distribution) of organic carbon into the ground, controlling microbial productivity and respiration, macropore location, distribution, and evolution, controlling the transport of most water that moves through soil (Beven and Germann, 1982), and spatial distribution of organic acids and root respiration (generation of CO₂; Jones, 1998). These factors will impact infiltration of meteoric water charged with carbonic acid (H₂CO₃), influencing the breakdown of minerals and the redox conditions under which metals can be mobilized.

Figure 4 Model-wise “level of integration” calculated from the summed scientific disciplines and modelled compartments for the corresponding model. Trend lines corresponding to the models associated with LTER (blue) and CZO (red), and models associated with data from both networks (black) are presented.

Figure 5 (a) Representative processes from different disciplines (different colours) in different modules that can be integrated into one model suite or a coupled reactive transport–land surface–hydrologic model RT-Flux-PIHM (Bao et al., 2017). The integration among processes from mechanistic bases in the model will allow systematic understanding at the watershed scale. (b) Model result of simulated difference in groundwater flow for a grassland (dotted line) and woody encroached (solid line) watershed (N04-D), at the Konza Prairie, KS (USA). Grassland simulations are parameterized with a 0.3 m rooting depth and enhanced horizontal macropore development, while woody encroached simulations have a rooting depth that extends to 1.0 m deep.

To explore the larger-scale consequences of changes in root system architecture on soil water and riverine chemistry requires integration of processes from different scientific disciplines into an integrated or coupled model. Here we show the example of RT-FluxPIHM, which integrates processes from a reactive transport (RT) with a land-surface and hydrologic model (FluxPIHM) (Fig. 5) (Bao et al., 2017; Li et al., 2017a). RT-FluxPIHM is being used to examine the hydrologic and biogeochemical ramifications of woody encroachment into grasslands at the Konza Biological Station (KS, USA), a well-characterized and well-monitored LTER site. Preliminary numerical experiments explore how differences in rooting depth and macroporosity distribution (vertically and horizontally) alter groundwater flow patterns, and thus stream water discharge and solute behaviour. The enhanced vertical macropore development through deeper roots of woody encroachment compared to grass led to higher groundwater flow (Fig. 5b). One limitation to such complex integrated numerical models is the numerous datasets needed for parameterization. However, working with datasets derived from LTER, CZOs, and NEON allows the evaluation of model performance against data that characterize key processes embedded within integrated models. These types of coupled models (see also Dhara model by Le and Kumar, 2017) offer a way to explore plant–water–biogeochemical feedbacks at the watershed scale and help guide future field experiments.

4.3 Open issues and implications

Fragmentation and lack of integration has limited our abilities to understand the formation and function of ESD at various spatial scales, and to predict system response to global change and interaction of processes and parameters from sites to continents (Grimm et al., 2013). In our survey, 10 % of the models already used data from both LTER and CZO networks in model applications. This implies that these models already integrated processes of interest to the two communities. However, inclusion of multiple disciplines and ecosystem compartments may increase model complexity and data requirements. Hence, integrated modelling may be limited by data availability but allows for more general applicability of conclusions (Basu et al., 2010; Li et al., 2017b). The numerical model applications necessary to test LTER and CZO conceptual model assumptions are integrated, process-based, spatially explicit models at the watershed scale that predict emergent behaviour. The high level of multidisciplinary model inputs requires numerically expensive models and more importantly a sharp learning curve of the users. It does not necessarily mean that the data are co-measured. Where CZOs mainly focus on understanding near-surface structure and dynamics, developing synergies with LTER might fill many of the ecological gaps in CZO studies by providing the scientific expertise, research platforms, and datasets necessary to analyse environmental change with a particular focus on ecological-driven processes. Whereas the conceptual models for LTER and CZO sites are oriented towards to fundamental process understanding using specific parsimonious models, integrated models remain an essential part of the mission. Model integration does not necessarily require an increase in model complexity. Parsimonious models can be integrated in a larger model platform (e.g. Duffy et al., 2014; Peckham et al., 2013) to investigate feedbacks over climatic and geographic gradients, and across disciplines.

5 Challenge III: complementarity and disciplinary segregation

5.1 Current status

Models that included data from both LTER and CZO sites cluster in the centre of the ordination (Fig. 6), which indicates that those models use a fairly similar group of variables. The models located in the centre of the ordination focus mostly on hydrology and geophysical processes. Models clustered in group A are associated with CZOs, are mostly located in the second quadrant, and are distinguished by a focus on modelling processes in the soil profile, regolith, and bedrock. Models clustered in group B are associated with LTERs, are mostly located in the first and fourth quadrants, and focus on processes related to ecosystems and biota. The horizontal axis of the ordination separates physics-oriented from biotic-oriented models, and below-ground (negative) from above-ground (positive). The vertical axis distinguishes highly integrated models, mainly hydrology based and containing a number of processes, from specific models, such as those focusing only on rock weathering.

Figure 6 Ordination of the models and data in the survey; clusters show significant differences for models of LTER and CZO. CZO cluster A focuses on variables of the saprolite and bedrock; LTER cluster B focuses on the variables of biota and biodiversity.

5.2 Open issues and implications

Available datasets provide the opportunity to integrate a larger number of compartments into ESD models, but the ordination reveals that this integration has not progressed very far. The survey revealed that models using data from both CZO and LTER networks generally cover a larger range of variables compared to models applied to only one network (Fig. 4). However, the outlier position of models focusing on biota and habitat variables in the ordination (Fig. 6) indicates the need to integrate the biotic compartment in models of coupled processes such as energy, carbon and nutrient cycling, and weathering (Filser et al., 2016; Richter and Billings, 2015; Vereecken et al., 2016b; Wall et al., 2015). In this survey, many biotic models were opposed to below-ground compartments in the ordination (Fig. 6). This result demonstrates the lack of models of biotic processes in the subsurface, e.g. the representation of the weathering microbiome or root system architecture and dynamics (Smithwick et al., 2014), despite the fact that the underground biota performs a crucial ecosystem functioning role (Deruiter et al., 1994; Wall et al., 2015). Therefore, it is particularly important to harmonize and standardize observations of biotic variables related to processes and feedbacks with hydrologic and biogeochemical cycles. Recent initiatives address the missing integration of below-ground biota in terrestrial Earth system science and models (Key to Soil Organic Matter Dynamics and Modelling – KEYSOM-BIOLINK project; Filser et al., 2016), and the provision of substantial datasets on soil biota and biodiversity (global soil biodiversity database; Ramirez et al., 2015). Some models are being developed that are capable of estimating the role of biotic activity in soil formation, decomposition–mineralization processes, and predicting the carbon and nutrient cycles in specific soil types (Komarov et al., 2017; Wieder et al., 2017). Nevertheless, joining discipline-specific data with the largely site-to-catchment-based but discipline-specific modelling expertise of the CZO and LTER communities would lay the ground for new findings.

6 Outlook

6.1 Satisfying cross-disciplinary data demand with ESD models

The relationship between models and data is a relationship of mutual benefits. Data enable scientists to develop and test hypotheses (e.g. Braud et al., 2014; Clark et al., 2011), but models may also help scientists to better design data collection strategies and tactics for observation networks (Brantley et al., 2016). With the increase in computational capabilities, stochastic methods such as data assimilation, global sensitivity analysis, and optimization algorithms are becoming more widely used. Commonly, these methods are used for parameter and state estimation. Additionally, stochastic analyses open the way to determine the observation requirements to reduce model uncertainties and test hypotheses. Stochastic analyses can be used to identify key physical processes and their impacts if variables are subject to change. Thus, models can improve observation network strategies by quantifying process sensitivity to observed variables and parameters, as well as measurement frequency, and resolution and extent needs in space and time (Lin, 2010). Model-based assessments of observability, predictability, and the impact of heterogeneity on processes at the relevant scales could improve network data collection efficiency and complementarity. Merging data and modelling through data assimilation may also enable testing predictions from small-scale process understanding in larger-scale, simplified model representations (Heffernan et al., 2014; Vereecken et al., 2016a).

In particular, the reanalysis concept addresses the benefits of data application in ESD modelling. Although it is widely used in meteorological models (e.g. Compo et al., 2011; Dee et al., 2011), reanalysis has only sparingly been used in terrestrial Earth system science or ecology. For performing reanalysis, a physics-based model is fed with observations through a data assimilation scheme over a sufficiently long time period to update model states and parameters over time. Model states and parameters are optimized with the data assimilation method based on the observation, considering uncertainty in observations, model structure, initial states, and forcings. Application of reanalysis in Earth system models could generate gap-filled, multi-compartment, and coherent physics-based time series of terrestrial states, fluxes, and parameters including variables characterizing biological processes and biodiversity. Based on often non-continuous and sparse in situ observations from long-term observation networks such as CZOs and LTER, critical zone or ecosystem reanalysis would need to specifically target biological and biodiversity-related processes in Earth system models. The generated continuous reanalysis data could inform further modelling processes or be used to test existing hypotheses. However, to undertake reanalysis for ESD, many questions need to be resolved, including the choice of the Earth system model, the data assimilation method, model parameterization and forcing data, validation data, and ultimately the representation of biotic and abiotic processes.

A related challenge is how to represent the roles of biota in integrated ESD models (Deruiter et al., 1994; Richter and Billings, 2015). Such efforts must be based on improved understanding of biotic–abiotic interactions, feedbacks, and thresholds (e.g. forest dieback; Breshears et al., 2005). Integrated ESD models must include phenomena such as community assembly, evolution, the emergence of pests and pathogens, and invasion by invasive species, which are not currently included in CZO models. New initiatives have been launched recently to integrate the biotic component in Earth system science and models (Filser et al., 2016), including, for example, modelling the roles of biota (e.g. bacteria, fungi, roots) in the subsurface (Grandy et al., 2016). At the same time, improved representation of processes such as hydrologic and geochemical cycles may improve the integration of LTER models.

6.2 Integration of models

Integrated models covering different disciplines and compartments are needed to objectively increase process understanding and develop predictive capabilities on the effects of climate and land use changes on ESD. In our survey, the few land surface–atmosphere integrated process-oriented models like PIHM and Parflow-CLM were exceptional in the level of integration and application of observation network data. Along with a few other examples (e.g. Boone et al., 2009; Lafaysse et al., 2017), land surface models are rarely used to model processes with an integrated approach embracing biotic and abiotic variables using LTER and CZO data. A stronger communication between land surface modellers and LTER–CZO communities would enhance the integration of in situ observations in models. However, this communication must overcome the disparity in scales of analysis between the continental-to-global focus of land surface models and the site-to-region focus of LTER–CZO observatories.

The majority (80 %) of the surveyed modellers also supported the idea of creating a model platform, but they were divided about what services should be provided on such a platform. Integration of parsimonious models into an integrated process-based model could be one service under such a model platform. A model platform could promote the understanding of organizing principles, classifications, and general rules of coupling processes and environmental conditions (Sawicz et al., 2011; Sivapalan, 2003; Sivapalan et al., 2003). Insights can also be gained through cross-site comparison and synthesis studies of observation data across different sites under gradients of climate, Earth surface characteristics (e.g. soil type, lithology, topography, vegetation), and human impact (e.g. pristine, agriculture, urban) conditions, which observation networks are well positioned to carry out.

6.3 Strategies for steering synergies in Earth observatory networks

With respect to investigating specific aspects of ESD, the interactions of biotic and abiotic processes as well as below-ground and above-ground processes are key links, where geosphere-focused research by CZO and ecology-focused research by LTER could benefit observation and model integration, optimizing the joint use of resource-intensive observatories by more than one research community. Leveraging the CZO and LTER data across networks and scales implicates enhanced ESD modelling capabilities. Desirable data harmonization across networks could be achieved based upon blending conceptual frameworks such as the ecosystem integrity (Muller et al., 2000) and the essential biodiversity variables (Pereira et al., 2013) as suggested by Haase et al. (2018), and the CZO approach (Chorover et al., 2015; Brantley et al., 2016). Data harmonization among networks and co-location of sites by different networks allow for more efficient allocation of resources and increases multi-compartment datasets at co-located sites. Co-location is the joint use of individual research sites by two or more networks. The merger of data from different networks to calibrate and validate ESD models, following the example of the Coupled Model Intercomparison Project (Meehl et al., 2005), may help existing networks to identify missing variables and potential additional observation sites in a resource-efficient manner. Continued efforts to integrate ESD models and data will help advance ESD process understanding. Furthermore, the interaction of observatory networks increases the spatial coverage of multi-compartment observations, allowing ESD models to address research questions and test hypotheses over larger scales, gaining full benefit of multi-compartment CZO and LTER data.

Considerations about steering observatory network synergies need to consider differences in the organizational structure, where CZOs have been mainly based on scientific networks and projects, while LTER has established formal governance structures regionally and globally. The degree of implementation and formalization of observatory networks also varies with geography, ranging from regions with well-established networks (US-LTER, US CZO, and NEON) to regions where research and observation networks are based on the initiative of individual sites, observatories, or projects. Some existing Earth observation networks such as NEON are already more systematic in spatial coverage and constitution, offering opportunities for advanced geographical ESD analysis. A notable European initiative is the Integrated European Ecosystem, Critical Zone and Socio-Ecological Research Infrastructure (eLTER RI), which includes the focal aspect of CZO research and requirements of widely used ecosystem models. In these attempts to steer synergies, the role of discussion amongst stakeholders, decision makers, funding agencies, and the broader scientific community cannot be overstated.

7 Concluding recommendations

The CZO and LTER networks could promote interdisciplinary research that improves process-based models spanning the geosciences and biosciences (Brantley et al., 2017; Rasmussen et al., 2011), and modelling efforts may feed back to help improve observation network design. To be effective, a stronger dialogue is needed between the observatory networks. More work is needed to apply CZO and LTER data in ESD models and thus strengthen the crucial role of the observatory networks in raising understanding of ESD processes and deriving predictive capabilities for drivers, impacts, and responses to global change. The rapidly increasing technological capabilities in computational power, ground-based instrumentation, and unmanned automated remote sensing require all stakeholders to decide on which aspects the future observational requirements shall focus. Given today's grand challenges, the communities need to focus on expanding observation efforts towards cross-community harmonized methods and datasets. The communication and exchange about services and tools for making data available through web platforms offer obvious opportunities in this sense.

Finally, there is an essential need to educate and train the next generation of Earth system scientists for modelling across disciplines. This indicates the need for dedicated Earth system science university courses, online teaching materials on model usage, and a coordinated, community-driven modelling platform.

Appendix A

Figure A1 Distribution of respondents associated with LTER, CZO, or both sites (a) to geographic region modelled (b), purpose of the modelling (c), timescale modelled (d), disciplines (e), and compartments (f) integrated.

Table A1 Summary table of variables, associated ecosystem compartment, and the times it was associated with a model, the source for the data being either from site observation or from another source: i.e. remotely sensed, modelled, external database, literature.