Functional convergence of biosphere–atmosphere interactions in response to meteorology

Understanding the dependencies of the terrestrial carbon and water cycle is a prerequisite to anticipate their behaviour under climate change conditions. However, terrestrial ecosystems and the atmosphere interact via a multitude of variables, timeand space scales. Additionally the interactions might differ among vegetation types or climatic regions. Today, novel algorithms aim to disentangle the causal structure behind such interaction from empirical data. Visualising the estimated structure in networks, the nodes represent relevant meteorological determinants and land-surface fluxes, and the links 5 dependencies among them possibly including their lag and strength. Here we show that biosphere–atmosphere interactions are strongly shaped by meteorological conditions. For example, we find that temperate and high latitude ecosystems during peak productivity exhibit very similar biosphere–atmosphere interaction networks as tropical forests. In times of anomalous conditions like drought though, both ecosystems behave more like Mediterranean ecosystems during their dry season. Our results demonstrate that ecosystems from different climate or vegetation types have similar biosphere–atmosphere interactions 10 if their meteorological conditions are similar. We anticipate our analysis to foster the use of network approaches as they allow a more comprehensive understanding of the state of ecosystem functioning. Long term or even irreversible changes in network structure are rare and thus can be indicators of fundamental functional ecosystem shifts.

. Two-dimensional embedding coloured by underlying mean exchange rates and meteorological conditions. The mean values are calculated over the respective time periods used for the network estimation. Each network is estimated on a three month window of daily time series data. Values are cut off at the highest and lowest percentile.
Type 3 exhibits the same strong links among Rg, VPD and H as Type 2 but T is weakly or not connected and the opposite for LE and NEE. R g and T are considerably lower than in Type 2 but because of sufficient water availability the Bowen ratio is 190 between 0 and 1. Typical ecosystems in this state are mid to high latitude forests during spring or autumn, e.g. Harvard Forest EMS Tower (US-Ha1, deciduous broadleaf forest (DBF)), Roccarespampani 1 (IT-Ro1, DBF), Vielsalm (BE-Vie, mixed forest (MF)) and Hyytiälä (FI-Hyy, ENF).
Type 4 is fully and strongly connected. Both energy input and water availability are high leading to Bowen ratios around 1.
This network state is typically present in tropical forests like the Guyaflux site in French Guiana (GF-Guy) (evergreen 195 broadleaf forest (EBF)) but can temporarily be also reached by a variety of other ecosystems, e.g. mid and high latitude forests like Hainich (DE-Hai, DBF), Tharandt (DE-Tha, ENF), BE-Vie (MF), FI-Hyy (ENF) as well as woody savannas (WSA) as Howard Springs (AU-How) and grasslands as Daly River Savanna (AU-Dap).
The archetypes of networks are located at the edges of the two-dimensional space and thus could define two imaginary axes.
From a physical point of view, energy is required for each process and interaction to occur, e.g. photosynthesis or evaporation 200 (Bonan, 2015). Therefore, transitions along the axis connecting the network types 1 and 4 might be interpreted as energy  controlled as dependencies among all variables fade or increase consistently. Transitions along the axis connecting network types 2 and 3 are explainable by a combination of water availability and a temperature gradient. Low water availability but high temperatures lead to low carbon and water fluxes and thus low connectivity. On the other hand sufficient water and medium temperatures allow for fluxes but likely reduce the influence of varying temperatures leading to connected NEE and

Ecosystems' median trajectories
Each point in the reduced t-SNE space represents a biosphere-atmosphere interaction network for a given month and ecosystem. Hence, we can trace an ecosystem's trajectory through time. An ecosystem's median annual trajectory (see Sect. 2.5) within the low dimensional space reflects seasonal patterns of meteorological conditions (Fig. 4). For example, mid-latitude sites like FR-Pue (EBF), DE-Hai (DBF) and FI-Hyy (ENF) exhibit a strong seasonal variation of R g and span a long distance in the t-SNE space. In contrast, tropical ecosystems like GF-Guy (EBF) constantly have high R g and exhibit predominantly 220 network type 4 indicative of high productive conditions -while DE-Hai or FI-Hyy reach this connectivity pattern only during peak growing season. US-SRM (WSA), however, has similar or even higher R g values throughout the year but barely manages to deviate from type 2 which is in accordance with its low water availability. The amount of precipitation further aligns with differences and characteristics of the trajectories of FR-Pue, DE-Hai and FI-Hyy. For example, FI-Hyy shows some deviation towards edge 2 in February and March, FR-Pue in June, July and August. For both, mean precipitation is lowest during these 225 month. The strong control by energy and water availability is in line with a recent analysis showing that variability in landsurface processes is largely explained by productivity measures as well as water and energy availability. Both, water and energy availability, need to be high for high productive states, yet the lack of either of them leads to low productivity (Kraemer et al., 2020). This biosphere state triangle is found in our analysis by the network type 1 (cold), 2 (dry) and 4 (high productivity). Yet, a fourth network type (type 3) is naturally occurring in the t-SNE space as we here include interactions with the atmosphere.

Deviations from ecosystem median trajectories
Climatic extremes are visible in an ecosystem's trajectory as strong deviations from the median trajectory. Figure 5 shows the trajectories of ecosystems during anomalous dry or wet conditions. During the European heatwave of 2003, in July and August the trajectories of two temperate central European forests, DE-Hai and DE-Tha, no longer manage to establish a network structure resembling network type 4, typical for these ecosystems during their high productive phase. Instead they are shifted 235 towards network type 2, associated with drier conditions (Fig. 5a, b). Similarly, the ecosystem BR-Sa3 (EBF) in the Brazilian tropical rainforest shows substantial deviations towards network type 2 during the exceptional dry season of 2001 (Aug, Sep, Oct) (Marengo et al., 2018) (Fig. 5c). In contrast, US-Wkg is a grassland accustomed to dry conditions and thus predominantly exhibits low water and carbon fluxes resulting in network structures as of network type 2, i.e. water and carbon fluxes are barely or even disconnected. Carbon and water fluxes of semi-arid ecosystems, however, are known to respond quickly and strongly 240 to sufficient precipitation (Potts et al., 2019;Leon et al., 2014;Reynolds et al., 2004). This sensitivity is found to carry over to the network structure as well. The network structure of US-Wkg becomes fully connected (network type 4) in September 2014 with above average precipitation (NOAA) (Fig. 5d). The relevance of climatic conditions in controlling biosphere-atmosphere interactions on three monthly time windows thus shows also on ecosystem level as they are strong enough to explain deviations from an ecosystem's median trajectory and lead to the detection of climatic extremes. (right). As networks are calculated using a centred three month moving window, each month is ascribed a network. Thus, the behaviour of an ecosystem can be tracked by its monthly networks, which form trajectories for each year. An ecosystem's median trajectory is composed of the two dimensional monthly median networks (see Sect. 2.5 for details).

Functional convergence of biosphere-atmosphere interactions
We have seen that networks representing biosphere-atmosphere interactions are strongly shaped by prevailing mean meteorological conditions. Moreover, the visualisation of ecosystem trajectories within the t-SNE space ( 1, however, suggests the opposite. The strongest gradients are given by the links NEE-LE and Rg-LE and transitions along the axis connecting type 2 and 3 (cf. Fig. 3) are dominated by changes in biosphere connectivity, i.e. LE and NEE.
In fact, the dominance of climatic drivers in controlling the temporal evolution of ecosystem functioning emerges also in 260 other studies (Musavi et al., 2017;Schwalm et al., 2017) as they showed that carbon fluxes are primarily controlled by climatic factors. Yet, these and others also show the role of biotic factors in shaping the responses of ecosystem processes to climatic variability. For example, Musavi et al. (2017) revealed in a global ecosystem study that species diversity and ecosystem age decrease inter annual variability of GPP. Similarly, Wagg et al. (2017) discovered biodiversity to increase long-term stability of ecosystem productivity. In regional studies Wales et al. (2020) found the stability of net primary production to be affected 265 by the kind and severity of disturbances. Tamrakar et al. (2018) showed that seasonal carbon fluxes were more sensitive to environmental conditions in a homogeneous forest compared to a heterogeneous one. It would be of interest to investigate, to which degree the effects of biotic factors also translates to the sensitivity of the network structure.
Furthermore, extreme heat and drought events (Sippel et al., 2018) or compound events in general (Zscheischler et al., 2020) can severely disrupt ecosystem functions. The time of recovery from such disturbances is a crucial parameter in assessing 270 ecosystem resilience. Schwalm et al. (2017) showed that the recovery time measured as the recovery in GPP is primarily influenced by climate but secondarily by biodiversity and CO 2 fertilisation. Assessing the recovery time via GPP already puts the ecosystem functioning into focus. The here presented framework, i.e. the sensitivity of an ecosystem's network structure to meteorological conditions, might be a valuable asset to study reaction time and strength to and recovery from extreme events as it not only utilises one variable but the interactions of a set of variables, thereby capturing more comprehensively an ecosystem 275 state. A drawback is the reduced temporal resolution (a certain time period of daily or even half hourly measurements is aggregated to one network) which can be offset by the here used moving window approach to a certain degree. Especially with regard to climatic extreme conditions in recent years with observed vegetation dieback in, for example, DE-Hai (Schuldt et al., 2020), further studies could also shed light on the role of adaptation in shaping biosphere-atmosphere interactions. Our study suggest that adaptation to a lesser degree limits the range of possible interactions but enables to sustain and persist certain 280 conditions for longer periods. The focus of further studies thus could be to elucidate the role of biotic factors in influencing ecosystem trajectories as well as the role of adaptation and the response to extreme events.

Conclusions
We analysed the functional behaviour of a variety of ecosystems using the FLUXNET database of eddy covariance measurements. In particular, we examined the interaction structure between biosphere-atmosphere fluxes as well as atmospheric state 285 variables using PCMCI, an algorithm to estimate causal relationships from empirical time series. In total we included 119 measurement sites with cumulative 1067 measurement years leading to 10038 monthly networks. Using non-linear dimensionality reduction, we found four archetypes of network states defining the edges of the low dimensional embedding. They are characterised on the one hand by a fully connected and almost unconnected network structure and on the other hand by an antagonistic coupling of carbon and water flux with temperature -when one is strongly coupled, the other is decoupled. The transitions between these states correlate well with gradients of meteorological drivers, i.e. radiation and water availability.
The movement of an ecosystem within that space therefore strongly aligns with changes in meteorological conditions. This, however, also leads to similar behaviour under similar conditions for strongly contrasting ecosystems. For example, forests of mid or even high latitudes exhibit similar interaction structure as tropical forests given high radiation and water availability during summer. Yet, this state can also be reached by predominantly dry ecosystems like steppe grasslands given sufficient