Bridge to the future: Important lessons from 20 years of ecosystem observations made by the OzFlux network

Abstract In 2020, the Australian and New Zealand flux research and monitoring network, OzFlux, celebrated its 20th anniversary by reflecting on the lessons learned through two decades of ecosystem studies on global change biology. OzFlux is a network not only for ecosystem researchers, but also for those ‘next users’ of the knowledge, information and data that such networks provide. Here, we focus on eight lessons across topics of climate change and variability, disturbance and resilience, drought and heat stress and synergies with remote sensing and modelling. In distilling the key lessons learned, we also identify where further research is needed to fill knowledge gaps and improve the utility and relevance of the outputs from OzFlux. Extreme climate variability across Australia and New Zealand (droughts and flooding rains) provides a natural laboratory for a global understanding of ecosystems in this time of accelerating climate change. As evidence of worsening global fire risk emerges, the natural ability of these ecosystems to recover from disturbances, such as fire and cyclones, provides lessons on adaptation and resilience to disturbance. Drought and heatwaves are common occurrences across large parts of the region and can tip an ecosystem's carbon budget from a net CO2 sink to a net CO2 source. Despite such responses to stress, ecosystems at OzFlux sites show their resilience to climate variability by rapidly pivoting back to a strong carbon sink upon the return of favourable conditions. Located in under‐represented areas, OzFlux data have the potential for reducing uncertainties in global remote sensing products, and these data provide several opportunities to develop new theories and improve our ecosystem models. The accumulated impacts of these lessons over the last 20 years highlights the value of long‐term flux observations for natural and managed systems. A future vision for OzFlux includes ongoing and newly developed synergies with ecophysiologists, ecologists, geologists, remote sensors and modellers.

other flux networks and research communities of the importance of data sharing.
Like OzFlux, this global network of micrometeorological 'flux towers' that use the eddy covariance method, provide observations to advance the understanding and simulation of processes across the past, present and future for a wide array of the world's ecosystems.
These continuous, long-term and standardised measurements are critical for detecting ecosystem stress, recovery from disturbance, and resilience to climate change, as well as exploring the causes and effects of longer-term climate trends and interannual variability-a goal unattainable with short-term records (Baldocchi et al., 2018).
In-situ flux tower and remote sensing observations are being combined to upscale from site to regional and global scales (e.g. Cleugh et al., 2007;Jung et al., 2020;Schimel & Schneider, 2019), contributing valuable data-driven diagnoses of how climate change affects terrestrial carbon and water cycles (e.g. Piao et al., 2020). Similarly, combining in situ flux tower measurements, manipulation experiments and satellite remote sensing are advancing knowledge of how climate extremes affect the carbon cycle .   Chapin et al. (2006) for definitions of carbon cycle terms used in this paper. FLUXNET's global database of ecosystem-scale observations are being used to evaluate and improve the processes represented in many ecophysiological, hydrological and land surface models (LSMs), improving the regional and global Earth System models used around the world (e.g. Ziehn et al., 2020).
Vegetation of Australian and New Zealand ecosystems have evolved in geographic isolation, geological stability, long-term aridity and fire-prone environments. In Australia, these conditions have resulted in a unique flora with scleromorphic properties enabling existence in arid climates on old, highly weathered, low-nutrient soils and frequent fire (Fox, 1999). As a result, endemism in Australian flowering plants and gymnosperms is extremely high at 93% and 96% relative to global floras (Chapman, 2009). The Australian climate envelope differs from that of Europe, most of North America, Asia and South America, being, on average, warmer and drier (both in terms of rainfall and vapour pressure deficit; VPD) but also subject to larger interannual variations in rainfall and VPD than expe-  The aim of this paper is to describe the unique and most important insights, and new knowledge contributed by the OzFlux network over its 20-years of operation. Through a series of short 'lessons', we show how Australian and New Zealand ecosystems and landscapes interact with land management practices, climate variability and climate change, with a focus on the following: (1) ecosystem response, resistance and resilience to disturbance and stress; (2) ecosystem processes that modulate water availability, runoff and productivity and (3) net greenhouse gas emissions and the potential for these ecosystems to mitigate climate change and support ecosystem services and food production in the future. This aim reflects that our primary audience for these lessons is the ecosystem research community, however we anticipate that those 'next users' of the knowledge, information and data that networks such as OzFlux support may also find benefit from these insights. In distilling the key lessons learned, we also identify where further research is needed to fill knowledge gaps and improve the utility and relevance of the outputs from OzFlux.

| THE G ENE S IS OF OzFlux
The OzFlux journey began in the early 1990s when Australian and New Zealand researchers embarked on longer-term micrometeorological field campaigns and studies in agricultural, natural and modified forest, native grassland and wetland ecosystems. This research revealed gaps in our knowledge of ecosystem dynamics and feedbacks with climate and hydrology at multiple timescales, across the diverse landscapes of New Zealand and Australia (Campbell & Williamson, 1997;Cleugh et al., 2007;Hollinger et al., 1994;Leuning et al., 2004). Through long-term international collabora-  (Finnigan et al., 2003;Leuning et al., 1982;Webb et al., 1980). High quality, in situ measurements of ecosystem fluxes and stores of water, carbon and nutrients were also being sought to calibrate and validate remotely sensed observations in these unique landscapes and ecosystems. Flux data were also being incorporated into biophysically realistic LSMs, such as the CABLE LSM within Australia's global climate and Earth system model (Australian Community Climate and Earth System Simulator, Ziehn et al., 2020).
These foundational flux tower sites sowed the seeds of OzFlux, which expanded to a continental network when TERN (Terrestrial Ecosystem Research Network) was funded in 2009. This funding provided the capital and institutional investment needed to support the 'hard' infrastructure of around a dozen flux towers and supersites across Australia Karan et al., 2016). Equally important, it provided the dedicated and sustained support for 'soft' infrastructure needs such as training for early career researchers; the data management infrastructure to comply with FAIR data principles (Wilkinson et al., 2016); data curation and data processing to ensure consistency across the network; data quality control and assurance; and data discoverability and data access Isaac et al., 2017).
With the addition of new flux towers in ca. 2010 and the development of integrated data processing systems , OzFlux has run as a truly regional network since 2010. Historically, Australian OzFlux researchers have largely focussed on natural and forested ecosystems, whereas New Zealand OzFlux research has concentrated on greenhouse gas budgets and emissions from agricultural systems, including drained peatlands. The long-term investment in OzFlux has led to significant and diverse research outcomes and impacts as summarised in Figure 2. The following sections explore some of the key lessons and outcomes from OzFlux in more detail, and how they have contributed to global understanding in their respective scientific space.
F I G U R E 2 Summary of the significant scientific and technical outcomes from the OzFlux network after two decades: Blue relates to discovery, information and knowledge outcomes; grey outcomes relate to assessments across site, regional and global scales; yellow refers to the capacity building outcomes for researchers and green indicates technical outcomes for observations and modelling Australian regions have distributions of mean annual precipitation (MAP) variability that are much higher than the rest of the world ( Figure 3), and OzFlux sites measure across a very large range of MAP and in areas with higher MAP co-efficient of variation not captured by FLUXNET sites (Figure 3). Moreover, OzFlux includes sites with a very large spatial range in VPD, greater than 6 kPa (Renchon et al., 2018), allowing exploration of vegetation responses to high VPD that goes well beyond the conditions currently experienced by most ecosystems in the Northern Hemisphere (Grossiord et al., 2020). It is sometimes argued that Australian and New Zealand vegetation and its management is unique, with the implication that it is difficult to use data from these ecosystems to inform our understanding of vegetation function on other continents (see also  and ecosystem respiration (ER) Cleverly et al., 2019;Griebel et al., 2017;Haverd, Ahlström, et al., 2016;Hinko-Najera et al., 2017;Li et al., 2017;Renchon et al., 2018;Xie et al., 2019). They also result in seasonal fluctuations between mild and wet maritime winds and hot and dry continental winds from the Australian mainland. These shifts not only affect plant productivity, but also provide methodological challenges for comparing annual budgets that have been constructed from flux tower observations .
Recent heatwaves during a prolonged drought across south-  (Griebel, Bennett, et al., 2020;van Gorsel et al., 2016). These results highlight that the potential for temperate forests and woodlands to remain net carbon sinks will not only depend on the responses of photosynthesis to warmer temperatures, but also on soil water availability and on the concomitant responses of ER.
High temperatures and associated deficits in atmospheric vapour pressure provide challenges for the ability of plants to regulate water loss and to maintain photosynthesis. A synthesis across 17 OzFlux wooded ecosystems demonstrates strong alignment between the thermal optima of GPP and mean daytime air temperatures, indicating ecosystem scale photosynthesis has adjusted to past thermal regimes (Bennett et al., 2021). Although it currently seems that GPP in Australian broadleaf evergreen forests is buffered against small increases in air temperature, the shape of this relationship and the response of ER to rising temperatures will determine the sustainability of Australian carbon sinks into the future (Bennett et al., 2021;Duffy et al., 2021;Griebel, Bennett, et al., 2020;van Gorsel et al., 2016).  Beringer et al. (2016) showed that only a few had a low GPP/ER ratio, despite several sites in the network with a history of disturbance.
While much of the network was established in undisturbed sites, many have been subject to natural or managed disturbance over the past 20 years. The apparent resilience of these ecosystems to disturbance is an important aspect of their longer-term carbon balance in response to global change, which is discussed further in lesson 4.
Bushfire is one of the most widespread causes of ecosystem disturbance across Australia, having shaped adaptations in vegetation across the continent for over 80 million years, similar to southern Africa and in contrast to the more recent development of fire in the Mediterranean region and the Americas (Carpenter et al., 2015;Cleverly et al., 2019). In tropical Northern Australian mesic savannas, bushfires are frequent, with 30% of the total savanna land area burned annually . This fire regime directly affects carbon emissions and productivity due to canopy loss . Global climate change is expected to further increase extreme fire weather, and thus greenhouse gas emissions, which will further reduce the savanna carbon sink (Beringer et al., 2003;Duvert et al., 2020). By contrast, land management, which reduces fire frequency and intensity (e.g. by shifting fires from the late to the early dry season) is reducing greenhouse gas emissions at landscape scales in the tropical savanna (Edwards et al., 2021).  . By contrast, NEP in mesic tropical savanna ecosystems of northern Australia returns to pre-fire status in 3-4 months post-fire . The knowledge provided from this research into bushfires in Australia, including regional differences between the northern and southern parts of the continent, is important for understanding how these ecosystems adapt to changing climates.
It is particularly useful for determining whether they remain carbon sinks in the long-term as fire frequency and intensity changes, and for informing and improving Earth system models, many of which are poor at simulating fire.
Tropical cyclones largely affect OzFlux sites in northern Australia and occur infrequently, but when they do, they often cause great can be difficult to fill because many but not all disturbances require the serendipity of being in the right place at the right time. This reinforces the need for continuous measurements over many decades, to increase the chances of being in the right place at the right time.

| LE SSON 3-THE EFFEC T OF DROUG HT AND HE AT S TRE SS ON ECOSYS TEM C ARBON AND WATER BAL AN CE S
The primary stress events in natural and managed ecosystems  (Prior et al., 2004) and was used to convert DBH to age. Figure reproduced with permission from Hutley and Beringer (2011) to capture the response of native and managed ecosystems to the occurrence of these emerging trends in interannual and more frequent stress events Moore et al., 2018) (see lessons 1, 4 and 8).
The impact of drought has been particularly evident in semiarid Australia, where ecosystems have shifted from weak CO 2 sinks into CO 2 sources Qiu et al., 2020). The pivot point at which an ecosystem switches from a CO 2 sink to a CO 2 source can depend on the vegetation properties; for example, the Acacia spp. dominated woodland near Alice Springs, in the arid centre of Australia, remain a net CO 2 sink as long as the annual rainfall exceeds 260 mm (site average is 300 mm yr −1 ), whereas the nearby hummock grasslands become a CO 2 source if the annual rainfall falls below the pivot point of 506 mm yr −1 (Tarin, Nolan, Eamus, et al., 2020).
Ecosystems can also respond to drought stress by regulating their water use via phenotypic plasticity as observed in Eucalyptus obliqua at the Wombat State Forest in south-eastern Australia, where leaf water potential at the turgor loss point was lowered through osmotic adjustment during a short-term summer drought (Pritzkow et al., 2020). Other drought response mechanisms include partial drought deciduousness, where LAI is reduced to minimise the surface area for water loss, which also increases the Huber value (ratio of sapwood area to leaf area) during extended drought Pritzkow et al., 2020). Individual species may also behave differently when subject to similar stresses, as shown at Cumberland Plain, where the melaleuca stand maintained higher canopy conductance and transpiration under VPD and moisture stress than the neighbouring eucalypt stand (Griebel, Metzen, Boer, et al., 2020).
Drought events in New Zealand, although less intense than those typically experienced in Australia, can still reduce ecosystem carbon uptake. For example, a short-term meteorological drought turned an intensively grazed dairy pasture into a net CO 2 source (Kim & Kirschbaum, 2015;Kirschbaum et al., 2015;Rutledge et al., 2015). The intensive grazing that characterises these systems regularly removes pasture dry matter. Pasture regrowth and carbon uptake via photosynthesis following grazing is limited during drought conditions, leading to net carbon loss (Kirschbaum et al., 2017;Wall et al., 2019). In contrast to highly managed agroecosystems, native peatland bogs in New Zealand's Waikato region are able to maintain a strong carbon sink even during drought (Goodrich et al., 2017) likely due to ample soil moisture stores.
Temperate and semi-arid ecosystems in Australia display different mechanisms to tolerate prolonged water stress. For Mulga dominated semi-arid ecosystems, extensive expression of ecophysiological adaptations allows survival through decadal scale droughts Eamus et al., 2013;Tarin et al., 2020b) and are usually reliant on single rainfall events to boost their CO 2 uptake .
Temperate ecosystems in non-water limited regions of Australia are able to tolerate several years of below average rainfall through access to greater soil moisture reserves (Griebel, Bennett, et al., 2020;Keith et al., 2012;Kirschbaum et al., 2007). Access to soil moisture reserves helps buffer wet sclerophyll ecosystems against heatwaves, as illustrated by the combined drought and heatwave event in 2012/2013 that led to water-limited woodland ecosystems becoming CO 2 sources due to a reduction in photosynthesis caused by elevated water stress van Gorsel et al., 2016), while wetter forest systems were much less affected . Model analysis of the more recent 2018/2019 heatwave showed reduced productivity for most ecosystems across continental Australia (Qiu et al., 2020). Four sites in southeast Australia also show reduced CO 2 sink strength during this period ( Figure 5). Some of these OzFlux observations are leading to much-needed and rapid improvements in the CABLE LSM to better incorporate groundwater-vegetation interactions (Mu et al., 2021;Mu et al., 2021).
Drought can interact with disturbance (lesson 2) or other stress as was demonstrated at the temperate, wet sclerophyll, managed forest at Tumbarumba, where long-term drought coincided with an insect attack (Kirschbaum et al., 2007). The forest was impacted by this attack, but it became a CO 2 sink again when the insect attack had abated, despite continued and even intensifying drought conditions . A future that consists of more frequent heatwaves in combination with drought could deplete soil moisture reserves beyond the tipping point for many ecosystems and result in greater ecosystem stress.

| LE SSON 4 -ECOSYS TEM RE S ILIEN CE , ADAP TATION AND V ULNER AB ILIT Y TO INTER ANNUAL CLIMATE VARIAB ILIT Y
Ecosystems can be resilient to climate variability by maintaining a stable carbon budget during and shortly following the imposition of stress (Holling, 1973) (Karan et al., 2016). For example, while the strong interannual variability in arid and semi-arid Australian ecosystems reduces productivity, its recovery does not appear to be limited by previous sequences of drought, swinging rapidly between states of net CO 2 source and sink, sometimes from one year to the next Cleverly, Eamus, van Gorsel, et al., 2016;Tarin, Nolan, Medlyn, et al., 2020). Australian ecosystems also show resilience to drought and fire in their leaf phenology. For example, in Australia's mesic savannas, fire usually only consumes the seasonal grassy understorey, whereas canopy trees mostly remain intact (Lehmann et al., 2014). By contrast, in Australia's tropical drylands, a highly resilient leaf phenology allows strong growth during wet years despite the absence of a growing season in previous dry years (Ma et al., 2013). Similarly, Australian tropical rainforest trees are considered to be somewhat resilient to high-temperature stress and heatwaves due to the very high temperature at which leaf dark respiration reaches a peak (60°C) (Weerasinghe et al., 2014), although they may be instead vulnerable to high VPD stresses (Fu et al., 2018). However, a loss of resilience has been predicted for Australian drylands with the increased occurrence of future woody dieback and megadrought events (Ma et al., 2013), and the continued resilience of many ecosystems in Australia and New Zealand is not assured with global change (van Gorsel et al.,

2016).
Other examples of carbon-function resilience to disturbance and drought are evident in managed and natural ecosystems of New Zealand. Here, dairy farm pastures have shown rapid recovery to a net positive carbon balance within one week following intensive grazing events. In these systems, grass is maintained in a continuously juvenile state through repeated grazing and defoliation by cattle (Hunt et al., 2016). In contrast, northern New Zealand's peat-forming wetland ecosystems display resilience through the continuous accumulation of deep peat deposits over millennia, despite existing in a warm maritime climate zone with frequent seasonal water deficits.
In the few remaining intact peat wetlands, resilience to drought is a product of the ecosystem's conservative evaporation regime and highly dynamic peat surface level (Campbell & Williamson, 1997;Fritz et al., 2008), both of which contribute to maintaining a stable and shallow water table, limiting respired CO 2 losses (Goodrich et al., 2017;Ratcliffe et al., 2019). However, imposing artificial drainage diminishes their ability to self-regulate, leading to a shift in ecosystem structure and function, resulting in larger component CO 2 fluxes (Ratcliffe et al., 2019(Ratcliffe et al., , 2020. Furthermore, resilience is completely lost when drained peatlands are used for dairy grazing, where annual CO 2 losses can be extremely large, particularly during dry conditions .   Water fluxes are of critical concern in agroecosystems, where irrigation decisions are informed by balancing crop water use with yield-based revenue, irrigation costs and regulatory limits for nutrient leaching. There are concerns that the practice of irrigation, increasingly widespread in NZ, may lead to net carbon losses, and soil-core sampling studies point in this direction . However, flux measurements over irrigated pasture did not find any carbon losses throughout the three years of measurements (Laubach & Hunt, 2018). In another study, capturing flux measurements over lucerne, it was found that total evaporation and drainage increased in response to irrigation, relative to a nearby non-irrigated lucerne crop, with the benefit of larger biomass production at the cost of Another possible approach lies in the development of low-cost measurement systems (Hill et al., 2016). Communication between disciplines and with industry and policy makers will be central to OzFlux and the global flux community to help transition agricultural practices towards climate-smart food systems.

| LE SSON 6 -ADVAN CE S MADE VIA SYNERG IE S WITH REMOTE S ENS ING
The initiation of OzFlux was shortly preceded by NASA's Earth Observing System (EOS) that introduced the first suite of satellitebased global ecology products for long-term monitoring of ecosystem functioning, phenology, disturbance and plant stress (Xiao et al., 2019). The validity and robustness of these first biophysical products from remote sensing were challenged by the diversity of landscapes and extreme environments of Australia (Hill et al., 2006;Kanniah et al., 2009;Sea et al., 2011). For example, Leuning et al. (2005) reported that the moderate resolution imaging spectrometer (MODIS) LAI product overestimated in-situ LAI more than twofold over the moderately open, wet sclerophyll forest at the Tumbarumba OzFlux site. These native forests are known for their highly clumped crown architecture and vertical leaf inclination angle (Anderson, 1981).
The MODIS GPP product estimated the annual amplitude of tower GPP fluxes quite well but performed less well in estimating the seasonal phase of variation (Leuning et al., 2005). These assessments with relatively high accuracy where ecosystem processes are phenologically driven, such as in Australian wet to dry tropical savannas, grasslands and croplands Glenn et al., 2011;Ma et al., 2013;Moore et al., 2017;Zhang et al., 2008). However, in temperate and Mediterranean evergreen Australian forests/woodlands, the VI and LAI products were seasonally out of phase with GPP and found to be better proxies of photosynthetic 'infrastructure' capacity ( Australia's climate extremes (Ma et al., 2015. Annually during the big wet and reported that semi-arid Australian net CO 2 uptake was highly transient and rapidly dissipated by subsequent drought. The accuracies of the remotely sensed CO 2 retrievals and the atmospheric transport models are approaching the levels needed to constrain CO 2 fluxes to estimate net biome productivity (NBP) from the natural biosphere (Buchwitz et al., 2017;Kondo et al., 2016).
The OzFlux network capitalises on skills and infrastructure through strong collaborations of people both at a national level and through international networks (Figure 2), including SpecNet (Gamon et al., 2006), https://specn et.info/tumba rumba/) and the Australian Phenocam Network (http://pheno cam.org.au/). SpecNet sites are equipped with hyperspectral instruments and play important roles in linking in situ optical measures (fPAR, VIs and SIF) from tower platforms with flux observations, to explore mechanistic and scaling relationships Woodgate et al., 2020).
The phenocam network enables high temporal image-based recognition of understory/overstory dynamics at species levels, and thus enables leaf level demography, ontogeny and phenology analyses (Moore, Brown, et al., 2016;Wu et al., 2016). These sub-daily, nearground spectral and phenocam measurements bridge temporal, spatial and spectral scales with airborne and satellite remotely sensed proxies of canopy and ecosystem function.
Capturing the range of global variability in ecosystems is critical for accurately calibrating, validating and upscaling satellite algorithms and modelled outputs using high-quality ground-level data.
In a global flux tower analysis using MODIS satellite products and meteorological drivers, Tramontana et al. (2016)   . For example, the use of SIF and VIs together can be used to disentangle controls of canopy structure from physiology on GPP (Magney et al., 2019;Springer et al., 2017;Verma et al., 2017).

This is particularly important for evergreen canopies (dominant in
Australia and New Zealand) where GPP is often decoupled from VIs .
The current generation geostationary satellites (e.g. Himawari-8) provide sub-daily, 10-min image acquisition frequencies in near realtime across Australia, enabling integration with diurnal fluxes for refined insights into ecosystem dynamics. A metric of canopy structure, canopy clumping index, was recently retrieved from sub daily measures from the Deep Space Climate Observatory (DSCOVR) satellite and evaluated at OzFlux sites (Pisek et al., 2021). The International Space Station (ISS) has three instruments that provide regionaldiurnal measures of (1) Evapotranspiration from the ECOsystem

| LE SSON 7-ADVAN CE S MADE VIA SYNERG IE S WITH MODELLING
One of the most important outcomes from OzFlux has been the ability to constrain models used to quantify and predict terrestrial carbon and water fluxes, from site-scales (Kirschbaum et al., 2007 to the continent (Decker, 2015), using multi-annual, continuous data from around Australia and sampling a range of bioclimates. Foremost among these outcomes was the construction of a full continental carbon budget for Australia (Haverd, Raupach, Briggs, Canadell, Davis, et al., 2013). This work used multiple data sources, including OzFlux data, to constrain the CABLE LSM (Wang et al., 2011). The data-constrained estimate of Australia's NBP for 1990-2011 was 36 ± 29 Tg C yr −1 (Haverd, Raupach, Briggs, Canadell, Davis, 2013;Haverd, Raupach, Briggs, Canadell, Isaac, et al., 2013), with annual net primary productivity (NPP) quantified at 2.2 ± 0.4 Pg C yr −1 .
Similarly, OzFlux data underpin operational water modelling in  A key question relates to how the carbon and water cycles will change in the future; answering this will require longevity across the OzFlux and the wider FLUXNET network.

| LE SSON 8 -THE IMP ORTAN CE OF LONG -TERM ME A SUREMENTS TO DE TEC T DEC ADAL SC ALE E VENTS AND CLIMATE CHANG E EFFEC TS
Given the geographical extent of the Australian and New Zealand regions and the associated large range of climate drivers, climatic variability is naturally high (King et al., 2020), and this variability is increasing due to changes in climate and land use (Head et al., 2014;King et al., 2020). Regional climate variability is also driven by complex, large-scale ocean-atmosphere influences that operate at frequencies from weeks to decades and have a strong influence on rainfall (King et al., 2020;, and therefore drives variability of ecosystem dynamics ) (See also lesson 2). The net result is a climate system which operates in widely varying states spatially and temporally, driving periods of drought, flood and heatwaves (Freund et al., 2017;Kiem et al., 2016;Perkins-Kirkpatrick et al., 2016) that are increasing in severity with climate change (Cai et al., 2014(Cai et al., , 2021. Extreme events have a disproportionate effect on annual carbon exchange at regional to continental scales (Zscheischler et al., 2017) and long-term monitoring of ecosystem carbon exchange, water use and resource use efficiency is required to understand and predict ecosystem responses to the changing climatic range. This is particularly important in Australia, which is a global hot spot for variability-especially in semi-arid ecosystems, which exhibit large and 'asymmetrical' responses of GPP to rainfall variability (Haverd, Ahlström, et al., 2016). This large interannual variability makes detecting long-term trends from short records extremely difficult (Baldocchi et al., 2018). On the other hand, Australia may also provide an example to inform other continents about how ecosystems will adapt to increased climate variability with resource availability hard to predict.
A comprehensive understanding of interannual and interdecadal variability of the carbon cycle and its drivers requires long-term data (>50 years) (Fu et al., 2019;He et al., 2019;Jung et al., 2017;von Buttlar et al., 2018;Zscheischler et al., 2016). Continued operation of existing sites and the expansion of the global eddy covariance monitoring network (Baldocchi, 2019), together with the increasing length of the satellite record, will provide the observational constraints to gain this understanding. The two decades of observations in the OzFlux network span several significant ENSO events (Figure 7), and this length of record can be used to detect change in ecosystem properties as a function of short-term or highfrequency disturbances such as fire, insect attack, drought and cyclones Hutley et al., 2013;Keith et al., 2012).  (Eamus et al., 2001;. Frequent savanna fires (2 in 3 years) scorch the woody canopy and post-fire canopy reconstruction results in high respiratory losses (Cernusak et al., 2006) with the ecosystem a net source of CO 2 for months after fire, whereas LE recovers within weeks . This post-fire recovery phase is a period of lower WUE, and the savanna ecosystem has a lower-than-expected WUE because of these ecosystem characteristics.
Trends in WUE and RUE are highly statistically significant at AU-Tum (p < 0.01), and WUE increased by 16% over 18 years, whereas the tropical savanna site only increased by 6% (Figure 8). Over the period of observation, atmospheric CO 2 concentrations increased by about 10%, and the trend in WUE at AU-How is consistent with theoretical expectations of increased photosynthesis and WUE (Kirschbaum & McMillan, 2018;Walker et al., 2021)  . As such there is a growing imperative to use and build on our knowledge of ecosystem processes and emergent phenomena (Karan et al., 2016). These processes must be studied across a range of temporal and spatial scales to be properly understood and integrated into modelling. Synergistic network science has allowed these emergent processes to be un- Ecosystems are expected to experience continued long-term climate change and greater variability along with increased disturbance leading to a loss of ecosystem services. To best maintain our ecosystems and their services, we must anticipate and plan for these changes using predictive modelling and ecological forecasting. Developing this capability is crucial and will require forecasting (over the near term) and projections over multidecadal time scales) using real-time flux and soil at a given site, rather than these factors in isolation which control the risk of drought mortality (Feng et al., 2018(Feng et al., , 2019. In the future, measurements of hydraulic traits across the OzFlux network (Peters et al., 2021), coupled with eddy covariance data, could facilitate the development and testing of new theories governing plant controls on transpiration.
A significant proportion of Australia's total ecosystem biomass (ca. 30%-50%, Spawn et al., 2020) is found in the subsurface, yet our understanding of how the subsurface environment changes and influences ecosystems is lagging. Newly funded critical zone observatories (CZOs), co-located at several OzFlux sites, are now installing the F I G U R E 8 Timeseries of observed ecosystem water use and radiation use efficiency from two OzFlux sites with 20-year records: tropical savanna at the Howard Springs site and temperate Eucalypt forest at the Tumbarumba site. Trend lines are given for significant time series (p < 0.05) using the non-parametric Mann Kendal test equipment to monitor water, carbon and energy throughout deep soil profiles. By integrating observations of subsurface variation with the surface fluxes measured by OzFlux, these CZOs will offer better understanding of the interdependencies of carbon and water cycles across timescales and across the full vertical span of Australian ecosystems.
Ecosystem observatories are moving beyond CO 2 and water cycles to monitoring other greenhouse gases, especially emissions of CH 4 from wetlands and N 2 O from agricultural systems as highlighted in the lessons above. These potent greenhouse gases can now be measured at temporal and spatial scales that are relevant to land management and planning for mitigation of climate change.
There is currently a high demand for new researchers with skills in environmental monitoring, sensors and data analysis; however, it is a challenge to sustain training of postgraduate students and our capacity in the discipline of global change biology. Recruitment of new talent needs to start at the undergraduate level or earlier, to ensure a flow of quantitatively skilled researchers who are passionate about ecosystem science. Educational collaborations among engineers, atmospheric scientists, hydrologists, ecologists, physicists and others will set the stage for the next generation of environmental leadership and stewardship. OzFlux will continue to play a major role in training this next generation and in providing the ecosystem data which scientists, the public and managers/government can rely on in understanding our rapidly changing environment in Australia and New Zealand.

ACK N OWLED G EM ENTS
We dedicate this paper to the memory of Dr Vanessa Haverd, who

DATA AVA I L A B I L I T Y S TAT E M E N T
The OzFlux data that support the findings of this study are openly avail- ). Data to support production of Figure 5 were also provided by the Australian Bureau of Meteorology via www.bom.gov.au/climate/enso/enlist.