Search

Spatially explicit prediction of soil organic carbon (SOC) serves as a crucial foundation for effective land management strategies aimed at mitigating soil degradation and assessing carbon sequestration potential. Here, using more than 1000 in situ observations, we trained two machine learning models (a random forest model and a k-means coupled with multiple linear regression model) and one process-based model (the vertically resolved MIcrobial-MIneral Carbon Stabilization, MIMICS, model) to predict the SOC stocks of the top 30 cm of soil in Australia. Parameters of MIMICS were optimised for different site groupings using two distinct approaches: plant functional types (MIMICS-PFT) and the most influential environmental factors (MIMICS-ENV). All models showed good performance with respect to SOC predictions, with an R2 value greater than 0.8 during out-of-sample validation, with random forest being the most accurate; moreover, it was found that SOC in forests is more predictable than that in non-forest soils excluding croplands. The performance of continental-scale SOC predictions by MIMICS-ENV is better than that by MIMICS-PFT especially in non-forest soils. Digital maps of terrestrial SOC stocks generated using all of the models showed a similar spatial distribution, with higher values in south-eastern and south-western Australia, but the magnitude of the estimated SOC stocks varied. The mean ensemble estimate of SOC stocks was 30.3 t ha−1, with k-means coupled with multiple linear regression generating the highest estimate (mean SOC stocks of 38.15 t ha−1) and MIMICS-PFT generating the lowest estimate (mean SOC stocks of 24.29 t ha−1). We suggest that enhancing process-based models to incorporate newly identified drivers that significantly influence SOC variation in different environments could be the key to reducing the discrepancies in these estimates. Our findings underscore the considerable uncertainty in SOC estimates derived from different modelling approaches and emphasise the importance of rigorous out-of-sample validation before applying any one approach in Australia.

Climate projections from global circulation models (GCMs), part of the Coupled Model Intercomparison Project 6 (CMIP6), are often employed to study the impact of future climate on ecosystems. However, especially at regional scales, climate projections display large biases in key forcing variables such as temperature and precipitation. These biases have been identified as a major source of uncertainty in carbon cycle projections, hampering predictive capacity. In this study, we open the proverbial Pandora's box and peer under the lid of strategies to tackle climate model ensemble uncertainty. We employ a dynamic global vegetation model (LPJ-GUESS) and force it with raw output from CMIP6 to assess the uncertainty associated with the choice of climate forcing. We then test different methods to either bias-correct or calculate ensemble averages over the original forcing data to reduce the climate-driven uncertainty in the regional projection of the Australian carbon cycle. We find that all bias correction methods reduce the bias of continental averages of steady-state carbon variables. Bias correction can improve model carbon outputs, but carbon pools are insensitive to the type of bias correction method applied for both individual GCMs and the arithmetic ensemble average across all corrected models. None of the bias correction methods consistently improve the change in simulated carbon over time compared to the target dataset, highlighting the need to account for temporal properties in correction or ensemble-averaging methods. Multivariate bias correction methods tend to reduce the uncertainty more than univariate approaches, although the overall magnitude is similar. Even after correcting the bias in the meteorological forcing dataset, the simulated vegetation distribution presents different patterns when different GCMs are used to drive LPJ-GUESS. Additionally, we found that both the weighted ensemble-averaging and random forest approach reduce the bias in total ecosystem carbon to almost zero, clearly outperforming the arithmetic ensemble-averaging method. The random forest approach also produces the results closest to the target dataset for the change in the total carbon pool, seasonal carbon fluxes, emphasizing that machine learning approaches are promising tools for future studies. This highlights that, where possible, an arithmetic ensemble average should be avoided. However, potential target datasets that would facilitate the application of machine learning approaches, i.e., that cover both the spatial and temporal domain required to derive a robust informed ensemble average, are sparse for ecosystem variables.

The vegetation's response to climate change is a significant source of uncertainty in future terrestrial biosphere model projections. Constraining climate–carbon cycle feedbacks requires improving our understanding of both the immediate and long-term plant physiological responses to climate. In particular, the timescales and strength of memory effects arising from both extreme events (i.e. droughts and heatwaves) and structural lags in the systems (such as delays between rainfall and peak plant water content or between a precipitation deficit and down-regulation of productivity) have largely been overlooked in the development of terrestrial biosphere models. This is despite the knowledge that plant responses to climatic drivers occur across multiple timescales (seconds to decades), with the impact of climate extremes resonating for many years. Using data from 12 eddy covariance sites, covering two rainfall gradients (256 to 1491 mm yr−1) in Australia, in combination with a hierarchical Bayesian model, we characterised the timescales and magnitude of influence of antecedent drivers on daily net ecosystem exchange (NEE) and latent heat flux (λE). By focussing our analysis on a single continent (and predominately on a single genus), we reduced the degrees of variation between each site, providing a novel chance to explore the unique characteristics that might drive the importance of memory. Model fit varied considerably across sites when modelling NEE, with R2 values of between 0.30 and 0.83. λE was considerably more predictable across sites, with R2 values ranging from 0.56 to 0.93. When considered at a continental scale, both fluxes were more predictable when memory effects (expressed as lagged climate predictors) were included in the model. These memory effects accounted for an average of 17 % of the NEE predictability and 15 % for λE. Consistent with prior studies, the importance of environmental memory in predicting fluxes increased as site water availability declined (ρ=-0.73, p for NEE, ρ=-0.67, p for λE). However, these relationships did not necessarily hold when sites were grouped by vegetation type. We also tested a model of k-means clustering plus regression to confirm the suitability of the Bayesian model for modelling these sites. The k-means approach performed similarly to the Bayesian model in terms of model fit, demonstrating the robustness of the Bayesian framework for exploring the role of environmental memory. Our results underline the importance of capturing memory effects in models used to project future responses to climate change, especially in water-limited ecosystems. Finally, we demonstrate a considerable variation in individual-site predictability, driven to a notable degree by environmental memory, and this should be considered when evaluating model performance across ecosystems.

Eddy covariance flux towers measure the exchange of water, energy, and carbon fluxes between the land and atmosphere. They have become invaluable for theory development and evaluating land models. However, flux tower data as measured (even after site post-processing) are not directly suitable for land surface modelling due to data gaps in model forcing variables, inappropriate gap-filling, formatting, and varying data quality. Here we present a quality-control and data-formatting pipeline for tower data from FLUXNET2015, La Thuile, and OzFlux syntheses and the resultant 170-site globally distributed flux tower dataset specifically designed for use in land modelling. The dataset underpins the second phase of the Protocol for the Analysis of Land Surface Models (PALS) Land Surface Model Benchmarking Evaluation Project (PLUMBER), an international model intercomparison project encompassing >20 land surface and biosphere models. The dataset is provided in the Assistance for Land-surface Modelling Activities (ALMA) NetCDF format and is CF-NetCDF compliant. For forcing land surface models, the dataset provides fully gap-filled meteorological data that have had periods of low data quality removed. Additional constraints required for land models, such as reference measurement heights, vegetation types, and satellite-based monthly leaf area index estimates, are also included. For model evaluation, the dataset provides estimates of key water, carbon, and energy variables, with the latent and sensible heat fluxes additionally corrected for energy balance closure. The dataset provides a total of 1040 site years covering the period 1992–2018, with individual sites spanning from 1 to 21 years. The dataset is available at http://doi.org/10.25914/5fdb0902607e1 (Ukkola et al., 2021).

In response to a warming climate, temperature extremes are changing in many regions of the world. Therefore, understanding how the fluxes of sensible heat, latent heat and net ecosystem exchange respond and contribute to these changes is important. We examined 216 sites from the open access Tier 1 FLUXNET2015 and free fair-use La Thuile data sets, focussing only on observed (non-gap-filled) data periods. We examined the availability of sensible heat, latent heat and net ecosystem exchange observations coincident in time with measured temperature for all temperatures, and separately for the upper and lower tail of the temperature distribution, and expressed this availability as a measurement ratio. We showed that the measurement ratios for both sensible and latent heat fluxes are generally lower (0.79 and 0.73 respectively) than for temperature measurements, and the measurement ratio of net ecosystem exchange measurements are appreciably lower (0.42). However, sites do exist with a high proportion of measured sensible and latent heat fluxes, mostly over the United States, Europe and Australia. Few sites have a high proportion of measured fluxes at the lower tail of the temperature distribution over very cold regions (e.g. Alaska, Russia) or at the upper tail in many warm regions (e.g. Central America and the majority of the Mediterranean region), and many of the world's coldest and hottest regions are not represented in the freely available FLUXNET data at all (e.g. India, the Gulf States, Greenland and Antarctica). However, some sites do provide measured fluxes at extreme temperatures, suggesting an opportunity for the FLUXNET community to share strategies to increase measurement availability at the tails of the temperature distribution. We also highlight a wide discrepancy between the measurement ratios across FLUXNET sites that is not related to the actual temperature or rainfall regimes at the site, which we cannot explain. Our analysis provides guidance to help select eddy covariance sites for researchers interested in understanding and/or modelling responses to temperature extremes.

The rationale for using multi-model ensembles in climate change projections and impacts research is often based on the expectation that different models constitute independent estimates; therefore, a range of models allows a better characterisation of the uncertainties in the representation of the climate system than a single model. However, it is known that research groups share literature, ideas for representations of processes, parameterisations, evaluation data sets and even sections of model code. Thus, nominally different models might have similar biases because of similarities in the way they represent a subset of processes, or even be near-duplicates of others, weakening the assumption that they constitute independent estimates. If there are near-replicates of some models, then treating all models equally is likely to bias the inferences made using these ensembles. The challenge is to establish the degree to which this might be true for any given application. While this issue is recognised by many in the community, quantifying and accounting for model dependence in anything other than an ad-hoc way is challenging. Here we present a synthesis of the range of disparate attempts to define, quantify and address model dependence in multi-model climate ensembles in a common conceptual framework, and provide guidance on how users can test the efficacy of approaches that move beyond the equally weighted ensemble. In the upcoming Coupled Model Intercomparison Project phase 6 (CMIP6), several new models that are closely related to existing models are anticipated, as well as large ensembles from some models. We argue that quantitatively accounting for dependence in addition to model performance, and thoroughly testing the effectiveness of the approach used will be key to a sound interpretation of the CMIP ensembles in future scientific studies.

The FLUXNET dataset contains eddy covariance measurements from across the globe and represents an invaluable estimate of the fluxes of energy, water, and carbon between the land surface and the atmosphere. While there is an expectation that the broad range of site characteristics in FLUXNET result in a diversity of flux behaviour, there has been little exploration of how predictable site behaviour is across the network. Here, 155 datasets with 30 min temporal resolution from the Tier 1 of FLUXNET 2015 were analysed in a first attempt to assess individual site predictability. We defined site uniqueness as the disparity in performance between multiple empirical models trained globally and locally for each site and used this along with the mean performance as measures of predictability. We then tested how strongly uniqueness was determined by various site characteristics, including climatology, vegetation type, and data quality. The strongest determinant of predictability appeared to be that drier sites tended to be more unique. We found very few other clear predictors of uniqueness across different sites, in particular little evidence that flux behaviour was well discretised by vegetation type. Data length and quality also appeared to have little impact on uniqueness. While this result might relate to our definition of uniqueness, we argue that our approach provides a useful basis for site selection in LSM evaluation, and we invite critique and development of the methodology.

End users studying impacts and risks caused by human-induced climate change are often presented with large multi-model ensembles of climate projections whose composition and size are arbitrarily determined. An efficient and versatile method that finds a subset which maintains certain key properties from the full ensemble is needed, but very little work has been done in this area. Therefore, users typically make their own somewhat subjective subset choices and commonly use the equally weighted model mean as a best estimate. However, different climate model simulations cannot necessarily be regarded as independent estimates due to the presence of duplicated code and shared development history. Here, we present an efficient and flexible tool that makes better use of the ensemble as a whole by finding a subset with improved mean performance compared to the multi-model mean while at the same time maintaining the spread and addressing the problem of model interdependence. Out-of-sample skill and reliability are demonstrated using model-as-truth experiments. This approach is illustrated with one set of optimisation criteria but we also highlight the flexibility of cost functions, depending on the focus of different users. The technique is useful for a range of applications that, for example, minimise present-day bias to obtain an accurate ensemble mean, reduce dependence in ensemble spread, maximise future spread, ensure good performance of individual models in an ensemble, reduce the ensemble size while maintaining important ensemble characteristics, or optimise several of these at the same time. As in any calibration exercise, the final ensemble is sensitive to the metric, observational product, and pre-processing steps used.

Previous research has shown that land surface models (LSMs) are performing poorly when compared with relatively simple empirical models over a wide range of metrics and environments. Atmospheric driving data appear to provide information about land surface fluxes that LSMs are not fully utilising. Here, we further quantify the information available in the meteorological forcing data that are used by LSMs for predicting land surface fluxes, by interrogating FLUXNET data, and extending the benchmarking methodology used in previous experiments. We show that substantial performance improvement is possible for empirical models using meteorological data alone, with no explicit vegetation or soil properties, thus setting lower bounds on a priori expectations on LSM performance. The process also identifies key meteorological variables that provide predictive power. We provide an ensemble of empirical benchmarks that are simple to reproduce and provide a range of behaviours and predictive performance, acting as a baseline benchmark set for future studies. We reanalyse previously published LSM simulations and show that there is more diversity between LSMs than previously indicated, although it remains unclear why LSMs are broadly performing so much worse than simple empirical models.

The savanna complex is a highly diverse global biome that occurs within the seasonally dry tropical to sub-tropical equatorial latitudes and are structurally and functionally distinct from grasslands and forests. Savannas are open-canopy environments that encompass a broad demographic continuum, often characterised by a changing dominance between C3-tree and C4-grass vegetation, where frequent environmental disturbances such as fire modulates the balance between ephemeral and perennial life forms. Climate change is projected to result in significant changes to the savanna floristic structure, with increases to woody biomass expected through CO2 fertilisation in mesic savannas and increased tree mortality expected through increased rainfall interannual variability in xeric savannas. The complex interaction between vegetation and climate that occurs in savannas has traditionally challenged terrestrial biosphere models (TBMs), which aim to simulate the interaction between the atmosphere and the land surface to predict responses of vegetation to changing in environmental forcing. In this review, we examine whether TBMs are able to adequately represent savanna fluxes and what implications potential deficiencies may have for climate change projection scenarios that rely on these models. We start by highlighting the defining characteristic traits and behaviours of savannas, how these differ across continents and how this information is (or is not) represented in the structural framework of many TBMs. We highlight three dynamic processes that we believe directly affect the water use and productivity of the savanna system: phenology, root-water access and fire dynamics. Following this, we discuss how these processes are represented in many current-generation TBMs and whether they are suitable for simulating savanna fluxes.Finally, we give an overview of how eddy-covariance observations in combination with other data sources can be used in model benchmarking and intercomparison frameworks to diagnose the performance of TBMs in this environment and formulate road maps for future development. Our investigation reveals that many TBMs systematically misrepresent phenology, the effects of fire and root-water access (if they are considered at all) and that these should be critical areas for future development. Furthermore, such processes must not be static (i.e. prescribed behaviour) but be capable of responding to the changing environmental conditions in order to emulate the dynamic behaviour of savannas. Without such developments, however, TBMs will have limited predictive capability in making the critical projections needed to understand how savannas will respond to future global change.

Flux towers measure ecosystem-scale surface–atmosphere exchanges of energy, carbon dioxide and water vapour. The network of flux towers now encompasses ∼ 900 sites, spread across every continent. Consequently, these data have become an essential benchmarking tool for land surface models (LSMs). However, these data as released are not immediately usable for driving, evaluating and benchmarking LSMs. Flux tower data must first be transformed into a LSM-readable file format, a process which involves changing units, screening missing data and varying degrees of additional gap-filling. All of this often leads to an under-utilisation of these data in model benchmarking. To resolve some of these issues, and to help make flux tower measurements more widely used, we present a reproducible, open-source R package that transforms the FLUXNET2015 and La Thuile data releases into community standard NetCDF files that are directly usable by LSMs. We note that these data would also be useful for any other user or community seeking to independently quality control, gap-fill or use the FLUXNET data.

The savanna ecosystem is one of the most dominant and complex terrestrial biomes, deriving from a distinct vegetative surface comprised of co-dominant tree and grass populations. While these two vegetation types co-exist functionally, demographically they are not static but are dynamically changing in response to environmental forces such as annual fire events and rainfall variability. Modelling savanna environments with the current generation of terrestrial biosphere models (TBMs) has presented many problems, particularly describing fire frequency and intensity, phenology, leaf biochemistry of C3 and C4 photosynthesis vegetation, and root-water uptake. In order to better understand why TBMs perform so poorly in savannas, we conducted a model inter-comparison of six TBMs and assessed their performance at simulating latent energy (LE) and gross primary productivity (GPP) for five savanna sites along a rainfall gradient in northern Australia. Performance in predicting LE and GPP was measured using an empirical benchmarking system, which ranks models by their ability to utilise meteorological driving information to predict the fluxes. On average, the TBMs performed as well as a multi-linear regression of the fluxes against solar radiation, temperature and vapour pressure deficit but were outperformed by a more complicated nonlinear response model that also included the leaf area index (LAI). This identified that the TBMs are not fully utilising their input information effectively in determining savanna LE and GPP and highlights that savanna dynamics cannot be calibrated into models and that there are problems in underlying model processes. We identified key weaknesses in a model's ability to simulate savanna fluxes and their seasonal variation, related to the representation of vegetation by the models and root-water uptake. We underline these weaknesses in terms of three critical areas for development. First, prescribed tree-rooting depths must be deep enough, enabling the extraction of deep soil-water stores to maintain photosynthesis and transpiration during the dry season. Second, models must treat grasses as a co-dominant interface for water and carbon exchange rather than a secondary one to trees. Third, models need a dynamic representation of LAI that encompasses the dynamic phenology of savanna vegetation and its response to rainfall interannual variability. We believe that this study is the first to assess how well TBMs simulate savanna ecosystems and that these results will be used to improve the representation of savannas ecosystems in future global climate model studies.

We implement a new stomatal conductance scheme, based on the optimality approach, within the Community Atmosphere Biosphere Land Exchange (CABLEv2.0.1) land surface model. Coupled land–atmosphere simulations are then performed using CABLEv2.0.1 within the Australian Community Climate and Earth Systems Simulator (ACCESSv1.3b) with prescribed sea surface temperatures. As in most land surface models, the default stomatal conductance scheme only accounts for differences in model parameters in relation to the photosynthetic pathway but not in relation to plant functional types. The new scheme allows model parameters to vary by plant functional type, based on a global synthesis of observations of stomatal conductance under different climate regimes over a wide range of species. We show that the new scheme reduces the latent heat flux from the land surface over the boreal forests during the Northern Hemisphere summer by 0.5–1.0 mm day−1. This leads to warmer daily maximum and minimum temperatures by up to 1.0 °C and warmer extreme maximum temperatures by up to 1.5 °C. These changes generally improve the climate model's climatology of warm extremes and improve existing biases by 10–20 %. The bias in minimum temperatures is however degraded but, overall, this is outweighed by the improvement in maximum temperatures as there is a net improvement in the diurnal temperature range in this region. In other regions such as parts of South and North America where ACCESSv1.3b has known large positive biases in both maximum and minimum temperatures (~ 5 to 10 °C), the new scheme degrades this bias by up to 1 °C. We conclude that, although several large biases remain in ACCESSv1.3b for temperature extremes, the improvements in the global climate model over large parts of the boreal forests during the Northern Hemisphere summer which result from the new stomatal scheme, constrained by a global synthesis of experimental data, provide a valuable advance in the long-term development of the ACCESS modelling system.

Stomatal conductance (gs) affects the fluxes of carbon, energy and water between the vegetated land surface and the atmosphere. We test an implementation of an optimal stomatal conductance model within the Community Atmosphere Biosphere Land Exchange (CABLE) land surface model (LSM). In common with many LSMs, CABLE does not differentiate between gs model parameters in relation to plant functional type (PFT), but instead only in relation to photosynthetic pathway. We constrained the key model parameter "g1", which represents plant water use strategy, by PFT, based on a global synthesis of stomatal behaviour. As proof of concept, we also demonstrate that the g1 parameter can be estimated using two long-term average (1960–1990) bioclimatic variables: (i) temperature and (ii) an indirect estimate of annual plant water availability. The new stomatal model, in conjunction with PFT parameterisations, resulted in a large reduction in annual fluxes of transpiration (~ 30% compared to the standard CABLE simulations) across evergreen needleleaf, tundra and C4 grass regions. Differences in other regions of the globe were typically small. Model performance against upscaled data products was not degraded, but did not noticeably reduce existing model–data biases. We identified assumptions relating to the coupling of the vegetation to the atmosphere and the parameterisation of the minimum stomatal conductance as areas requiring further investigation in both CABLE and potentially other LSMs. We conclude that optimisation theory can yield a simple and tractable approach to predicting stomatal conductance in LSMs.

Recent studies have identified the first-order representation of microbial decomposition as a major source of uncertainty in simulations and projections of the terrestrial carbon balance. Here, we use a reduced complexity model representative of current state-of-the-art models of soil organic carbon decomposition. We undertake a systematic sensitivity analysis to disentangle the effect of the time-invariant baseline residence time (k) and the sensitivity of microbial decomposition to temperature (Q10) on soil carbon dynamics at regional and global scales. Our simulations produce a range in total soil carbon at equilibrium of ~ 592 to 2745 Pg C, which is similar to the ~ 561 to 2938 Pg C range in pre-industrial soil carbon in models used in the fifth phase of the Coupled Model Intercomparison Project (CMIP5). This range depends primarily on the value of k, although the impact of Q10 is not trivial at regional scales. As climate changes through the historical period, and into the future, k is primarily responsible for the magnitude of the response in soil carbon, whereas Q10 determines whether the soil remains a sink, or becomes a source in the future mostly by its effect on mid-latitude carbon balance. If we restrict our simulations to those simulating total soil carbon stocks consistent with observations of current stocks, the projected range in total soil carbon change is reduced by 42% for the historical simulations and 45% for the future projections. However, while this observation-based selection dismisses outliers, it does not increase confidence in the future sign of the soil carbon feedback. We conclude that despite this result, future estimates of soil carbon and how soil carbon responds to climate change should be more constrained by available data sets of carbon stocks.

Soil carbon storage simulated by the Coupled Model Intercomparison Project (CMIP5) models varies 6-fold for the present day. Here, we confirm earlier work showing that this range already exists at the beginning of the CMIP5 historical simulations. We additionally show that this range is largely determined by the response of microbial decomposition during each model's spin-up procedure from initialization to equilibration. The 6-fold range in soil carbon, once established prior to the beginning of the historical period (and prior to the beginning of a CMIP5 simulation), is then maintained through the present and to 2100 almost unchanged even under a strong business-as-usual emissions scenario. We therefore highlight that a commonly ignored part of CMIP5 analyses – the land surface state achieved through the spin-up procedure – can be important for determining future carbon storage and land surface fluxes. We identify the need to better constrain the outcome of the spin-up procedure as an important step in reducing uncertainty in both projected soil carbon and land surface fluxes in CMIP5 transient simulations.

Reliable projections of future climate require land–atmosphere carbon (C) fluxes to be represented realistically in Earth system models (ESMs). There are several sources of uncertainty in how carbon is parameterised in these models. First, while interactions between the C, nitrogen (N) and phosphorus (P) cycles have been implemented in some models, these lead to diverse changes in land–atmosphere fluxes. Second, while the first-order parameterisation of soil organic matter decomposition is similar between models, formulations of the control of the soil physical state on microbial activity vary widely. For the first time, we address these sources of uncertainty simultaneously by implementing three soil moisture and three soil temperature respiration functions in an ESM that can be run with three degrees of biogeochemical nutrient limitation (C-only, C and N, and C and N and P). All 27 possible combinations of response functions and biogeochemical mode are equilibrated before transient historical (1850–2005) simulations are performed. As expected, implementing N and P limitation reduces the land carbon sink, transforming some regional sinks into net sources over the historical period. Meanwhile, regardless of which nutrient mode is used, various combinations of response functions imply a two-fold difference in the net ecosystem accumulation and a four-fold difference in equilibrated total soil C. We further show that regions with initially larger pools are more likely to become carbon sources, especially when nutrient availability limits the response of primary production to increasing atmospheric CO2. Simulating changes in soil C content therefore critically depends on both nutrient limitation and the choice of respiration functions.

The ocean's role in modulating the observed 1–7 Pg C yr−1 inter-annual variability in atmospheric CO2 growth rate is an important, but poorly constrained process due to current spatio-temporal limitations in ocean carbon measurements. Here, we investigate and develop a non-linear empirical approach to predict inorganic CO2 concentrations (total carbon dioxide (CT) and total alkalinity (AT)) in the global ocean mixed layer from hydrographic properties (temperature, salinity, dissolved oxygen and nutrients). The benefit of this approach is that once the empirical relationship is established, it can be applied to hydrographic datasets that have better spatio-temporal coverage, and therefore provide an additional constraint to diagnose ocean carbon dynamics globally. Previous empirical approaches have employed multiple linear regressions (MLR) and relied on ad hoc geographic and temporal partitioning of carbon data to constrain complex global carbon dynamics in the mixed layer. Synthesizing a new global CT/AT carbon bottle dataset consisting of ~33 000 measurements in the open ocean mixed layer, we develop a neural network based approach to better constrain the non-linear carbon system. The approach classifies features in the global biogeochemical dataset based on their similarity and homogeneity in a self-organizing map (SOM; Kohonen, 1988). After the initial SOM analysis, which includes geographic constraints, we apply a local linear optimizer to the neural network, which considerably enhances the predictive skill of the new approach. We call this new approach SOMLO, or self-organizing multiple linear output. Using independent bottle carbon data, we compare a traditional MLR analysis to our SOMLO approach to capture the spatial CT and AT distributions. We find the SOMLO approach improves predictive skill globally by 19% for CT, with a global capacity to predict CT to within 10.9 μmol kg−1 (9.2 μmol kg−1 for AT). The non-linear SOMLO approach is particularly powerful in complex but important regions like the Southern Ocean, North Atlantic and equatorial Pacific, where residual standard errors were reduced between 25 and 40% over traditional linear methods. We further test the SOMLO technique using the Bermuda Atlantic time series (BATS) and Hawaiian ocean time series (HOT) datasets, where hydrographic data was capable of explaining 90% of the seasonal cycle and inter-annual variability at those multi-decadal time-series stations.

Land models, which have been developed by the modeling community in the past few decades to predict future states of ecosystems and climate, have to be critically evaluated for their performance skills of simulating ecosystem responses and feedback to climate change. Benchmarking is an emerging procedure to measure performance of models against a set of defined standards. This paper proposes a benchmarking framework for evaluation of land model performances and, meanwhile, highlights major challenges at this infant stage of benchmark analysis. The framework includes (1) targeted aspects of model performance to be evaluated, (2) a set of benchmarks as defined references to test model performance, (3) metrics to measure and compare performance skills among models so as to identify model strengths and deficiencies, and (4) model improvement. Land models are required to simulate exchange of water, energy, carbon and sometimes other trace gases between the atmosphere and land surface, and should be evaluated for their simulations of biophysical processes, biogeochemical cycles, and vegetation dynamics in response to climate change across broad temporal and spatial scales. Thus, one major challenge is to select and define a limited number of benchmarks to effectively evaluate land model performance. The second challenge is to develop metrics of measuring mismatches between models and benchmarks. The metrics may include (1) a priori thresholds of acceptable model performance and (2) a scoring system to combine data–model mismatches for various processes at different temporal and spatial scales. The benchmark analyses should identify clues of weak model performance to guide future development, thus enabling improved predictions of future states of ecosystems and climate. The near-future research effort should be on development of a set of widely acceptable benchmarks that can be used to objectively, effectively, and reliably evaluate fundamental properties of land models to improve their prediction performance skills.

This work examines different conceptions of land surface model benchmarking and the importance of internationally standardized evaluation experiments that specify data sets, variables, metrics and model resolutions. It additionally demonstrates how essential the definition of a priori expectations of model performance can be, based on the complexity of a model and the amount of information being provided to it, and gives an example of how these expectations might be quantified. Finally, the Protocol for the Analysis of Land Surface models (PALS) is introduced – a free, online land surface model benchmarking application that is structured to meet both of these goals.

The CSIRO Mk3L climate system model, a reduced-resolution coupled general circulation model, has previously been described in this journal. The model is configured for millennium scale or multiple century scale simulations. This paper reports the impact of replacing the relatively simple land surface scheme that is the default parameterisation in Mk3L with a sophisticated land surface model that simulates the terrestrial energy, water and carbon balance in a physically and biologically consistent way. An evaluation of the new model's near-surface climatology highlights strengths and weaknesses, but overall the atmospheric variables, including the near-surface air temperature and precipitation, are simulated well. The impact of the more sophisticated land surface model on existing variables is relatively small, but generally positive. More significantly, the new land surface scheme allows an examination of surface carbon-related quantities including net primary productivity which adds significantly to the capacity of Mk3L. Overall, results demonstrate that this reduced-resolution climate model is a good foundation for exploring long time scale phenomena. The addition of the more sophisticated land surface model enables an exploration of important Earth System questions including land cover change and abrupt changes in terrestrial carbon storage.