Human actions are perturbing the Earth system beyond its planetary boundaries, particularly in terms of biodiversity, climate, and flows of phosphorus and nitrogen, while we also need to provide sustainable social livelihoods across the globe (Fanning et al., 2022; Lade et al., 2020; Steffen et al., 2015). Agricultural production is a main driver of environmental problems due to land use change, depletion of freshwater resources, and pollution of aquatic and terrestrial ecosystems (Springmann et al., 2018). In addition, modern agriculture will have to adapt to global limits on mineral phosphorus supply (Blackwell et al., 2019) and increasing regulation of pesticide use (Tang and Maggi, 2021). This means land-bound agriculture will have to increasingly rely on the internal functional capacity of soils – e.g. to recycle nutrients and suppress diseases – and, thus, on soil health. Likewise, regulation of the climate through carbon sequestration and reducing greenhouse gas emissions (Lehmann et al., 2020) and the provision of habitats for aboveground biodiversity to bend the curve of biodiversity loss (Leclère et al., 2020) are directly and indirectly linked to soil health. Furthermore, soil biodiversity contributes importantly to climate change adaptation by facilitating water storage in soils through modifying soil organic matter (Lal, 2020) and achieving ONE Health through removal of contaminants and by preventing disease spread (Wall et al., 2015). Indeed, soil and soil health are at the heart of achieving many of the UN Sustainable Development Goals for 2030 (Keesstra et al., 2016; Lal et al., 2021) and the European Green Deal (Montanarella and Panagos, 2021).

Soil health, defined here as the continuing capacity of soils to deliver the multiple soil functions on which society depends, takes centre stage in policy and practice with respect to soils worldwide (Van der Putten et al., 2023; Veerman et al., 2020), and I use the term interchangeably with soil multifunctionality. However, currently the field is divided with respect to the methods we use to measure and assess soil functioning and indicators thereof. Only three indicator variables (soil organic matter (SOM), acidity, and available P) were commonly measured (> 70 % of schemes) across 65 schemes that aim to measure soil health or quality, and no biological measure is implemented in more than 30 % of the 65 schemes (Bünemann et al., 2018). Indeed, until very recently, there was no national- or European-level monitoring system that could address the key functions of soils comprehensively (Creamer et al., 2022; Van Leeuwen et al., 2017), although steps in this direction are now being taken (Norris et al., 2020; Orgiazzi et al., 2022; Zwetsloot et al., 2021), for instance in the EU's Soil Health Benchmarks project (https://soilhealthbenchmarks.eu, last access: 20 May 2025). It is clear that further harmonization in methods and quantification is urgently needed.

Partly, I think this plethora of methods and approaches stems from an oversimplified, often correlational understanding of the causal linkages driving soil multifunctionality, equipment availability in the laboratories involved, and a decades-old policy pressure to deliver easy-to-implement indicators fast (Creamer et al., 2022), which prevented the zooming-out needed to better understand the soil systematically (Harris et al., 2022). Indeed, what we need are “new analytical and conceptual approaches … that capture systems characteristics of soil health, in order to operationalize … monitoring soil health” (Lehmann et al., 2020). However, systemic perspectives that integrate soil functions and responses are in their infancy (Vogel et al., 2018). It is unclear how to manage the soil functions (Baveye et al., 2016) and how to link functions to soil processes (Vogel et al., 2018, but see Creamer et al., 2022). Integrating all soil processes is highly complex because soil properties are spatially heterogenous, and the interactions in soil are typically non-linear (Vogel et al., 2018). Soil biology is a key missing ingredient, and its complexity is paralysing the soil health literature (Creamer et al., 2022; Lehmann et al., 2020; Van Leeuwen et al., 2017). We know that soil biodiversity drives soil multifunctionality (Delgado-Baquerizo et al., 2016; Wagg et al., 2014), but the causal relation to soil functioning for many organisms is not clear (Creamer et al., 2022). Many soil microbial variables measured are hard to interpret and are insufficiently benchmarked to allow inferences about soil health (Fierer et al., 2021). Furthermore, most research focusses on soil health in an agricultural context (Debeljak et al., 2019; Fierer et al., 2021), but we also need to understand and quantify it in forestry, nature management, drinking-water production areas, and industrial and urban areas, which are strongly underrepresented (Norris et al., 2020; Orgiazzi et al., 2022).

To move forward, we first need to know what kind of information society needs from soil science. In this context, I think the main research tasks are as follows:

Determine which soils are best for which function (FAO and ITPS, 2015) and which functions can be combined (synergies) and which cannot (trade-offs).

To do these tasks well, we need to get organized as a scientific community. First and foremost, we need to set a common measurement system for the multiple functions of soil. We need a balanced set of indicators that reflect soil biology, chemistry, and physics but that are geared towards soil functioning (Lehmann et al., 2020). So far, selection of soil biological indicators was driven by well-known methods, feasibility in general laboratories, and costs, but this should be based on a sound understanding of how the indicators link to soil functioning mechanistically (Creamer et al., 2022; Lehmann et al., 2020; Vogel et al., 2018). New proposals typically try to go from soil processes to functions in one go, but soil is complex (Young and Crawford, 2004), and, so far, this approach has been defeated by this complexity. In many cases, the drivers of soil functions, either direct or indirect, are used implicitly or explicitly as proxies for the functions themselves. For example, soil nutrient content is used as a proxy for soil fertility (Daou and Shipley, 2019), and microbial biomass is used as a proxy for carbon storage (Wiesmeier et al., 2019); in both cases, these do contribute to the function but are not nearly a complete description of it. We can take steps forward by formally separating the causes and consequences and/or the predictors and the indicators of soil functioning and by linking them to the underlying processes and environmental and management context. I propose that we can do so by applying latent-variable models and structural causal modelling to soil multifunctionality research.

My aim with this paper is to propose a new methodology for measuring soil functioning and soil multifunctionality. It is based on the well-established technique of latent-variable modelling commonly used in psychometry, economics, and the social sciences at large. In parallel to my work presented here, Maaz et al. (2023) have also used latent-variable models to represent soil health; however, our approaches are quite distinct. They rely on a mixture of stocks, environmental conditions, and properties as indicators for soil health, while my aim is to link to the soil functions themselves. The next step after setting a valid measurement framework will be to develop a causal model of how trade-offs and synergies among soil functions are mechanistically regulated. If we define soil health as the continuing capacity of soils to deliver the multiple soil functions on which society depends then what are soil functions? Here, I define soil functions as soil processes, physical, chemical, or biological in nature, acting singly or in combination. These functions can be beneficial for human society but can also be involved in the internal functioning of ecosystems without direct human benefits, i.e. soil functioning for the sake of the ecosystem itself. For consistency, perhaps “soil functions on which society depends” should be called “soil services” as a specific form of ecosystem services.

Great mathematical frameworks now exist to combine multiple functions into one aggregate measure of multifunctionality (Byrnes et al., 2014, 2023), and they could be used to signal that “something is wrong” with soil functioning. However, understanding which soils perform all functions best in aggregate, e.g. the highest average soil function, is not informative enough to guide sustainable use of soils (Bradford et al., 2014; Lehmann et al., 2020). We need to know which soils perform which functions well and to what extend the functions can be combined or not in a single soil. So instead of focussing on univariate summary statistics of multifunctionality, we need to come up with a multivariate, but still simple and communicable, representation for soil multifunctionality (Lehmann et al., 2020; Zwetsloot et al., 2021). Multivariate models of multifunctionality have been developed, including network approaches that can be valuable in exploratory investigations (Siwicka et al., 2021). Others developed elegant multivariate models to estimate the influence of different drivers on functions and interrelations among functions (Dooley et al., 2015). However, all these approaches are correlational in nature, leaving the causal relationships that induced these correlations potentially unexamined (Shipley, 2016). I think this is problematic, because of (1) potential paradoxes in the data that no amount of big data can resolve (e.g. Simpson's paradox) and (2) difficulties in generalizing the results of analyses to other contexts. Posing a mechanistic model that links soil functions a priori, which is iteratively improved in the face of new data, can resolve both of these issues. In addition, hypothesizing about such mechanistic models will help in stabilizing the set of measured “functions” now rampant in the literature by excluding those indicators that are actually stocks or ecosystem properties and not processes (Garland et al., 2021; Lehmann et al., 2020). Confronting the hypothesized models with data and proposing improvements can be done with structural equation modelling (Box 1). However, how should the complexity of soils and soil functioning be organized into one model?

“Correlation is not causation” is a central piece of endemic wisdom we scientists throw at one another on a regular basis. However, its complement, “causation implies correlation”, is much less known due to Karl Pearson's (Pearson, 1911) crusade on causality. Nevertheless, it is the central concept in modern causal analysis (Pearl, 2009; Shipley, 2016). The modern causal revolution arose from the pioneering work of population geneticist Sewall Wright, who developed path analysis (Wright, 1921, 1934), a method to estimate causal effects from observational data. His method was ignored by statisticians and biologists for decades because it did not fit with the views of the dominant schools of statistics headed by Karl Pearson and Ronald A. Fisher (Shipley, 2016). Instead, the method was refined within economics, sociology, political science, and psychology (e.g. Jöreskog, 1967).

Path analysis was transformed into structural equation modelling (SEM), which uses maximum likelihood (ML) estimation to test causal multivariate hypotheses. The multivariate hypotheses are specified as a graph, specifically a directed acyclical graph, which captures the hypothesized causal relationships among the variables involved. The central idea is beautifully simple: if the specified causal hypothesis is true then we can predict which variables should be correlated and which should not be; the latter are considered to be conditionally independent. In fact, the method depends on predicting the covariance matrix of the variables, comparing it to the observed covariance matrix and testing the model fit (using an ML χ2 test). If the model does not fit the data (e.g. χ2p<0.05) then the hypothesized causal graph is rejected. If there is no lack of fit then one concludes that the data are consistent with the causal processes hypothesized (until, in the next paper, someone else proves you wrong, of course). For SEM to work, it needs to assume linear relationships and multivariate normal distributions of the variables involved, but it comes with the major advantage that it can estimate latent variables. Latent variables (LVs) are variables that were not measured or even cannot be measured. LVs are a way to measure the unmeasurable!

LVs are extremely important concepts as many things cannot be measured (Shipley, 2016). For instance, we cannot measure air temperature, which is the average kinetic energy of the molecules in the air; we can only measure its effects on, for example, the expansion of mercury in a capillary column (a mercury thermometer) or the change in electrical voltage in a thermocouple. These observed variables are of course causally linked to the latent quantity temperature, but they are observed with measurement error. Misspecifying this dependence relation in a causal model, thus conflating air temperature (“heat”) with the readings of your thermometer (translated to °C), can lead to an erroneous test of the causal model because it leads to a different expected covariance structure and, thus, different conditional independence claims. Latent-variable models (LVMs) are a way to get around this problem by specifying that the observed variable (thermocouple voltage) is caused by the latent quantity of interest (air temperature), but it is observed with error and therefore correlated but not identical. This situation is treated by “measurement models” (Fig. 1), a subsection of LVMs developed in the social sciences. To parameterize and test a single LV, four indicator variables need to be measured to have sufficient degrees of freedom, although this can be relaxed if the model entails multiple causally related LVs. LVs are also used to represent more hypothetical variables; e.g. concepts such as genes, atoms, and intelligence are examples of latent variables. These examples are successful latent concepts; there are also problematic ones, such as “ether”. Choosing, developing, and justifying latent variables are, perhaps, the most difficult aspects of structural equation modelling.

Recently, the SEM toolbox was expanded with a new estimation and testing method based on d-separation. D-separation is a criterion used to derive conditional independence claims, specifying which variables should not be correlated given the a priori specified causal model (Shipley, 2000, 2016). The d-separation-based approach is flexible and can fully accommodate non-normal data, non-linear functional relationships, and nested sampling structures as it works not with the whole covariance matrix but instead looks at each d-separation independence claim separately (using partial correlations in its most simple form) and combines this to test the whole causal model using a Fisher's exact C test (C for combined, the d-sep test; Shipley, 2000). The logic is the same as for ML-based SEM. Given an a priori causal model, one tests for the conditional independence of variables predicted by the model.

Note, the methods of SEM and LVMs are implemented mathematically as regression models, but it is important to realize that the interpretation of SEM is much stronger than for ordinary regression models. Ordinary regression models are simple tools aiming only to predict the effect of X on Y. The goal is prediction, not primarily understanding, although the latter is often attempted. Causal interpretation of regression models is problematic because parameter estimates and significance depend strongly on the included variables and even their order. In fact, misrepresenting the underlying causal structure can easily lead to entirely the wrong qualitative conclusions, e.g. in the situation called Simpson's paradox (see the supplementary code at https://github.com/JasperWubs/SoilMFv0.1, last access: 20 May 2025), which no amount of data will resolve. SEM, however, is different. It is different not because of its mathematics; rather, it is different because it relies on an a priori causal hypothesis to be tested with data. The a priori is crucial: when SEM software is used to find the “best”-fitting model by means of model selection tools (e.g. the Akaike information criterion (AIC)) then Wright's philosophy falls apart, and SEM becomes just another regression tool only to be used for explorative data analysis and hypothesis generation. So, as an analyst using SEM, you get one, and only one, epistemologically sound shot at testing your causal hypothesis. Of course, upon arduously collecting data and then rejecting your model, there is immense temptation to update the model by including new, not a priori specified, causal relationships and to present the updated model in the resultant paper as if it were the original a priori model. This is a posteriori discovery and, again, is only suitable for exploration and hypothesis generation and not for direct causal interpretation. Therefore, I am strongly in favour of implementing a strict requirement that SEM used for causal hypothesis testing is preceded by the publication of the a priori model in a curated, time-stamped repository. Any updates to the model should be fully reported in the paper because newly discovered links require further testing. In this way, our causal models can be transparently developed and updated. For both SEM and LVMs, excellent textbooks, reviews, and manuals exist (see Grace, 2006; Grace et al., 2010, 2012; Shipley, 2016); this is also the case for other tools in the causal analysis toolkit (Pearl, 2009). This summary is a condensed version of the key points in Shipley (2016).

Before we can model interacting soil functions mechanistically, we need a common framework to measure them. For this we have to move beyond using simple indicators since the processes driving the different functions of soils are complex. Soil fertility, for instance, is a complex soil function that drives the process of primary production. It is complex because many factors contribute to it (Daou and Shipley, 2019), and it changes through time. Higher nutrient availability but also water content and soil texture and structure interact to shape how well plants grow in a soil. Furthermore, plant species and cultivars respond differently to the different drivers of soil fertility; for example, some prefer nitrate over ammonium, others are salt- or drought-tolerant, some can puncture compacted soils, and other species cannot (Grime, 2001). So, while it is possible to build a soil fertility model for individual crops by accounting for their limiting factors for growth and estimating the functional relationships to these factors, this is much more difficult to quantify in general with predictive values for all plant and crop species in a community simultaneously (Daou and Shipley, 2019).

Nevertheless, we can borrow the data analytic machinery used in the social sciences to estimate these complex soil traits. In psychology, economy, and other social disciplines, complex properties are measured using latent-variable models – specifically a subsection called “measurement models” – that allow an analyst to infer the status of the complex property by modelling the responses that the property induces (Fig. 1). A well-known example is the IQ test that aims to quantify the complex and hard-to-measure trait intelligence (Spearman, 1904). It does this by fitting a measurement model to the measurable outcomes of intelligence, namely a person's ability to solve particular puzzles in a limited time. Daou and Shipley have successfully adapted this methodology for quantifying generalized soil fertility (Daou et al., 2021; Daou and Shipley, 2019, 2020), and I propose that we expand their framework to include all major functions of soil so we can study soil multifunctionality more systematically; I propose an IQ test for soils.

Following the functional land management (FLM) framework (Debeljak et al., 2019; Schulte et al., 2014; Zwetsloot et al., 2021), I focus on four main soil functions of direct importance to society (Fig. 2). The IQ test for soils will focus on the following soil functions: (1) primary production, driven by soil fertility; (2) climate regulation, consisting of carbon storage and reducing greenhouse gas (GHG) emissions (or net GHG consumption by soil); (3) water regulation, composed of water storage and purification of contaminants; and, finally, (4) provision of habitat for biodiversity, focussing initially on plant diversity. See proposals for expansion to other species groups in the Discussion section. I exclude nutrient cycling, which is included in the FLM, because I think it is not a soil function beneficial to society in and of itself. Instead, I see it as a structuring principle; nutrient cycling determines where nutrients are “invested” and thus which functions “thrive” (see also Schröder et al., 2016). Additionally, direct issues with nutrients for society, e.g. low soil fertility and nitrate leaching, are captured under the other soil functions – respectively, primary production and water purification in these examples.

Climate and water regulation are, respectively, further divided into the carbon storage and reducing greenhouse gas (GHG) emission subfunctions and water storage and purification subfunctions (light-blue circles) because of the very different causal mechanisms in play. The four soil functions are all interrelated – some trading off and others acting in synergy – because they all depend on the same basic resources (nutrients, energy, water). I hypothesize that the soil's plant–microbe–soil stoichiometry (green oval with orange operator sign) determines which functions are preferentially expressed by any given soil. How this regulation plays out is conditional on the geochemistry of the soil, mainly its mineralogy. Measuring the functions on a common scale and studying their interrelations using a common causal framework will help us determine how to manage soils for optimal multifunctionality.

The FLM framework was originally designed to integrate over relatively large spatial scales (Schreefel et al., 2022; Schulte et al., 2014) and uses decision trees, partly based on expert judgement, to generate assessments of the different soil functions on a semi-quantitative scale (low–medium–high; Soil Navigator Decision Support System, DSS; Debeljak et al., 2019). In addition, the assessment of different functions is partly based on the same information (Zwetsloot et al., 2021); e.g. SOM is a component in four out of five functions. How those pre-specified modelling relations affect the observed trade-offs and synergies among functions is unclear. While I think the efforts made using FLM (and the associated Soil Navigator Decision Support System, https://soilnavigator.eu/, last access: 20 May 2025) have great value for society in recommending changes based on the best knowledge today, I also believe we need to deepen our mechanistic understanding of the interrelations of the soil functioning and how they can be optimized. For this, I propose that we need a measurement and modelling framework that (1) allows quantitative assessment of soil functions based on independent data and (2) assesses functions and drivers at low spatial and temporal resolutions (Bradford et al., 2016, 2017; Fierer et al., 2021).

Many processes in soil depend on factors external to soil, such as temperature and water inputs. This contributes to the challenge in using many biological soil health indicators (Fierer et al., 2021) as they can become highly variable in time and space. To get around that, it was proposed that soils be incubated under standard conditions so that only factors internal to a soil would contribute to the observed functioning (Daou and Shipley, 2019). This is the approach I take here as well, and, as such, the proposed measurement system is focussed on estimating potential soil functioning and multifunctionality under a set of soil-external conditions optimal for plant growth. Below, I provide suggestions on how to link these measures to actual in situ rates of soil functioning. Nevertheless, I think this focus on the intrinsic – although not time-invariant – potential soil functioning is important as it can give the method predictive value for expected in situ soil functioning irrespective of the weather conditions that materialize during the growing season.

Here, I outline a proposal for a standardized soil multifunctionality assay that addresses the key soil functions in the functional land management framework (Schulte et al., 2014). The method is based on incubations of intact soil cores, subjected to several treatments, and measuring responses that are indicative of the underlying soil functions (Fig. 3; Table 1). The methods assume that all soils are sampled in the same way and incubated under standardized conditions, including temperature, light, watering regime, and air humidity, to ensure comparability (see Table 2 for a proposal). The goal is to estimate the intrinsic capacity of each soil for performing each soil function.

A soil sampling team will collect 32 soil monoliths (60 mm × 25 cm deep, ∼ 22.6 L soil) per soil. The monoliths are used to quantify primary production (a – eight green monoliths, two per bio-assay plant species), climate regulation (b – four orange monoliths, one for each substrate addition treatment), and water regulation (c – four blue monoliths for water storage and purification measurements) and for supporting plant biodiversity (d – 16 coloured monoliths, with each colour representing an indicator plant; this is the same as in (a), for which direct and indirect plant–soil feedback (PSF) is estimated in phase 2 (P2) in each of the four soils conditioned during phase 1 (P1)). The monoliths are incubated for 90 d under standard incubation conditions (Table 2). As such, the measurements target the capacity of a soil to deliver key soil functions under optimal conditions for plant growth. For both primary production and biodiversity functions, plant harvest days are fixed (indicated in days after the species name) and based on plant dry mass. Likewise, upon substrate addition (t0), gaseous efflux of CO₂, N₂O, and CH₄ are measured on fixed days, with intensive sampling in the first 14 d and then less frequent sampling until day 90. In addition, microbial C and C in soil fractions (aggregates) are measured after 70 and 90 d. The water regulation measurements can be done independently in this setup and can potentially be shifted in time but are now placed at the end of the 90 d period to spread the workload over time. However, infiltration and leaching measurements will be conducted over a fixed time period.

Generally, the practical and logistical choices for method selection in soil quality assessments vary depending on the objectives: mechanistic understanding, functional land management, and/or large-spatial-scale monitoring (Creamer et al., 2022). However, the scheme I propose here aims to strike a balance between these three objectives. The use of measurement models linked to functions important to land management, standardized measurements that can be compared across labs and thus potentially scaled up, and a flexible framework that allows the integrated study of underlying mechanisms make this three-way integration possible. The question is this: how well will it do all three?

Currently, I propose that samples be collected as intact soil cores to preserve soil structure and macroscopic features of soil so that the real vertical and horizontal variations are reflected in the measurements. These replicate cores need to come from small homogeneous areas, accounting for variations in microclimate, soil type, and land use and management. However, it was shown that intact cores and homogenized soils generate almost identical pictures of soil fertility (Daou and Shipley, 2020), which would make for much easier sample collection and handling. Similarly, earlier studies using substrate additions sometimes incubate as little as 80 g of soil (Doetterl et al., 2015); this would strongly minimize substrate and soil requirements and may be an improvement over what I propose here. Likewise, Daou and Shipley (2019) conduct their work in a highly controlled growth cabinet, but could the data still be measured with acceptable error variances in a glasshouse, a screenhouse, or a common garden setup? In a common garden, of course, temperature and light cannot be controlled, but maybe their impact can effectively be approximated by using growing degree days as measured by a local weather station or temperature loggers?

Loosening up this constraint will be important for application of the method in the Global South where high-tech facilities are strongly limiting. In general, the approach may be challenging to implement “as is” in the Global South and potentially in other labs as well. Currently, the method relies on some advanced lab analytical equipment to get all the required measurements. Further work needs to focus on gaining meaningful measurements using simpler approaches, but they need to be validated against the robust methods identified here.

What about sampling time? Do we need to include seasonal dynamics, e.g. reflecting the massive turnover of bacterial and fungal communities over the year (Schadt et al., 2003), or can we select a single most predictive period? I think it would be most valuable if we could sample in the seasonally cold and/or dry period when plant growth is most limited. Then we could compare in situ soil functioning data in the field during the subsequent growing season to our prior off-season ex situ estimates. These linkages could be used to build predictive models. An alternative would be to sample during the peak season, but then, often, (1) farmers are busy on their field, (2) crops are damaged by sampling and walking, and (3) researchers are occupied with other field experiments and observations.

Here, I propose incubating soils under standard soil-external conditions optimal for plant growth (see Table 2). However, can these conditions be applied to all soils? What about soils that experience regular waterlogging? What about soils from low- or high-temperature conditions: will the shift to mesic conditions cause unnatural behaviour of these soils? Can we shorten the protocol? For biodiversity regulation, I propose conducting two-phase plant–soil feedback experiments (Bever, 1994; Van der Putten et al., 2013), but from the first phase alone, we can also use the shoot biomass data to get an initial idea of the soil's ability to support plant diversity by looking at the evenness of the relative abundances (Pielou, 1966). Could that be predictive of phase-2 competitive hierarchies?

I am strongly in favour of reporting on the measured soil functions separately so that fellow scientists, policymakers, and the public can make their own assessment and overlay their own priorities with respect to the multiple functions of soil. However, can these measure not be combined into a single indicator? If they are combined with reports of the individual functions, I think they can be. There is a huge body of literature on multi-objective optimization methods (Pereira et al., 2022) where combining objectives is operationalized using explicit rules and criteria. Such optimization should be done with maximum transparency about how functions are weighted and combined for the aggregate index to have any practical use. Also, the weighing should be informed by involving multiple stakeholder group consultations, e.g. using focus group discussions (Bampa et al., 2019; Schulte et al., 2019).

The methods I propose are too cumbersome to be used directly in commercial soil testing but are crucial to advance our foundational understanding. In order to be useful, indicators need to be conceptually relevant; sensitive to changes; informative for management; and effective, e.g. cheap and fast (Lehmann et al., 2020). I argue that my method is conceptually relevant and sensitive and that, when the measurements are explicitly linked to environmental and management data, the results can be used to inform management decisions. The effectiveness is something requiring further testing; see the preceding discussion for steps I want to take. Additionally, we should explore how these soil functioning measurements can be approximated by high-throughput screening techniques such as near-infrared spectroscopy, X-ray fluorescence, and potentially eco-acoustics and environmental DNA.

Finally, to scale up and inform spatial planning and management choices worldwide, the measurements need to be integrated into a strong framework, explaining the potential, the synergies, and the trade-offs among functions mechanistically (Fierer et al., 2021). Including biology in these models is key (Creamer et al., 2022; Fierer et al., 2021). As recently as 2004, a map of known soil threats and degradations published by Science listed only physical and chemical forms of soil degradation and was solely focussed on agricultural production (AAAS, 2004). We have moved on but into unknown territory. The mechanistic machinery is, for an important part, there in the literature but needs to be conceptually brought together, e.g. by using plant–microbe–soil stoichiometry as an organizing principle.

In the wake of the Green Revolution, seeing widespread application of chemical fertilizers and pesticide control, the importance of soil science has dwindled. Now, due to the threats exerted on human societies by climate change and biodiversity loss, soil has been revalued as a central nexus integrating many aspects of human wellbeing (Sigl et al., 2023). I believe that the study of soil multifunctionality and, thus, soil health should lie at the heart of this new valuation of soil and soil biodiversity and should be a key focus area in order to bring humanity within the planetary boundaries (Steffen et al., 2015) while simultaneously developing sustainable livelihoods for all (Dearing et al., 2014; Fanning et al., 2022). This also means that we have to put the study of soil multifunctionality on solid empirical and theoretical footing, for which this paper develops a concrete proposal (Sect. 4; Fig. 3).

A key improvement is that I separated the causes and consequences of the soil functions. Focussing on the consequences allows standardized measurements that can be adopted across laboratories, both foundational and applied research oriented, and allows them to be linked flexibly, via the estimated latent variables, to competing mechanistic frameworks through structural equation models. Linking the ex situ functional measurements by mechanistic causal models is also important to understand the results within their environmental context. It is well known that soil health indicators need to be interpreted in site-specific ways (Creamer et al., 2022; Vogel et al., 2018), and that means that a global understanding needs to account for the relevant site specificities. For instance, clay content determines what range of values to expect for organic matter content (Lehmann et al., 2020), while soil texture shapes ecosystem recovery trajectories (Bach et al., 2010). A key question will be “how unique are the properties and functions in this soil?” compared to the soils in our reference set. To what extent can we extrapolate our results meaningfully and based on which (minimum) set of parameters? To answer these questions we need to bring soil functional and contextual measurements together in a common global database.

Here, I have worked out a simple but causally consistent methodology to quantify soil multifunctionality and, thus, soil health. The system is based on latent-variable modelling (LVM), with each LV capturing one crucial soil function: primary production, climate regulation (split into carbon storage and GHG emission reduction), water regulation (split into water storage and purification capacity), and biodiversity regulation (captured as plant diversity potential). This system makes explicit the fact that soil functions are complex soil properties, contingent on many drivers, that cannot be measured directly using any device. It also explicitly separates the causes and consequences of each soil function. Using the consequences as indicators, we can estimate the LV factors that approximate the soil intrinsic capacity to perform each function. For example, we can estimate soil fertility from plant growth. I hope this can be a common point of departure in the soil health field to allow scientists to band together and to organize soil multifunctionality and soil health research more mechanistically.