Estimating the attainable soil organic carbon deficit in the soil fine fraction to inform feasible storage targets and de-risk carbon farming decisions

Dataset

The data used in this study were derived primarily from a national dataset collected under the Soil Carbon Research Program (SCaRP), representing 4180 farmer paddock sites across Australia’s cropping and pasture regions (Baldock et al. 2013b). The SCaRP dataset (https://doi.org/10.25919/5ddfd6888d4e5) included SOC fractionation data for 312 soils and was augmented with fractionation data derived for an additional 163 SCaRP soils, fractionated after the completion of the original SCaRP project (n = 475). The dataset included gravimetric concentrations of total organic carbon and the following three component fractions: (1) particulate organic carbon (POC) representing the SOC associated with >50 μm particles after removal of resistant organic carbon (ROC); (2) humus organic carbon (HOC) representing the SOC associated with ≤50 μm particles after removal of resistant organic carbon (ROC); and (3) ROC representing the SOC associated with >50 μm and ≤50 μm particles having a polyaromatic chemical structure. The POC fraction has a turnover time of 1–2 years, while the HOC fraction is expected to remain in soils for up to 100 years. The ROC fraction assumes a stable form of carbon, which has a turnover time of centuries to millennia.

The SOC contained within the HOC fraction was used to provide measured values for the current carbon concentration of FF_{SOC_Actual} in each fractionated soil in units of mg FF_SOC g⁻¹ ≤2 mm soil. In addition to the data available within the published SCaRP dataset, the mass of soil fractionated and the mass of the fine fraction for each fractioned soil was obtained from archived SCaRP data files and were used to calculate the gravimetric concentration of the FF_Soil in units of g FF_Soil g⁻¹ ≤2 mm soil. A detailed description of the distribution of the sampling sites, sampling design, and the procedure used to derive the total SOC, POC, HOC and ROC concentrations is described by Baldock et al. (2013a).

Quantile regression analysis and calculation of the attainable soil organic carbon deficit in the fine fraction of the soils

Initially, the FF_{SOC_Attainable} analysis was performed as a global 90th quantile regression analysis in which soils (n = 475) collected from sites with annual precipitation values ranging from 277 mm to 1809 mm and collected from different depths (0–0.05 m; only for a few sites in New South Wales, sites; 0–0.10 m; 0.10–0.20 m; and 0.20–0.30 m) were pooled and a single model was derived. This represents the typical approach taken when calculating SOC deficit. However, after exploring the initial results, it was evident that an approach that acknowledged variations in mean annual precipitation and soil depth would be more appropriate.

The significance of separating data based on precipitation and depth prior to fitting 90th quantile regressions (Eqn 1), was tested by converting both precipitation and depth into dummy variables. The dummy variables were created based on soil depth (0–0.10 m or 0.10–0.30 m) and mean annual precipitation (≤600 mm or >600 mm) from 1991 to 2010 at the site from which the soils were collected. The 90th quantile regression model was fitted including a higher order interaction term (precipitation class × depth class), which was significant (P < 0.0001), indicating an inhomogeneity of slopes and/or of the intercepts. The inference is carried out using the ‘bootstrapped’ approach and the quantile regression modelling is performed using R package quantreg (Koenker et al. 2023). The 90th quantile regression analysis (Eqn 1) of the gravimetric content of FF_{SOC_Actual} (mg FF_SOC g⁻¹ ≤ 2 mm soil) as a function of the gravimetric content of FF_Soil (g FF_Soil g⁻¹ ≤ 2 mm soil) was applied to all soils for each depth and precipitation combination (Table 1), and the slope (β) and intercept terms were defined. The 90th quantile predicted values for FF_{SOC_Actual} were used to define the FF_{SOC_Attainable} for each soil, and the FF_SOC__{Attainable_Def} of each soil was calculated as the difference between the FF_{SOC_Attainable} and the measured FF_{SOC_Actual} (Eqn 2).

where β and intercept refer to quantile regression analysis parameters associated.

Chemometric model development for the attainable soil organic carbon deficit

During SCaRP, infrared (IR) spectra were acquired and archived for all 475 soils and the additional 20 020 soils that SCaRP analysed (see Baldock et al. (2013c) for details about the acquisition of the IR spectra). The IR spectral datasets covered both the MIR and part of the NIR spectrum sections of the electromagnetic region. The derived FF_{SOC_Attainable_Def} values and the concomitant IR spectra acquired for the 475 soils were used to assess the ability to generate predictive IR/PLSR algorithms to predict the FF_SOC__{Attainable_Def} values of each soil.

The raw spectra acquired by SCaRP were pre-processed as follows: (1) reflection was converted into absorbance (absorbance = log(1/reflectance); (2) the spectra were truncated and resampled between 6000 and 600 cm⁻¹ with a resolution of 4 cm⁻¹; and (3) a baseline offset transformation was applied (i.e. baseline corrected). In the baseline offset transformation, the lowest absorbance value of the spectrum was subtracted from all other spectral values. Baldock et al. (2013c) built successful PLSR prediction models with MIR spectra to which only baseline correction was applied and noted that including additional pre-processing methods did not improve the resultant PLSR models. Soil samples were finely grounded prior to scanning using Fourier transform MIR spectroscopy, and minimum pre-processing of the spectra is performed in the current study.

Due to the limited number of calibration soils, instead of splitting the dataset into calibration and validation datasets, the chemometric models were developed as bootstrapped IR/PLSR models. The optimum number of factors was determined using the ‘onesigma’ approach as explained in the ‘pls’ R package through a leave-one-out-cross validation approach (Liland et al. 2022). A total of 100 bootstrapped models were developed and were validated using ‘in-the-bag’ and ‘out-of-bag’ datasets. The out-of-bag validation approach can be considered as an independent validation since the soils included were not included in the model calibration process. The models were evaluated by the root mean square error (RMSE), as a measure of the model accuracy, and Lin’s concordance correlation coefficient (LCCC) (Lawrence 1989), as a measure of how good the fit between measured and predicted values was. Model prediction quality was considered superior when RMSE and LCCC of the developed models were close to zero and one, respectively.

The developed IR/PLSR models were applied to the remaining 20 020 SCaRP MIR spectra to predict their FF_SOC__{Attainable_Def} values. The IR/PLSR prediction uncertainty was assessed by calculating the lower (i.e. 0.05) and upper (i.e. 0.95) percentiles derived from the 100 bootstrapped model outputs for each sample.

Spatial modelling of soil carbon saturation deficit

Spatialisation of the point FF_SOC__{Attainable_Def} concentrations was performed using the IR/PLSR predicted values for all SCaRP soils. Before spatial modelling, the FF_SOC__{Attainable_Def} concentrations were converted into FF_SOC__{Attainable_Def} stocks with Eqn 3.

where FF_SOC__{Attainable_Def} stock is expressed in Mg C ha⁻¹, FF_SOC__{Attainable_Def} is expressed as mg C g⁻¹ soil, bulk density is expressed as g soil cm⁻³, depth/thickness is expressed in cm, gravel correction factor (1 - gravel fraction), P_rt correction for the proportion of the land area within the sampling unit allocated to rocks and/or trees. Measured values for bulk density, gravel correction factor and P_rt were provided for each soil as part of the SCaRP dataset.

Spatialisation was then performed by an ensemble approach using a Random Forest (RF) model (Breiman 2001). Two global spatial models (i.e. 0–0.10 m and 0.10–0.30 m) were fitted with estimates generated from four different IR/PLSR model estimates associated with depth and mean annual precipitation.

Quantifying fine fraction soil organic carbon storage parameters

The global analysis performed to derive the FF_{SOC_Attainable} is in Fig. 3. It was revealed that not considering key environmental drivers of carbon inflows and outflows into soils; e.g. mean annual precipitation and other factors such as the sampling depth, generated unrealistic estimates of FF_SOC__{Attainable_Def}. As seen in Fig. 3, the 90th quantile regression line (Blue colour) that defined the FF_{SOC_ Attainable} was governed by the FF_{SOC_Actual} of the soils collected from the 0–0.10 m soil layer and from sites with higher mean annual precipitation. As a result, the estimated FF_SOC__{Attainable_Def} was overestimated for regions with low mean annual precipitation and lower depth intervals even with the same FF_Soil mass. The overestimation of the FF_SOC__{Attainable_Def} is due to fitting a global 90th quantile regression where the upper limit of the FF_SOC__{Attainable_Def} is defined using higher values of the FF_{SOC_Actual} reported in topsoils and higher precipitation regions. The likelihood of achieving the global estimated FF_{SOC_Attainable} might not be physically possible under current environmental constraints that affect the carbon inflows and outflows.

Fig. 3.

Relationship between soil organic carbon concentrations in the fine fraction of soils (FF_Soil) with the fine fraction mass. The colour gradient in dots represents the variation in the mean annual precipitation associated with each sampling site and the symbols represent the different soil depths from which soils were collected. The blue colour line indicates 90th quantile regression line.

Given the limitations associated with the global analysis (Fig. 3), and the significant (P < 0.0001) higher order interaction terms, showing an inhomogeneity of slopes and/or of the intercepts, were obtained when the precipitation class and depth class were included in the 90th quantile regression analysis, performing 90th quantile regression separately for each combination precipitation and depth was justified (Table 1). As a result, FF_{SOC_Attainable} and subsequent calculations of the FF_SOC__{Attainable_Def} were performed separately for soils derived from different depths and from sites with different mean annual precipitations. The results of the quantile regression analyses completed for each combination of soil depth (0–0.10 m and 0.10–0.30 m) and mean annual precipitation (≤600 mm and >600 mm) are in Fig. 4. The FF_{SOC_Attainable}, as defined by the 90th quantile regression line, in the 0–0.10 m soil layer was approximately twice that associated with the 0.10–0.30 m soil layer. The FF_{SOC_Attainable} associated with soils collected from sites with >600 mm mean annual precipitation was approximately twice that associated with soils collected from sites with ≤600 mm mean annual precipitation.

Fig. 4.

The summary of the regression analysis that is used to derive the FF_{Attainable_SOC}. (a) Depth 0–0.10 m and LRZ. (b) Depth 0.10–0.30 m and LRZ. (c) Depth 0.0–0.10 m and HRZ. (d) Depth 0.10–0.30 m and HRZ.

The distribution of FF_SOC__{Attainable_Def} concentrations and some summary statistics are in Fig. 5 and Table 2. The density plot for the calculated FF_SOC__{Attainable_Def} associated with each combination of depth and mean annual precipitation is depicted in Fig. 5. Based on the median FF_SOC__{Attainable_Def} values, soils from the HRZ had higher concentrations of FF_SOC__{Attainable_Def} than soils from the LRZ for both depth intervals (Table 2). Further, a higher standard deviation of FF_SOC__{Attainable_Def} was found for soils from the 0–0.10 m depth than the 0.10–0.30 m depth within the HRZ. In contrast to the HRZ, the magnitude of the difference between FF_SOC__{Attainable_Def} median values between the two soil depth layers was higher for the LRZ. In fact, the 0–0.10 m depth layer had FF_SOC__{Attainable_Def} values three times larger than the 0.10–0.30 m depth layer. The quantile regression models estimated parameters are summarised in the Supplementary Table S2.

Fig. 5.

Density plots for the calculated FF_SOC__{Attainable_Def} concentrations by depth/precipitation classes.

Chemometric modelling of attainable soil organic carbon deficit concentrations in the fine fraction of soil

The four separate IR/PLSR calibration models (in-the-bag) for 0–0.10 m and 0.10–0.30 m in the LRZ and HRZ for both soil depth layers, respectively, had mean LCCC values ≥0.85 (Table 3). The out-of-bag validation of the fitted models, except for IR/PLSR model fitted for depth interval 0.10–0.30 m and LRZ, reported an LCCC value greater than 0.75. Model performance appeared to correspond with the number of samples included suggesting that increasing the number of soils in each class through the fractionation of additional soils covering wider spatial variation could increase model performance. In-the-bag and out-of-bag mean RMSE values varied over the combinations of soil depth and mean annual precipitation (Table 3). This was due to variations in the concentrations of the FF_{SOC_Actual} across those groupings. Nonetheless, these results suggested that the IR/PLSR algorithms derived could be used to confidently predict FF_SOC__{Attainable_Def}.

The mean β coefficients associated with 100 bootstrapped for each of the four IR/PLSR calibration models show which spectral components contributed to the prediction of the FF_SOC__{Attainable_Def} (Fig. 6). Examination of these β coefficients showed distinct positive contributions from spectral features at 3640–3604 cm⁻¹, 2296–1872 cm⁻¹, 1284 cm⁻¹, 1180 cm⁻¹, 980 cm⁻¹ and 812 cm⁻¹, corresponding with mineral-associated peaks of gibbsite, quartz, clay and O-Si-O. Conversely, strong negative effects were noted from spectral features at 3700 cm⁻¹, 3208–2806 cm⁻¹, 1624–1552 cm⁻¹ associated with OH−, organic matter, and amide C=O bonds. The negative contribution from organic matter to the predictions was also observed in the β coefficients of FF_SOC__{Attainable_Def} models in Baldock et al. (2019) and as they discuss, is consistent with FF_SOC__{Attainable_Def} declining as organic carbon content increased.

Fig. 6.

The mean β coefficients derived from the 100 bootstrapped partial least squares regression (PLSR) equations for (a) 0–0.10 m in low rainfall zone (LRZ; ≤600 mm); (b) 0–0.10 m in high rainfall zone (HRZ; >600 mm); (c) 0.10–0.30 m in the LRZ; and (d) 0.10–0.30 m in the HRZ.

Spatial modelling of attainable soil organic carbon deficit stocks in the fine fraction of soil

The in-the-bag and out-of-bag spatial machine learning model validation statistics are in Table 4. Both fitted models reported similar model quality assessment statistics with a model accuracy of ~5 Mg C ha⁻¹ as defined by RMSE out-of-bag values. The top 10 environmental covariates for both depth intervals were identified and were predominately associated with climate (Fig. 7). This demonstrated the importance of climatic variables that affect biomass production and carbon flow into the soils. The SOC stabilisation was spatially represented by digitally mapped clay mineral types and the summation of the clay and silt fractions across the study area. The SOC stabilisation properties were ranked 12th and 14th in the VIP plots generated for the fitted spatial random forest models, respectively, for 0–0.10 m and 0.10–0.30 m depth intervals. Having climate-driven variables, clay-silt content (fine fraction) and clay mineral (i.e. illite, kaolinite and smectite) as environmental covariates for the prediction of FF_SOC__{Attainable_Def} demonstrated their importance on the inflow and stabilisation of the SOC within the FF_Soil.

Fig. 7.

The estimated mean importance value for the considered environmental covariates used to estimate the FF_SOC__{Attainable_Def} through fitted 100 bootstrapped spatial Random Forest models. Table S1 describes the covariates. (a) Variable importance of covariates, 0–0.10 m. (b) Variable importance of covariates, 0.10–0.30 m.

Derived spatial estimates are in Fig. 8, and prediction uncertainty was calculated as 0.05 and 0.95 percentiles (see Supplementary Fig. S1). When the FF_SOC__{Attainable_Def} values were closer to zero or negative, those regions have less potential to enhance the SOC in the FF_Soil. The spatial pattern of the FF_SOC__{Attainable_Def} demonstrates that some of the land areas under both LRZ and HRZ have FF_SOC__{Attainable_Def} close to zero or negative. For example, in the state of Victoria, the HRZ Gippsland region generally reported less area under the positive stocks of FF_SOC__{Attainable_Def} for the depth interval 0–0.10 m compared to 0.10–0.30 m depth interval estimates.

Fig. 8.

Distribution of the soil organic carbon attainable deficit stocks across major agricultural production regions of Australia. The spatial estimates were made for specified two depth intervals: (a) 0–0.10 m; and (b) 0.10–0.30 m.

Similarly, in LRZ, the Mallee region reported less area under the positive stocks of FF_SOC__{Attainable_Def} indicating less opportunity to increase the stable form of SOC in the FF_Soil. The eastern boundary of Australia, mainly southern Queensland and New South Wales, reported higher positive stock values for the FF_SOC__{Attainable_Def} indicating an opportunity to further enhance the SOC in the FF_Soil. The predicted FF_SOC__{Attainable_Def} stocks across Australia revealed an opportunity to increase current FF_SOC by 3.47 GT and 3.24 GT for the 0–0.10 m and 0.10–0.30 m depth intervals, respectively. The average estimated positive FF_SOC__{Attainable_Def} stock values were 13.35 Mg C ha⁻¹ and 12.49 Mg C ha⁻¹, respectively, for the 0–0.10 m and 0.10–0.30 m depth intervals. There was also discontinuity of spatial estimated FF_SOC__{Attainable_Def} stock values in the north-south direction in the east of Australia caused by changes in the rainfall gradients of the input covariates.

Developing a framework for rapid and cost-effective estimation of FF_SOC__{Attainable_Def} concentrations using IR/PLSR modelling framework

Chemometric models are frequently used to predict various chemical, physical and biological soil properties (Janik et al. 1998; Viscarra Rossel et al. 2006; Stenberg et al. 2010). A range of studies have now demonstrated the ability to predict SOC and its fraction contents from IR spectra using chemometric modelling approaches (Baldock et al. 2013c, 2018). For the SCaRP soil samples used in this study, Baldock et al. (2013c) demonstrated that SOC, SOC fractions, and total nitrogen contents could be predicted using IR spectral data and chemometric modelling. Baldock et al. (2018) extended the use of IR/PLSR estimated SOC contents and SOC fractions to characterise the vulnerability of SOC to subsequent loss in agricultural soils from New Zealand.

Only a few studies (see Baldock et al. 2019) have used a combined IR dataset and chemometric modelling approach, similar to that developed in this study, to derive a unique prediction algorithm for quantifying the FF_SOC__{Attainable_Def} of the FF_Soil. Baldock et al. (2019) demonstrated an ability to derive the values of key soil properties identified as drivers of FF_{SOC_Attainable} by Beare et al. (2014) as well as an ability to directly predict measured FF_{SOC_Attainable} and FF_SOC__{Attainable_Def} from the acquired IR spectra. However, neither Baldock et al. (2019) nor the supporting work of Beare et al. (2014) or McNally et al. (2017) examined the potential impacts that rainfall and soil depth or similar drivers of SOC accumulation would have on the derivation of both FF_{SOC_Attainable} and FF_SOC__{Attainable_Def}. Further, these previous studies (Beare et al. 2014; McNally et al. 2017; Baldock et al. 2019) did not define the deficit as FF_SOC__{Attainable_Def}. Instead, the application of a 90th quantile regression to the FF_{SOC_Actual} across a range of soils was assumed to provide a measure of the potential upper limit that FF_{SOC_Actual} could reach. This upper limit of FF_{SOC_Actual} was considered an appropriate value on which to base the calculation of saturation deficit irrespective of soil properties, environmental conditions, or applied management practices that can all influence the amount of carbon captured by plants and added to the soil.

Deriving the soil organic carbon attainable deficit concentrations using broad drivers

A range of environmental factors and soil properties affect the flow of carbon into the soil, how that carbon can be stabilised within the soil matrix, and the fraction that is lost from the soil back to the atmosphere. In Australian dryland agricultural systems, potential dry matter production of crops and pastures is dominantly dictated by rainfall. While an attempt was made to develop different quantile regression models considering soil depth (i.e. 0–0.10 m vs 0.10–0.30 m) and using a broad classification of environmental drivers that affect carbon flows (i.e. LRZ vs HRZ), we acknowledge that, if available, a higher number of measurements could and enable FF_SOC__{Attainable_Def} to be derived based on local properties that govern the inputs and their subsequent biological processing as well the stability of SOC. Although the capacity of soils to store SOC in a stable form is commonly attributed to the FF_Soil (clay + silt), there are other properties such as specific surface area, extractable aluminium (pyrophosphate) content of soils, and soil pH that affect the SOC stabilisation capacity (Beare et al. 2014; McNally et al. 2017). In contrast, the emphasis in this study was focused on quantifying the loading of organic carbon within the FF_Soil (i.e. the FF_{SOC_Actual} concentration) within soils collected from different depths from sites having different mean annual precipitations.

The framework adopted allowed us to create scalable products on FF_SOC__{Attainable_Def}, from point estimates to spatial estimates using readily available datasets. In contrast to the current study, McNally et al. (2017) developed an empirical model to predict the SOC stabilisation based on the broader soil types (allophanic vs non-allophanic soils) covering topsoils (0–0.15 m) of New Zealand permanent pasture and cropping soils. Our analysis revealed that estimated quantile regression coefficients reported different SOC loadings of the FF_SOC mass (y-axis in Fig. 4) and also the rate of stabilisation (slope of the regression respective models, supplied in the supplement Table S2) with change in FF_Soil mass after considering the broad rainfall classification and depth intervals. This result demonstrated that performing quantile regression analysis to estimate the FF_SOC__{Attainable_Def} is interlinked to the key drivers that govern the flow of carbon into soil systems. In fact, we found unrealistic values of the FF_SOC__{Attainable_Def} if the values were derived from the global analysis (Fig. 3), due to input datasets consisting of different precipitation gradients and sampling depth intervals. In this study, due to limited measurement dataset, only broad rainfall classification and depth intervals were considered in defining FF_SOC__{Attainable_Def}. We concluded that local analysis to derive FF_SOC__{Attainable_Def} is a more appropriate way forward compared to the use of a single global analysis in future studies.

Are attainable deficit estimates static or dynamic?

The key assumption made in the current analysis is that the upper 90th quantile regression, used to define the FF_{SOC_Attainable}, should be applied separately to groups of soils differentiated based on the magnitude of FF_{SOC_Actual} drivers. The magnitude of carbon input to soil represents an additional driver that was not included in the analyses completed due to a lack of site-specific data. Although the results were strongly influenced by mean annual precipitation, plant productivity and thus carbon inputs to the soil within dryland agricultural regions of Australia are significantly impacted by management practice. Hochman et al. (2016) reported that the average wheat yield in Australia was 1.7 Mg ha⁻¹ (1996–2010), and the average simulated water-limited yield potential was 3.5 Mg ha⁻¹. Their analysis revealed an average wheat yield gap of 1.8 Mg ha⁻¹ representing 51% of the potential yield. If the yield gap can be reduced without altering the harvest index or root/shoot ratios and stubbles are retained the flow of carbon to the soil can be increased. Under such conditions, calculated FF_SOC__{Attainable_Def} values for a given soil depth and mean annual precipitation might change due to increased carbon flow to the soil. Thus, values obtained for FF_{SOC_Attainable} and FF_SOC__{Attainable_Def} should be considered dynamic and continually recalculated as improvements in productivity are realised through the use of improved genetic material and management practices.

The above discussion then leads to the question of what is the appropriate manner to refer to the SOC deficit when values are derived from a range of soils considered to be representative of current agricultural management practices? Do the values derived provide an indication of the potential upper limits of FF_{SOC_Actual} that may be possible for the soils or are they only representative of what can be obtained using previous and current management practices and types of crop/pasture types grown? As a result, we argue that the approach used in this study and previous studies (e.g. Beare et al. 2014; McNally et al. 2017; Six et al. 2002) defines an FF_{SOC_Actual} stock that is attainable (i.e. the FF_{SOC_Attainable}) given the current soil properties, environmental conditions and applied management practices and not a potential upper limit of FF_{SOC_Attainable_Def}. The potential upper limit of FF_{SOC_Attainable_Def} should be considered as conceptual, unknown and undefinable. Thus, when attempting to quantify the FF_{SOC_Attainable_Def} based on FF_{SOC_Actual} values obtained across a range of soils, it should be defined as an attainable deficit rather than a potential deficit.

Identification of spatial drivers of FF_SOC__{Attainable_Def}

Previous work on Australian continental scale spatial modelling of SOC includes mapping of SOC stocks (Viscarra Rossel et al. 2014), and SOC fractions (Viscarra Rossel et al. 2019; Román Dobarco et al. 2022); however, no literature was found on spatial modelling of the FF_{SOC_Attainable_Def}. Therefore, a comparison of the spatial drivers of FF_{SOC_Attainable_Def} found in this study was completed against previously published Australian literature on total SOC and HOC. Similar to the current analysis of FF_SOC__{Attainable_Def} spatial modelling (0–0.10 m and 0.10–0.30 m), Viscarra Rossel et al. (2014) reported that for the Australian cool temperate and temperate-Mediterranean regions rainfall was a key driver of the SOC stocks (0–0.30 m). Viscarra Rossel et al. (2014) used annual cumulative rainfall as a driver of SOC stocks, while in the current analysis, season rainfall was used in addition to cumulative annual rainfall. Further, Viscarra Rossel et al. (2019) identified climate drivers, namely, mean annual temperature, mean annual precipitation, and potential evapotranspiration, as the highest ranking important variables for estimating SOC using a global variable importance analysis for the continental scale mapping of the HOC. Similar climate drivers governed the spatial distribution of FF_SOC__{Attainable_Def} in the current study for both depth intervals (Fig. 7). Román Dobarco et al. (2022), used isometric log-ratio transformation (ilr) for the SOC fractions, namely POC, HOC and ROC, instead of mapping those fractions using individual models. Therefore, direct comparison with the current study is not feasible. However, their two ilr models revealed that radiometric and climate variables were key drivers for the respective models fitted for the considered depth intervals, namely 0-0.05 m, 0.05–0.15 m, and 0.15–0.30 m. Only one radiometric variable was included in the current study when considering the top 15 variables based on VIP ranking. Overall, it can be concluded that key drivers of the FF_SOC__{Attainable_Def} followed a similar pattern to drivers of the SOC and its fraction stocks for the Australian context.

How the current proposed framework differs from the existing approaches used to quantify soil organic carbon storage

As discussed earlier, our approach was based on defining the FF_SOC__{Attainable_Def} associated with the FF_Soil. Further, we deployed an empirical approach through the fusion of: (1) measurements of FF_{SOC_Actual} and FF_Soil masses; (2) statistical modelling (i.e. quantile and IR/PLSR regression modelling); and (3) spatial upscaling through an ensemble machine learning approach. From the carbon accounting perspective, the total SOC is used to define the carbon storage capacities. This can cause volatility due to residence time of different SOC fractions, particularly labile forms of SOC. On the other hand, the current approach is conservative and focuses on the more stable forms of SOC resulting in less vulnerability and potential volatility. For example, Padarian et al. (2021) used an empirical approach to quantify the SOC storage capacity using total SOC stocks through the development of a machine learning quantile regression model. In contrast in the current analysis, the FF_SOC__{Attainable_Def} was derived using the loading of carbon in the FF_Soil.

Implications

Current analysis revealed the successful use of the IR/PLSR models beyond the prediction of concentration of SOC and its component fractions and their vulnerability to loss. In fact, IR/PLSR models developed herein for FF_SOC__{Attainable_Def} can be used along with other IR tools to evaluate the current status of the SOC, its composition, vulnerability and attainable space, adding extra value for the landholder and carbon aggregators. These IR-derived datasets will provide a better assessment of the attainable SOC deficit and aid in the derivation of realistic carbon sequestration targets within future carbon projects.

Further, derived spatial estimates and their prediction uncertainties can be used to evaluate the potential of land parcels to increase SOC sequestration in the FF_Soil. We have derived the spatial estimates at a resolution of 90 m, which will provide a meaningful insight at the regional scale, but applications are limited to the paddock scale. These derived estimates will provide insight into the attainable mineral-associated SOC that can be further increased. Although these estimates are derived based on the previous and current land management practices, the estimates will provide firsthand information for the aggregators to explore land parcels that have a higher potential to increase SOC, higher ability to retain carbon/permanence through mineral association and minimise the risk of reversal. With the spatialised outputs generated at a 90 m spatial resolution, the outputs could be used as guidance information rather than explicit information. However, the framework developed in the current study could be used to generate project-specific information through a combination of IR/PLSR and spatial modelling using the project specific datasets. For policymakers, these spatial estimates will enable the identification of regions where targeted policies can be implemented when designing SOC projects. For project and policy scale applications, the generated spatial outputs should be used while considering their uncertainties. The graphical summary of the pathways to market and policy directions is depicted in Fig. 9.

Fig. 9.

Pathways to market and linkage to carbon policies for the derived FF_SOC__{Attainable_Def} datasets.