Introduction

Soil organic carbon (SOC) constitutes the largest continuously cycling carbon pool in terrestrial ecosystem1,2. SOC is chemically or physically associated with soil minerals, such as iron‒aluminium (Fe-Al) oxides, which are crucial for soil carbon accumulation and long-term storage3. Generally, SOC can be divided into mineral-associated organic carbon (MAOC) and particulate organic carbon (POC), which have distinct formation processes, functions, and turnover times4. POC is a mixture of partially decomposed plants and their decomposition byproducts and has a short mean residence time5,6. In contrast, MAOC, which is protected by chemical bonds associated with minerals and is classified as a stable form of C, has a longer turnover time and enhanced chemical stability7,8. Fe-bound organic carbon (OC-Fe), an essential component of MAOC, plays a crucial role in the accumulation and preservation of SOC through the formation of Fe‒organic complexes6,9. Consequently, an increase in the amount of OC-Fe can contribute to a reduction in soil C emissions under climate change10. However, the global patterns in the OC-Fe budget, spatial distribution and regulatory mechanisms are still unclear, which limits our capacity to tailor soil management practices in accordance with climate change mitigation pathways.

The formation of OC-Fe primarily originates from the interaction between minerals and SOC10. Fe-Al oxides control SOC storage, turnover, and mineralization11. Fe contributes to SOC accumulation through three primary mechanisms: the promotion of soil aggregate formation, adsorption and coprecipitation with dissolved organic carbon, and alterations in microbial activity9,12,13. Similarly, Al can adsorb to reactive surface sites or coprecipitate with soil solid phases and thus enhance the stability of SOC14. Accordingly, Fe-Al oxides can accelerate the accumulation of OC-Fe and play a crucial role in SOC sequestration15,16,17,18,19. Specifically, the contribution of Fe-Al minerals to SOC ranges from 3% to 72%20 and mainly depends on differences in the degree of crystallinity of these compounds21 because poorly crystalline minerals have a larger specific surface area than crystalline minerals and higher chelation capacity22. Notably, poorly crystalline Fe (Feo, Fep, Alo and Alp) adsorbs six times more SOC than does crystalline Fe17, with poorly crystalline Fe-Al oxides globally sequestering a quarter of the SOC pool20. Although accurate estimation of SOC is an essential prerequisite for mitigating climate change, the insufficient inclusion of SOC-determining factors in predictive models leads to a degree of uncertainty in SOC forecasts19. Hence, numerous scholars have advocated for the incorporation of metal factors into SOC prediction models, as they have reported that mineral factors often have greater importance than climatic variables do14,23,24. However, only global data on the soil clay content25 and mineral composition26,27 are currently available, and information on the global distribution of Fe-Al oxides is lacking. Thus, analyses of the impacts of Fe-Al oxides on SOC within global terrestrial ecosystems may be limited. Therefore, estimating the global distributions of various forms of Fe-Al oxides is essential for improving the accuracy of SOC calculations.

OC-Fe associations are regulated by biogeochemical factors. The first is the activity and morphology of Fe- and Al-oxyhydroxides, which particularly affect the accumulation of OC-Fe12,23,28,29. Second, clay, silt and sand particles play important roles in the stabilization and formation of OC-Fe30 because they have the ability to strongly bind with Fe-Al oxides and dominate the surface area available for the sorption of SOC. Third, climate features, which drive changes in precipitation, temperature, and ecosystem type, significantly affect OC-Fe formation. For example, cold and arid climates coupled with good soil aeration and high redox potential promote OC-Fe formation in high-altitude regions of the Tibetan Plateau31. Moreover, soil pH regulates OC-Fe by influencing the solubility of metal cations32. Previous studies have shown the importance of Fe oxides in OC-Fe protection and mineralization, Fe oxides associate with SOC, through adsorption and co-precipitation, reduced the bioavailability and promoted the accumulation of OC-Fe33. However, the global mechanisms by which Fe-Al oxides affect OC-Fe have rarely been explored on the basis of multiple comprehensive factors. A critical gap remains in examining the roles of these dominant factors and demonstrating the general mechanism of OC-Fe formation.

Although previous studies have reported the contribution of OC-Fe to SOC and have revealed distinct variations across ecosystem types34, OC-Fe has rarely been studied at a global scale. For example, Zhao et al.28 reported that forest OC-Fe constituted 37.8 ± 20.0% of the SOC in 14 forest soils in the USA. Furthermore, Fang et al.19 reported that OC-Fe constituted 15.8 ± 12.0% of the SOC in plateau grasslands on the Tibetan Plateau. Wan et al.35 reported an SOC range of 6.2–31.2% for 12 cropland soils in east-central China. Although OC-Fe is important to SOC, few studies have been conducted at the regional level, and the global contribution of OC-Fe to SOC or MAOC in terrestrial ecosystems has not yet been estimated. This has led to a lack of understanding of the general stabilization mechanisms of SOC. Moreover, the maximum Fe-bound organic carbon capacity (OC-Femax), which affects the soil stabilized C sequestration capacity, remains unknown.

Thus, in this study, we synthesized 3,395 observations from 178 publications to (1) reveal the global distribution patterns of Fe-Al oxides and use these mineral factors to further assess the global stocks of SOC, MAOC, and OC-Fe; (2) investigate the factors regulating variations in soil OC-Fe concentrations; and (3) evaluate the contributions of OC-Fe to SOC and MAOC stocks at a global scale. The results of this study provide empirical insights that can help guide land management practices to increase carbon sequestration and mitigate global climate change.

Results

Variations in the soil OC-Fe concentration

To determine the importance of OC-Fe with respect to SOC, we first aimed to determine the spatial variations in these two variables across different ecosystems on the basis of an analysis of global data. SOC concentrations demonstrated significant variations across different ecosystems, with a decrease observed with increasing soil depth. The mean SOC concentrations were highest in wetlands (117.9 ± 117.1 g kg−1; mean ± s.d.), followed by permafrost (117.6 ± 52.3 g kg−1), shrubland (22.8 ± 32.7 g kg−1), forest (22.8 ± 24.1 g kg−1), grassland (22.7 ± 25.0 g kg−1), and other types (21.6 ± 27.2 g kg−1), with the lowest value found in croplands (17.2 ± 19.7 g kg−1) (Supplementary Fig. S1A). In the topsoil, the SOC concentration was 54.1 ± 74.8 g kg−1, whereas that in the subsoil was 11.5 ± 21.5 g kg−1 (Supplementary Fig. S1A). The highest SOCconcentration in the topsoil was found in permafrost at 123.0 ± 48.9 g kg−1, followed by that in wetlands at 122.9 ± 118.2 g kg−1, shrubland at 35.7 ± 36.8 g kg−1, forest at 33.4 ± 26.6 g kg−1, other types at 32.6 ± 32.0 g kg−1, grassland at 29.0 ± 9.8 g kg−1 and cropland at 19.5 ± 10.8 g kg−1 (Supplementary Fig. S1A). In contrast, the highest SOC concentration in the subsoil was found in wetlands at 53.7 ± 77.8 g kg−1, followed by permafrost at 32.9 ± 21.7 g kg−1, cropland at 10.8 ± 17.1 g kg−1, forest at 9.7 ± 10.7 g kg−1, other types at 8.9 ± 10.8 g kg−1, grassland at 9.8 ± 13.4 g kg−1, and finally shrubland at a mere 3.1 ± 4.6 g kg−1 (Supplementary Fig. S1A).

The global average OC-Fe concentration in the topsoil was 17.5 ± 23.8 g kg−1, which was accompanied by variation across different ecosystems. The OC-Fe distribution in the topsoil among ecosystems was highest in wetlands (23.9 ± 27.9 g kg−1), followed by permafrost (12.7 ± 3.3 g kg−1), forest (5.8 ± 9.4 g kg−1), other (5.5 ± 5.7 g kg−1), grassland (4.1 ± 5.0 g kg−1), and cropland (3.5 ± 3.5 g kg−1) (Fig. 1A). OC-Fe was significantly positively correlated with both SOC (R2 = 0.74, p < 0.001) and MAOC (R2 = 0.34, p < 0.001) (Fig. 1B, C). The distribution pattern of the OC-Fe concentration was similar to that of SOC and MAOC.

Fig. 1: Comparison of OC-Fe across terrestrial ecosystems in topsoil.
Fig. 1: Comparison of OC-Fe across terrestrial ecosystems in topsoil.
Full size image

A OC-Fe; B Correlations of SOC with OC-Fe; C Correlations of MAOC with OC-Fe. SOC, soil organic carbon; OC-Fe, Fe-bound organic carbon. MAOC, mineral-associated organic carbon. The numbers in parentheses indicate the numbers of samples. Box plots indicate the medians (horizontal lines), 1st and 3rd quartiles (boxes), 1.5× interquartile range (whiskers), means (diamonds), and outliers (dots outside whiskers). Shading depicts 95% confidence intervals on the slope.

Relationships between soil OC-Fe concentrations and regulatory factors

We selected correlation factors as model variables to estimate the global OC-Fe distribution and evaluate its significance for SOC. In this study, the climatic factors MAP (mean annual precipitation) and AI (aridity index) were significantly positively correlated with Fed, Fep, Ald, and Alp (p < 0.05) (Supplementary Fig. S2). The MAT (mean annual temperature), MAP, CS (clay and silt), and elevation were identified as the most important predictors of the Fe-Al oxide concentration (Supplementary Fig. S3). Additionally, the concentration of Fe-Al oxides was found to be greater in forested regions than in other ecosystems (Supplementary Table S1). Moreover, diverse forms of Fe-Al oxides (Supplementary Figs. S3–5) were predicted to exist, and the results revealed that each form had distinct effects on SOC (Supplementary Fig. S6). Poorly crystalline Fe-Al oxides (Alo + 0.5Feo and Alp + 0.5Fep) had a notably greater effect on SOC (R2 = 0.10, p < 0.001; R2 = 0.10, p < 0.001) (Supplementary Fig. S7B, C). An assessment of the relative importance of the MAOC variables revealed that Feo had greater impacts (Supplementary Fig. S7D–F; Supplementary Fig. S6C, D). Feo (R2 = 0.31, p < 0.001) had the most significant influence on OC-Fe (Fig. 2A; Supplementary Fig. S8). The effect of Alo + 0.5Feo (R2 = 0.55, p < 0.001) on OC-Fe was greater than that of Ald + 0.5Fed (R2 = 0.42, p < 0.001) and Alp + 0.5Fep (R2 = 0.42, p < 0.001) (Fig. 2E, F; Supplementary Fig. S7G). Soil physicochemical properties, such as the soil bulk density (SBD), total nitrogen (TN), clay content, and silt content, were also strongly associated with soil carbon (Supplementary Fig. S9). OC-Fe, fsOC-Fe (OC-Fe contributes to SOC), and fmOC-Fe (OC-Fe contributes to MAOC) exhibited significant negative correlations with PET (potential evapotranspiration) and positive correlations with AI (p < 0.05). Additionally, OC-Fe and fmOC-Fe were significantly positively correlated with the MAT (p < 0.05), whereas OC-Fe and fsOC-Fe were significantly negatively correlated with the MAT (p < 0.05), in contrast to the positive correlation observed for fmOC-Fe (p < 0.05) (Supplementary Fig. S10).

Fig. 2: Prediction performance and correlation plots of soil OC-Fe and controlling factors.
Fig. 2: Prediction performance and correlation plots of soil OC-Fe and controlling factors.
Full size image

A Relative importance for OC-Fe; B Predictions for OC-Fe (best data approach, random forest model), showing a good fit across the complete data range; C Correlations of OC-Fe with Feo: Fed; D Molar ratios of OC-Fe: Fed; Correlations of OC-Fe with 0.5Feo + Alo (E) and 0.5Fep + Alp (F). MAP mean annual precipitation (mm), MAT mean annual temperature (°C), PET potential evapotranspiration (mm), AI aridity index, TN soil total nitrogen (g kg−1), CS clay and silt (%), SBD soil bulk density (g cm−3), 0.5Feo + Alo weight-normalized contents of Feo and Alo, 0.5Fep + Alp weight-normalized contents of Fep and Alp. Fed and Ald dithionite-extractable Fe-Al, Feo and Alo oxalate-extractable Fe-Al, Fep and Alp sodium-pyrophosphate-extractable Fe-Al. Box plots indicate the medians (horizontal lines), 1st and 3rd quartiles (boxes), 1.5× interquartile range (whiskers), means (diamonds), and outliers (dots outside whiskers). Shading depicts 95% confidence intervals on the slope.

Both MAOC (R2 = 0.49, p < 0.001) and OC-Fe (R2 = 0.49, p < 0.001) demonstrated significant positive correlations with the degree of Fe oxide (Feo: Fed) activation (Fig. 2C; Supplementary Fig. S7H). The mean molar ratios of OC-Fe: Fed differed across ecosystems; the mean value for wetlands was highest at 14.1, followed by that for croplands at 5.5, that for forests at 3.7, that for permafrost at 3.6, and that for grasslands at 2.7 (Fig. 2D).

Global patterns of OC-Fe

The random forest machine learning algorithm yielded a good prediction of OC-Fe (R2 = 0.90) through fivefold cross-validation (Figs. 2B and  3A; Supplementary Table S2), and in this study, we present a global map of organic carbon and its fractions within the 0–30 cm soil depth (Fig. 3; Supplementary Figs. S11–14). The predictions of this study revealed a global total OC-Fe stock of 232.9 PgC in the topsoil (0–30 cm) (Table 1). Furthermore, the area-weighted average soil OC-Fe concentration was 10.2 ± 8.0 g kg−1, with a soil OC-Fe density of 2.5 ± 1.6 kg m−2 in the topsoil. A global map of the soil OC-Fe content revealed a distinct latitudinal pattern (Fig. 3A). The OC-Fe concentrations are roughly symmetrically distributed between the hemispheres, decreasing and then increasing from the equator towards the poles. Highlands and mountain areas at lower latitudes, such as the Tibetan Plateau, the Alps, the Greater Caucasus Mountains, and the highlands of central Africa, etc., have higher OC-Fe concentrations. Additionally, some equatorial plains with high biomass, such as the Amazonian plains, also have relatively high soil OC-Fe concentrations. The mean OC-Fe density was similar to the distribution patterns of the OC-Fe concentrations (Fig. 3A; Supplementary Fig. S14A).

Fig. 3: Modelled global patterns of soil OC-Fe density at soil depths ranging from 0 to 30 cm.
Fig. 3: Modelled global patterns of soil OC-Fe density at soil depths ranging from 0 to 30 cm.
Full size image

A OC-Fe density; B fsOC-Fe; C OC-Femax density; D OC-Fe saturation. Values were predicted and calculated using a data-driven random forest model with global climate, edaphic properties, and geographical factors. fsOC-Fe, OC-Fe contributes to SOC. OC-Femax, the maximum Fe-bound organic carbon capacity. The percent OC-Fe saturation was calculated as the current Fe-bound organic carbon (OC-Fe) divided by the maximum Fe-bound organic carbon capacity (OC-Femax). The blank areas depicted in the maps represent regions from which tundra and deserts have been excluded from the analysis.

Table 1 Modelled global stocks of OC-Fe and OC-Femax as well as global mean values of fsOC-Fe and fmOC-Fe within the 0–30 cm soil layer across different ecosystems
Full size table

The contribution of OC-Fe to SOC decreased from the equator and began to increase again at approximately 40° latitude in both hemispheres (Fig. 3B). The estimated OC-Femax decreased with increasing latitude in the Southern Hemisphere, peaked near the equator, and then initially decreased in the Northern Hemisphere before rising again, with a prominent peak at approximately 25°N (Fig. 3C). There was significant sequestration potential near the equator and at 25°N. Specifically, the greatest OC-Fe deficits (18.9 ± 9.5 kg m−2) were observed at approximately 25°N latitude (Fig. 3D; Supplementary Fig. S14B). Regional variations in OC-Fe saturation were also apparent, with higher saturation in the western part of North and South America in the highlands of southern Africa, whereas saturation was lower in central South America, central Africa and Southeast Asia (Fig. 3D).

The distribution of the OC-Fe stock varied across different ecosystems: forests had the largest reserve at 118.2 PgC, followed by grasslands at 68.3 PgC, shrublands at 22.3 PgC, and croplands at 20.2 PgC. Furthermore, the OC-Fe stock varied with depth: 31.2 PgC for the top 5 cm, 76.3 PgC for the next 10 cm, and finally 125.4 PgC for the deepest layer at 15–30 cm (Table 1). The fsOC-Fe in forests was the highest at 40.8 ± 15.6%, followed by that in wetlands at 21.3 ± 20.8%, that in grasslands at 29.9 ± 15.5%, that in croplands at 27.4 ± 9.0%, and that in shrublands at 26.4 ± 10.8% (Table 1). The fmOC-Fe was highest in forests (65.2 ± 27.6%), followed by wetlands (51.7 ± 32.7%), grasslands (45.7 ± 24.8%), croplands (42.2 ± 18.1%) and shrublands (39.0 ± 17.1%) (Table 1). The mean OC-Fe saturation was highest in wetlands (48.5 ± 57.9%), followed by shrublands (42.7 ± 37.8%), grasslands (37.1 ± 35.1%), forests (27.8 ± 35.6%), and croplands (25.5 ± 25.4%). The highest OC-Fe deficit implied that terrestrial ecosystems have abundant potential to store more OC-Fe, as was observed in forests at 516.9 PgC, followed by grasslands at 247.5 PgC, croplands at 143.0 PgC, shrublands at 80.6 PgC and wetlands at 7.0 PgC (Table 1). The global average saturation of OC-Fe formed through adsorption and co-precipitation was 20.9 ± 22.6% and 22.6 ± 20.9%, respectively (Supplementary Fig. S15).

Discussion

Fe-Al oxides are the primary minerals bound to SOC36. In this study, we assessed the distinct spatial patterns of various forms of Fe-Al oxides, and a model based on fivefold cross-validation was performed effectively (Supplementary Table S3). Similarities were observed in the latitudinal profiles of the mean concentrations of Feo with those of SOC, MAOC, and OC-Fe (Supplementary Figs. S4,  S11, S13, and 14). Among all forms of Fe oxides, Feo is the primary factor influencing SOC sequestration10. More specifically, poorly crystallized mineral phases of Fe oxides regulate the formation of OC-Fe, which subsequently impacts SOC sequestration on a global scale.

While there were similarities in the latitudinal profiles of Fe-Al oxides and SOC, differences were observed in the latitudinal profiles of various forms of the oxide minerals. The global distribution map of Fe-Al oxides (Supplementary Figs. S3 and 4) revealed distinct latitudinal variations among the different forms of these oxides. As the latitude decreased in both hemispheres, the mean concentrations of Fed and Ald clearly increased. In contrast, Feo and Alo gradually decreased from high latitudes, followed by an increase, peaking at lower latitudes. Additionally, the latitudinal gradient of Fep aligns with that of Feo but decreased less at higher latitudes, whereas Alp remained relatively unchanged along the latitudinal gradient. Most forms of Fe-Al oxides exhibit relatively high mean concentrations at relatively low latitudes. This is primarily attributed to wet tropical conditions, which promote more extensive weathering of primary minerals, leading to greater retention of Fe-Al oxides in the soil profile36,37,38,39. Soils in tropical regions are mainly composed of kaolinite and sesquioxides (e.g., gibbsite, haematite, and goethite) and are thus rich in Fe-Al oxides40. Additionally, at low latitudes, precipitation has a strong effect on Fep and Alp, facilitating Fe-Al complexation reactions23.

The stocks of Fe-Al oxides varied across different ecosystem types and increased with soil depth (Supplementary Table S1). The stocks of various forms of Fe-Al oxides displayed notable variability, with forests having the largest stocks, followed by grasslands, shrublands, and croplands in descending order (Supplementary Table S1), which was likely due to abundant rainfall and high net primary productivity, particularly in tropical forests. These conditions provide optimal environments for Fe-reducing and Fe-oxidizing bacteria41, leading to a greater presence of short-range-ordered oxides in forests20,23,42. Second, the low Fe-Al oxide stocks in croplands may result primarily from tillage operations, which bring deeper soils lacking chemical protection from Fe-Al oxides to the surface, thereby reducing the levels of Fe-Al oxides (e.g., Feo, Fep, and Alp)43.

Our results suggest that the primary determinants of soil OC-Fe are soil properties, mineralogical factors and climate features (Fig. 2A, B). Soil properties play a pivotal role in determining the concentration and spatial distribution patterns of OC-Fe, with pH, elevation, TN and CS emerging as the main influencing factors (Fig. 2A, B). A lower soil pH enhances the adsorption of SOC onto Fe oxides32. This occurred because a decrease in pH increases the concentration of Fe and Al cations in the soil solution, which subsequently enhances the formation of OC-Fe14,20,44. A global distribution map of OC-Fe highlights that colder and drier climate regions have relatively high OC-Fe levels (Fig. 3A), which can be attributed to well-aerated soil with high redox potential, which is conducive to Fe-Al oxidation and consequently promotes OC-Fe formation. These results are consistent with those of a study conducted in the permafrost regions of the Qinghai‒Tibet Plateau, where high altitudes promote the accumulation of OC-Fe19,31. Importantly, this study did not estimate OC-Fe reserves in tundra and permafrost soils because of the limited availability of data. These findings underscore the need for further research in tundra and permafrost ecosystems to inform future management strategies aimed at enhancing carbon storage. The effect of TN on OC-Fe may be indirect. Atmospheric nitrogen deposition and biological nitrogen fixation increase plant growth, resulting in increased plant-derived organic carbon. This organic carbon is then selectively protected by Fe oxides35, leading to elevated OC-Fe levels. Clay, silt, and sand particles are key contributors to OC-Fe stabilization and formation30. Their capacity to form strong bonds with Fe-Al oxides and their dominance in providing surface area for SOC sorption. The maximum amount of carbon that clay and silt fractions can retain is known as the soil carbon saturation. This concept was also employed in this study to estimate OC-Fe saturation7,45,46.

Feo was the primary mineral factor influencing the soil OC-Fe content (Fig. 2A). The main reason for this is that poorly crystalline Fe oxides have a greater capacity to promote SOC binding because of their larger specific surface areas and greater number of reactive sites12. Therefore, the crystallographic phase of minerals significantly influences the formation of OC-Fe. The formation of OC-Fe primarily results from the adsorption and coprecipitation of organic carbon with Fe oxides47. The molar ratio of OC-Fe to total Fe oxides (OC-Fe: Fed) indicates the mechanism by which organic carbon binds to Fe oxides48,49, with ratios less than 1 and greater than 6 indicating adsorption and coprecipitation, respectively44. Analysis of the average OC-Fe: Fed molar ratio across different ecosystems suggested that wetlands are predominantly coprecipitating, whereas croplands, forests and permafrost and grasslands are a combination of adsorption and coprecipitation (Fig. 2D). Climate variables (MAT and MAP) can influence the molar ratio by altering the content of Fed50. OC-Fe was significantly positively correlated with Feo: Fed (R2 = 0.49, p < 0.001) (Fig. 2C), indicating that highly reactive Fe oxides play a dominant role in the interactions between Fe and OC51. The importance of mineralogical and climatic factors in determining OC-Fe concentrations depends on the climate region. In arid regions, Fe oxides, particularly Feo, are more important predictors of OC-Fe than are climate factors. This result is the same as that of previous studies at lower elevations in arid regions15. Therefore, including mineralogical factors in models for predicting OC-Fe is important for improving predictions.

Warm, arid conditions, along with high soil pH and specifically high calcium content, promote the binding of Fe oxides to SOC via cation bridging15. Changes in rainfall or warming temperatures in dry zones may have little impact on organic carbon sorption to Fe reactive mineral surfaces20. However, in the tropics, climate strongly influences OC-Fe concentrations (Supplementary Fig. S10). The MAT is a key factor and is negatively correlated with OC-Fe concentrations in the tropics (Supplementary Fig. S10). Owing to warming, the organic carbon associated with Fe oxides has become vulnerable to degradation52, suggesting that further research on the OC-Fe response to global climate change is needed to improve the understanding of soil carbon stabilization mechanisms under future warming scenarios. Precipitation and soil moisture could impact the stabilization of SOC by altering the dissolution of Fe and Al from minerals, regulating their ability to bind with and protect organic carbon via complexation or adsorption14. Notably, wetlands are unique in that they experience seasonal variations in moisture53. Alterations of moisture greatly influence the formation and stability of OC-Fe due to the changes of soil oxygen status33,54, active Fe oxide content and enzymatic activity55, soil pH56. Given that data on this topic are rare, highlighting that future research should focus on the OC-Fe temporal dynamics in the changes of moisture condition to better comprehend the role of wetlands in the global carbon cycle.

This study revealed differences in the ways in which OC-Fe contributes to SOC (fsOC-Fe) across diverse ecosystems. In topsoil, the global stock of OC-Fe is 232.9 PgC and constitutes 33.5 ± 15.6% of the SOC, which is higher than the 25% reported in a previous global study conducted by Kramer and Chadwick20 on carbon retention by reactive minerals. The discrepancy may arise from the difference in soil sampling depths: previous research extended to 2 m, whereas the current study focused solely on the topsoil layer (0–30 cm). The results of this study revealed that the fsOC-Fe of forests was 40.8 ± 15.6%, which is close to the reported value of 37.8 ± 20.0% in a total of 14 forest soils sampled across the U.S.A28. Fang et al.19 reported that the proportion of OC-Fe to SOC in grassland soils on the Tibetan Plateau was 15.8 ± 12.0%, which was lower than the global mean value (29.9 ± 15.5%) for grasslands calculated in this study (Table 1). The global map generated in this study also indicated lower fsOC-Fe values in the Tibetan Plateau region (Fig. 3D). Wan et al.35 investigated the range of fsOC-Fe in 12 cultivated soils in east-central China, which ranged from 6.2% to 31.2%. They extrapolated the fsOC-Fe value for Chinese agricultural soils to 15.7 ± 6.4%, which is lower than the global cropland average reported in this study (27.4 ± 9.0%) (Table 1). The discrepancy arises because the previous study was conducted in East Asia, where the global fsOC-Fe map shows lower values (Fig. 3D). This regional difference contributes to the lower estimates reported in that study than those reported in the current research. This study highlights the general contribution of OC-Fe to SOC at the global scale and highlights its importance.

In addition to variations among different ecosystems, the fsOC-Fe distribution also varied with soil depth. The fsOC-Fe value increased gradually with increasing depth (Table 1). Notably, in our study, we analysed only the surface layer, and a more detailed vertical gradient analysis is warranted. The fsOC-Fe was significantly positively correlated with clay and silt (Supplementary Fig. S16), indicating that physical aggregation and clay minerals may jointly protect SOC34. SBD exhibited a significant positive correlation with the fsOC-Fe value (p < 0.05) (Supplementary Fig. S16), implying that tightly textured soil enhances OC-Fe stability and further affects the fsOC-Fe. The fsOC-Fe was significantly negatively correlated with the MAT and PET (p < 0.05) (Supplementary Fig. S10). These climatic factors indirectly influence soil carbon dynamics primarily by affecting plant biomass and mineral weathering processes57.

OC-Fe is the main component of MAOC6, but few researchers have analysed the contribution of OC-Fe to MAOC (fmOC-Fe). The fmOC-Fe value was highest in forests (65.2 ± 27.6%), followed by wetlands (51.7 ± 32.7%), grasslands (45.7 ± 24.8%), croplands (42.2 ± 18.1%), and shrublands (39.0 ± 17.1%) (Table 1). Although the fmOC-Fe value varied across different ecosystems, the trend closely aligned with changes in Fe oxide stocks (Supplementary Table S1; Supplementary Fig. S16). For example, forests presented elevated levels of Fe oxides41, and they also presented the highest fmOC-Fe values. These results suggested that Fe oxide stocks influence the fmOC-Fe (Supplementary Fig. S16). The fmOC-Fe value was considerably greater at depths of 0–5 cm and 15–30 cm but lower at depths of 5–15 cm in croplands, grasslands and shrublands (Table 1). Detailed investigations are required to elucidate the specific reasons behind this depth-based variation.

We compared the OC-Fe saturations of two pathways in Fe-C association and found that co-precipitation is more conducive to OC-Fe formation than adsorption at a global scale (Supplementary Fig. S15). Although there is a lack of direct correlational analyses to elucidate the controlling mechanisms for the OC-Fe deficit. We speculate that there are several main factors involved. First, OC-Fe sequestration is controlled through Fe minerals and depends on the specific surface area, Fe mineral crystallinity, and Fe oxidation state10,58. Second, soil physicochemical properties (e.g., clay and silt contents and pH) regulate OC-Fe through physicochemical interactions34. Third, climate can indirectly affect the formation or stability of OC-Fe associations by altering the weathering of Fe-bearing minerals, the formation and transformation of iron minerals, and microbial activity10,59. Moreover, plant root exudate C inputs (e.g., oxalic acid, a component of organic acids) could also promote carbon loss by releasing organic compounds from protective associations with Fe and stimulating a ‘priming effect’ to increase microbial mineralization of SOC60,61.

The implementation of soil carbon management practices requires pertinent guidelines that consider climate features and ecosystem properties. Thus, management of soil carbon sequestration in terrestrial ecosystems should target relatively low saturation levels with abundant potential areas. Here, OC-Fe saturation was greater at the low latitudes (15°N) of the Northern Hemisphere and midlatitudes of the Southern Hemisphere (30– a large OC-Fe deficit in the midlatitudes (25°N) and the equatorial region (Fig. 3C, D; Supplementary Fig. S14B), with abundant potential to maximize carbon accrual. Forests have the lowest OC-Fe saturation, and forests have greater plasticity in storing more OC-Fe; thus, forests should take action to store more OC-Fe. Despite the lower OC-Fe saturation of forests, large-scale afforestation is still a valuable measure for enhancing the soil carbon store because of its higher concentrations of OC-Fe than those of croplands and grasslands (Fig. 1). For cropland, increased organic manure fertilization could increase OC-Fe by promoting the formation of poorly crystalline Fe fractions17 and thus increase SOC accumulation62. The OC-Fe saturation of wetlands is the highest, indicating that strengthening the protection and restoration of wetlands is necessary63. In addition, maintaining a reasonable water table level55 and Sphagnum restoration and cultivation may be effective solutions for increasing SOC sequestration and persistence in wetland29. This study can help us understand which ecosystem may be most useful for developing SOC sequestration strategies or practices. However, the estimates of OC-Femax in this study were obtained from an idealized model, and further in-depth investigations are necessary to assess specific measures and associated benefits.

Despite our considerable efforts to compile databases and generate global predictions, our study has several limitations. First, sparse data have been obtained with similar spatial distributions of soil properties in certain regions, such as northern Canada, central and western Australia, the Russian region, and western Europe, potentially resulting in decreased prediction accurac64. Second, the number of observations for subsoil (>30 cm depth) was significantly smaller than that for topsoil (0–30 cm depth), leading to increased uncertainty in subsoil prediction, particularly with respect to unknown reservoirs of OC-Fe. Third, although we employed commonly used methods such as machine learning and incorporated more than a dozen predictor variables, a portion of our results remain unexplained. This discrepancy could stem from measurement errors or methodological limitations. Additionally, the inclusion of global raster data to fill in missing values further introduces uncertainty. These limitations highlight the necessity for additional measurements of Fe-Al oxides and OC-Fe, particularly from underrepresented regions, as well as the application of more sophisticated statistical techniques for spatial prediction. In addition, more attention should be given to OC-Fe in the subsoil, as it has been shown that the fsOC-Fe value varies substantially across soil layers34. The OC-Fe extraction method has certain uncertainties, particularly regarding its extraction efficiency, which remains unclear65. Future research should focus on developing more accurate and efficient extraction techniques while adapting methodologies for diverse environmental conditions to ensure more precise global quantification of the OC-Fe pool. Some propagation of error occurs from the estimated Fe-Al oxides to the predicted OC-Fe, which will be addressed in future research using more advanced methods. The deficit calculation will be significantly different when the values from coprecipitation versus sorption are used. However, for field samples, determining whether sorption or coprecipitation is the major dominant process is challenging. In the future, contributions from different processes could also be incorporated into estimates of global OC-Fe.

In conclusion, in this study, we provided global, spatially explicit estimates of Fe-Al oxides and OC-Fe for topsoil. Our results showed that forests have the largest reserves of Fe-Al oxides, which can promote metal‒organic associations. The global presentation of data concerning various forms of Fe-Al oxides addresses the absence of widespread global data products on extractable soil metal minerals. Analysing their correlation with SOC and integrating them as variables in SOC prediction models will deepen our conceptual comprehension of SOC stability and increase the precision of global soil carbon predictions. The global stock of OC-Fe in terrestrial ecosystems is 232.9 PgC, accounting for 33.5 ± 15.6% of the SOC and 52.3 ± 26.8% of the MAOC in the topsoil (0–30 cm). We found that OC-Fe constituted a significant proportion of SOC and MAOC. Upon estimating the saturation of OC-Fe, we observed that forests and croplands presented considerable potential for increasing OC-Fe stocks. Our estimates of OC-Femax can help guide land management decisions aimed at carbon sequestration. The results of this study provide empirical insights that can guide land management strategies to increase carbon sequestration efforts and mitigate the impacts of global climate change.

Methods

Literature collection and data extraction

To investigate the influence of Fe-Al oxides on SOC, we conducted a search in Web of Science for literature published up to March 2023. We utilized the keywords “soil organic carbon” OR “SOC” OR “soil organic matter” OR “SOM” OR “soil carbon” and “mineral-associated” OR “particulate organic” OR “MAOC” OR “MOM” OR “MAOM” OR “MOC” OR “POM” OR “POC” OR “clay” OR “silt” OR “physical separation” OR “physical fractionation” and “iron” OR “OC-Fe” OR “Iron-bound” OR “Fe” OR “iron oxide” OR “Fe oxide” OR “aluminium oxide” OR “Al oxide” OR “Al”. To prevent the absence of OC-Fe data, we used the keywords “OC-Fe” OR “Fe-bound OC” OR “Iron-bound organic carbon” to supplement the dataset. We retrieved pertinent data from the articles obtained. Although manganese (Mn) is recognized as a significant mineral factor affecting SOC sequestration, few studies have focused on Mn oxides. Moreover, the soil content of the Mn oxides was lower than that of the Fe and Al oxides. Consequently, Mn oxides were not considered in this study.

We applied seven criteria to select suitable studies: (1) SOC or soil organic matter (SOM) concentrations were reported, and SOM concentrations were converted into organic carbon concentrations via a standardized conversion factor of 1.72466; (2) information on sample locations and depths was provided; (3) at least one of the mineral factors (e.g., Fed) was reported; (4) Ald and Fed were extracted via the dithionite-citrate-bicarbonate (DCB) method67, Alo and Feo were extracted with acid-ammonium oxalate68, and Alp and Fep were extracted via sodium pyrophosphate69; (5) OC-Fe was extracted via the citrate-bicarbonate-dithionite method70; and (6) prior to POC and MAOC separation, soil aggregates were dispersed via sodium hexametaphosphate, sonication, or shaking with glass beads to ensure effective breakdown. The POC and MAOC fractions were then separated on the basis of either size or density. If separated by size, the MAOC fraction had to be smaller than 53 µm, whereas if separated by density, the MAOC fraction had to be heavier than 1.60–1.85 g cm−3; and (7) pot experiments, laboratory trials, and model simulations were excluded.

On the basis of the established criteria, we collected 3395 observations from 178 publications and provided point information (Fig. 4A) and biome details (Fig. 4B) for the data reported in the literature. For each study, we collected the following target variables: (1) soil organic carbon factors: SOC (g kg−1; n  =  3392), POC (g kg−1; n  =  395), MAOC (g kg−1; n  =  601) and OC-Fe (g kg−1; n  =  881), and (2) mineral factors: Fed and Ald (n  =  2841 and 1041, respectively), Feo and Alo (n  =  2297 and 1517, respectively), and Fep and Alp (n  =  931 and 597, respectively). Additionally, information on the sampling sites, such as (1) geographical information, including the location of the sampling site (latitude and longitude) and elevation (m; n  =  1657); (2) climatic factors, including the MAT (°C; n  =  2603) and MAP (mm; n  =  2622); and (3) soil physical and chemical properties, including TN (g kg−1; n  =  1126), CS (%; n  =  1897), SBD (g cm−3; n  =  1054), soil pH (n  =  2552) and soil depth (0–286 cm). Data were extracted directly from tables, texts, or publication graphs with the aid of the GetData Graph Digitizer (version 2.24, http://www.getdata-graph-digitizer.com/). Missing variables related to soil properties were supplemented by the equal-area splines method71 from published datasets (e.g., the SoilGrids database) when studies did not provide complete information72 (Supplementary Table S4). Owing to the limited availability of literature, values for PET (mm) and the AI were derived via site-specific coordinate information. Finally, we collected 3395 observations from 178 publications covering a wide range of ecosystem types, such as croplands, grasslands, forests, and shrublands (Fig. 4).

Fig. 4: Spatial distribution of the sampling sites considered in this study.
Fig. 4: Spatial distribution of the sampling sites considered in this study.
Full size image

A Global distribution of the data sites (937 sites) used in this study. Source data are provided as a Source Data file. B Whittaker biome distributions of all sites85.

Evaluating the global spatial patterns of Fe-Al oxides

Fe-Al oxides is thought to be mainly affected by soil physicochemical properties and climate3,73,74. Four climatic variables (MAT, MAP, PET and AI), four soil property variables (TN, CS, SBD and pH) and one geographical variable (elevation) were employed as independent variables to investigate their impacts on Fe-Al oxides. Eight models, including four linear regression models—the multiple linear regression model, multiple stepwise regression model75, least angle regression model76, and elastic net model77,78—and four nonlinear models—the cubist model79, boosted tree model75, bagged tree model80, and random forest model81—were evaluated to determine the best predictive model for Fe-Al oxides (Supplementary Tables S2 and S3). The predictive strength was assessed via a fivefold cross-validation method in which the whole training dataset was randomly divided into five groups. Four of the groups (80%) were designated as training data, with the remaining group (20%) serving as validation data. This process was repeated iteratively until each group had been utilized once as a validation dataset. R² and root mean square error (RMSE) values were used as performance evaluation metrics. In addition to the abovementioned variables, soil depth served as a covariate for predicting Fe-Al oxide concentrations at various depths. To avoid multicollinearity, the variable inflation factor (VIF) of each variable was calculated. If the VIF was greater than five, the variable with the highest VIF was removed, and this process was repeated until all variables had a VIF less than five. The Fe-Al oxide concentration was converted into reserves via soil bulk density, land area and soil depth parameters according to He et al.64.

The Fe-Al oxide concentrations were predicted across five standard depth intervals delineated in GlobalSoilMap25: 0–5 cm, 5–15 cm, 15–30 cm, 30–60 cm and 60–100 cm. The soil carbon predictions were limited to 0–5 cm, 5–15 cm, and 15–30 cm intervals because of the absence of OC-Fe subsoil data. For other depth intervals, such as 0–30 cm or 30–100 cm, mean values were derived by numerically integrating a weighted average of predictions within each interval through the trapezoidal function25. The weighted average of the SOC over the depth interval \([{{\rm{a}}},{{\rm{b}}}]\) can be computed via the following extended formula:

$${{{\rm{SOC}}}}_{{{\rm{avg}}}}\approx \frac{1}{\left({{\rm{b}}}-{{\rm{a}}}\right)}\cdot \frac{1}{2}\sum _{{{\rm{k}}}=1}^{{{\rm{N}}}-1}\left({{{\rm{x}}}}_{{{\rm{k}}}+1}-{{{\rm{x}}}}_{{{\rm{k}}}}\right)\left({{\rm{SOC}}}\left({{{\rm{x}}}}_{{{\rm{k}}}}\right)+{{\rm{SOC}}}\left({{{\rm{x}}}}_{{{\rm{k}}}+1}\right)\right)$$
(1)

where \({{\rm{a}}}\) and \({{\rm{b}}}\) are the boundaries of the depth interval (e.g., 0 cm and 30 cm), \({{\rm{N}}}\) is the number of standard depths within this interval, \({{{\rm{x}}}}_{{{\rm{k}}}}\) is the \({{\rm{k}}}\)-th depth (e.g., 0 cm, 5 cm, 15 cm, or 30 cm), and \({{\rm{SOC}}}({{{\rm{x}}}}_{{{\rm{k}}}})\) is the predicted soil organic carbon value at depth \({{{\rm{x}}}}_{{{\rm{k}}}}\). The term (\({{{\rm{x}}}}_{{{\rm{k}}}+1}-{{{\rm{x}}}}_{{{\rm{k}}}}\)) accounts for the depth difference between adjacent points. The sum aggregates the contributions of each depth segment to the total SOC.

Global soil carbon estimation

After the global data of the Fe-Al oxides were estimated, they were included as predictors of OC-Fe. Six mineralogical variables (Fed, Feo, Fep, Ald, Alo, and Alp), four climatic variables (MAT, MAP, PET, AI), four soil property variables (TN, CS, SBD, and pH), one geographical variable (elevation) and one ecological variable (ecosystem) were chosen as independent variables, and their effects on the OC-Fe concentration were analysed. The updated dataset was used to build eight models for predicting OC-Fe after inputting missing values from global raster data of Fe-Al oxides (obtained from the abovementioned model simulations). The predictive performance was evaluated using fivefold cross-validation, where the dataset was randomly split into five groups. R2 and RMSE were employed as evaluation metrics. The global stocks of OC-Fe were calculated via the global soil bulk density (obtained from the SoilGrids database). The formula is as follows:

$${{\rm{OC}}}-{{{\rm{Fe}}}}_{{{\rm{D}}}}=\sum _{i=1}^{n}{{{\rm{T}}}}_{{{\rm{i}}}}\times {{{\rm{\theta }}}}_{{{\rm{i}}}}\times {{\rm{OC}}}-{{{\rm{Fe}}}}_{{{\rm{i}}}}/10$$
(2)

where OC-FeD represents the carbon density in the soil (kg m−2), n is the number of soil layers, Ti is the thickness of each soil layer (cm), θi is the soil bulk density of each soil layer (g cm−3), and OC-Fei is the OC-Fe concentration of different soil layers (g kg−1). On the basis of the carbon density calculated via the above method, the OC-Fe stock was calculated according to the following equation:

$${{\rm{OC}}}-{{{\rm{Fe}}}}_{{{\rm{stocks}}}}={{\rm{OC}}}-{{{\rm{Fe}}}}_{{{\rm{D}}}}\times {{\rm{S}}}$$
(3)

where OC-Festocks is the OC-Fe stock, OC-FeD is the OC-Fe density (kg m−2), and S is the soil area (m2). Simultaneously, we conducted spatial analysis via ArcGIS 10.8 (http://www.esri.com/), employing a thematic vector map of global ecosystem types to quantify carbon stocks across diverse ecosystem types. We also used the same method used for OC-Fe to generate global datasets for SOC, MAOC and POC. Notably, datasets published by Georgiou et al.7 were incorporated for analysis to assess the impact of Fe-Al oxides on different fractions of organic carbon due to the limited availability of POC and MAOC data. We rigorously assessed the performance of the model and presented uncertainty ranges for the 90% prediction intervals, which were determined from the 5th and 95th quantiles. We employed a Monte Carlo simulation approach to quantify the uncertainty in OC-Fe predictions arising from error propagation in Fe-Al oxides. Using the mean and standard error rasters for each oxide (Fed, Feo, Fep, Ald, Alo, Alp), random perturbations were applied to generate new input rasters for each simulation. These modified inputs were then used in 100 iterations of OC-Fe prediction via a pre-trained random forest model. The resulting predictions were aggregated to calculate the standard deviation and coefficient of variation (CV) for each pixel, thereby quantifying the spatial uncertainty in OC-Fe associated with the variability in Fe-Al oxides (Supplementary Fig. S17).

Boundary line analysis is used to evaluate the maximum response of a variable to finite and independent factors45. The concept of SOC saturation suggests that the quantity of stable SOC is constrained and dictated by the presence of fine particles (clay + silt)82,83. According to the generalized climate classification scheme for aridity index values, areas are classified as arid (<0.2), semiarid (0.2–0.5), dry subhumid (0.5–0.65), or humid (>0.65). To obtain the maximum capacity of Fe-bound organic carbon (OC-Femax), a boundary line analysis method was used for each climate classification (arid, semiarid, dry subhumid and humid)84. Specifically, OC-Fe (y-axis; g kg−1) as a function of clay and silt (x-axis; %) was examined. Given the experimental uncertainty of individual observations, the maximum boundary slope was estimated by a 95th quantile fit (Supplementary Fig. S18). We used OC-Fe observations from clay and silt (<63 μm) size classes to derive OC-Femax7. The mineralogical capacity of OC-Femax was then compared with the predicted OC-Fe stocks to calculate the mineralogical OC-Fe deficit (OC-Femax minus OC-Fe) and OC-Fe saturation (OC-Fe divided by OC-Femax) globally.

Relative importance and linear regression analyses

On the basis of the results of the fivefold cross-validation, the random forest model was selected for concentration prediction (Supplementary Tables S2 and S3). The mean decrease in accuracy (% IncMSE) was used to evaluate the importance of each variable. Compared with other methods, the random forest method has been demonstrated to be effective in reducing overfitting to the training dataset and avoids issues associated with multicollinearity, as observed in multiple regression analysis7. The importance of a variable measure in the random forest model indicates the degree to which the predictor variables influence the outcomes of the model. We standardized each variable on a scale from 0 to 100% to represent their relative importance in relation to the model outcomes (% relative importance) to compare all the independent control variables effectively in the model.

We defined the fsOC-Fe value as the proportion of OC-Fe to SOC and the fmOC-Fe value as the proportion of OC-Fe to MAOC. The values in Table 1 and Supplementary Tables S1 and S5 were computed via “Zonal Statistics” in ArcGIS10.8 (http://www.esri.com/). Linear regression analyses were employed to analyse the relationships between the target variables and the environmental factors. Differences were considered significant at the level of p  <  0.05.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.