Substrate and climate determine terrestrial litter decomposition

Significance

Litter decomposition has profound implications for carbon fluxes and nutrient cycling, making it crucial to identify the key factors controlling decomposition rates for ecosystem management and environmental policy. Global litter decomposition rate has historically been hypothesized to be driven by climate. However, by integrating a global dataset from 1,432 terrestrial sites with machine learning approaches, we found that litter substrate and climate determine decomposition rates across spatial and temporal scales, with substrate exerting a stronger direct influence. Future climate change accelerates leaf litter decomposition rates, with smaller effect in low latitudes. Our results highlight the synergistic effect of litter substrate and climate on decomposition rates and have robust implications for modeling terrestrial-atmosphere carbon flux under current and future climate scenarios.

Abstract

Litter decomposition is a fundamental biogeochemical process for carbon flux and nutrient cycling in terrestrial ecosystems, yet the global variation in decomposition rates and their covariations with climate and substrate are not fully understood. Here, we synthesized a global dataset of 6,733 independent observations across six continents to illustrate the climatic and substrate controls over litter decomposition. The average decomposition rates of various litter types ranged from 0.74 to 4.01 y⁻¹ across polar to tropics, showing a large geographical span. Litter substrate and climate directly explained 36 and 30% of the variations in decomposition rates, with the carbon-to-nitrogen ratio identified as the best predictor. In the absence of climate variables, litter substrate can effectively explain the variation, while the model’s predictive capacity decreased significantly after litter substrate was excluded. Our synthesis highlights that a fundamental constraint on litter substrate leads to predictable global-scale patterns of terrestrial litter decomposition rates. Integrating litter chemistry parameters should be prioritized for parameter optimization in Earth system models.

Plant litter contributes approximately 40 to 80 Pg (1 Pg =10⁹ metric tons) carbon per year to soils (1). Litter decomposition contributes to soil organic matter by forming stable organic residues and supports plant growth by releasing nutrients that are taken up by roots (2, 3). By contrast, more than half of the carbon in plant litter is released into the atmosphere during decomposition (2, 4). Therefore, litter decomposition rate (k) not only reflects the turnover time of carbon and nutrients between plant and soil ecosystems, but also serves as a critical parameter for evaluating the atmospheric carbon budget (5). Although considerable studies have evaluated litter decomposition spanning multiple climatic zones in recent years (6–8), the global patterns of litter decomposition rates remain elusive.

Litter decomposition is governed directly by substrate and decomposer community, both of which strongly vary with climate. Variations in environmental factors along a broad range in climate modulate litter decomposition rates by changing soil decomposer activity (6, 9, 10) and litter chemical decomposability (11, 12). Climate shapes the unique environment for litter decomposition in different climatic zones (13, 14) and has traditionally been assumed as the determinant control over litter decomposition rate at global scale (15–17). However, recent studies across climatic zones suggest that this dominance may be overestimated (18–20), and several studies have found that the chemical composition of litter substrate is more important in explaining large-scale litter decomposition rate than climate does (21–23), although this understanding is not yet ubiquitously illustrated (24–26).

In current Earth system models (ESMs), modules simplify terrestrial carbon cycles by adopting the theoretical framework of CENTURY (27–29), which divides several soil and vegetation carbon pools to simulate carbon transfer (30). Each carbon pool is assigned plant functional types specific turnover rates that further modulated by soil temperature and moisture, determine the decomposition rate of organic matter (31, 32). Litter quality is implicitly involved in the split of plant residue into different litter pools based on the lignin-to-nitrogen ratio of the residue (27). However, the simplification of litter decomposition fails to adequately capture the dynamic decomposition rates that are jointly controlled by climate and substrate. This oversimplification limits the applicability and accuracy of these models in predicting carbon and nutrient release from decomposing litter and overlooks the critical role of litter decomposition in global carbon and nutrient flows. Therefore, integrating the coupling of climate and substrate controls over global litter decomposition rates in current ESMs represents a key step toward improving the predictive capacity and applicability.

Here we compiled a global dataset including 6,733 independent observations at 1,432 terrestrial sites (Figs. 1A and SI Appendix, Fig. S1) derived from in situ litter decomposition experiments without any artificial manipulations. These observations span six continents and encompass all major climate zones, enabling us to gain a deeper insight into the role of both climate and substrate in driving litter decomposition rates. We categorized the ecosystem, litter, and vegetation types identified in these experiments and obtained litter substrate data provided by the authors or by retrieving relevant publications at the same experimental sites. We also collected climate, edaphic factors, vegetation, and soil biota data at these study sites from various reliable global databases to assess their effects on litter decomposition rates (Methods). The k value represents the rate at which carbon and nutrients are released from dead litter materials during organic matter decomposition and is estimated using a first-order exponential model (33) by fitting litter mass remaining over time. The objectives are i) to elucidate the global distribution of litter decomposition rates, ii) to discern the direct and indirect drivers of climate, litter substrate, vegetation, and edaphic and biota factors, and iii) to forecast the global patterns of leaf litter decomposition rates.

Fig. 1.

Site locations and annual litter decomposition rates. (A) A total of 1,432 sites across six continents in various terrestrial ecosystems. (B–E) Variations in litter decomposition rates (k) estimated by first-order exponential model in different (B) climate zones, (C) ecosystems, (D) litter types, and (E) vegetation types. The k values were transformed by logarithm, and the boxplots show the median (center line) and interquartile ranges (box). The climate zones were classified by the Köppen climate classification system, including A: tropical/megathermal climates, B: arid climates, Cs: Mediterranean-type climates, Cw: dry-winter subtropical climates, D: continental/microthermal climates, E: polar climates. The term “others” in panel (D) includes flowers, fruits, leaf sheaths, etc.

Climate Effect on Litter Decomposition Rates.

Litter decomposition rates reached a peak at 4.01 y⁻¹ in tropical climates and a lowest rate of 0.74 y⁻¹ in polar climates (Fig. 1B and SI Appendix, Table S4). Environments with warmer temperature, higher humidity, and abundant nutrient inputs, such as in mangroves and agricultural ecosystems, showed higher decomposition rates of 9.59 and 2.95 y⁻¹, respectively. These environmental conditions support higher microbial activity and enzymatic functions, which accelerate litter decomposition rate (2, 9). In contrast, desert and tundra exhibited notably lower decomposition rates of 0.58 and 0.17 y⁻¹, respectively (Fig. 1C). The direct effect of climate on litter decomposition rate is primarily attributed to photodegradation by solar radiation and leaching from rainfall, which either dissolves low-molecular-weight organic matter or breaks lignin compounds down (17, 34). Structural equation modeling (SEM) suggests that climate directly explains 30% of the variation in litter decomposition rates (Fig. 2A). The variations explained by climate reached at 41%, when considering the indirect effect of climate through its influence on vegetation, litter substrate, soil properties, and decomposer community (SI Appendix, Table S5). Climate is coupled with other factors, such as litter carbon and nitrogen contents (35), soil cation exchange capacity (CEC), and soil bacterial and fungal diversity (10), in explaining decomposition rates, showing distinct distribution trends with latitude. Meanwhile, regions in warm and humid climates have higher plant biomass and biodiversity (SI Appendix, Figs. S4 and S13), which are associated with the complexity of plant carbon sources that contributes to litter decomposition. The intricate interactions between biotic and abiotic factors in high biodiversity regions give rise to more complex mechanisms, especially for detrital food webs that are involved in cascading biological interactions (synergistic or antagonistic effects) on litter decomposition rate (36, 37). This dual role of carbon (energy) flow originated from litter materials underscores the intricate interplay of climate and litter substrate in driving global-scale litter decomposition rate.

Fig. 2.

Factors influencing litter decomposition in terrestrial ecosystems. (A) A piecewise structural equation model integrates the direct and indirect effects of climate, plant, soil, soil biota, and litter substrate on litter decomposition rate (k, transformed by logarithm). Orange and green arrows indicate positive and negative effects, respectively. Numbers adjacent to the arrows are path coefficients from partial regressions that show standardized effect sizes, and the widths of arrows are proportional to the path coefficients. (B) Relative contributions of multiple predictors in explaining k. The circles show average parameter estimates (standardized regression coefficients) of model predictors, associated with 95% CI. The relative importance of each factor is shown as the percentage of explained variance. Asterisks denote significant effects (^*P < 0.05, ^**P < 0.01, and ^***P < 0.001). The marginal (m) and conditional (c) R² values represent the proportions of variance explained by all predictors without and with the random effect of sampling site. MAT: mean annual temperature, MAP: mean annual precipitation, bio8: mean daily mean air temperatures of the wettest quarter, NPP: net primary production, EVI: enhanced vegetation index, LAI: leaf area index, LCN: litter initial carbon and nitrogen ratio, LC: litter initial carbon content, LN: litter initial nitrogen content, SOC: soil organic carbon, SFP: soil fungal proportion, SBD: soil biodiversity, N: soil nitrogen content.

Substrate Dominance in Litter Decomposition Rates.

The influence of litter substrate on decomposition rates was evident by the significant variations among litter and vegetation types. High-quality litter, such as leaves, had higher decomposition rate of 1.65 y⁻¹. Conversely, litter with woody structures (e.g. twigs and branches) and litter with high proportions of recalcitrant compounds, showed lower decomposition rates of 0.24 and 0.48 y⁻¹, respectively (Fig. 1D). Furthermore, broadleaved tree species with lower C/N ratios showed significantly higher decomposition rates (1.77 y⁻¹) compared to coniferous tree species (0.38 y⁻¹) (Fig. 1E), while grasses had the highest decomposition rate of 1.95 y⁻¹. The diversity of litter materials may promote synergistic interactions among various soil decomposer communities (38) involved in litter decomposition by providing a greater variety of carbon compounds (39) or by facilitating nitrogen release (40). Additionally, mixed litter often exhibits higher decomposition rates, but such an increase in decomposition rate was only present in the mixture of different litter types (Fig. 1D and SI Appendix, Table S4).

Litter carbon and nitrogen contents and ratios are key variables contributing to the variations in litter substrate. Carbon serves as the primary energy source for microorganisms and directly determines microbial carbon use efficiency (2, 38, 41). When complex carbon compounds are difficult to degrade or nitrogen content is low, both litter decomposition rates and microbial growth are limited by substrate availability. Consequently, litter substrate directly explained 36% of the variation in litter decomposition rates (Fig. 2A). The initial carbon and nitrogen contents and C/N ratio had significant effects on decomposition rates (P < 0.001, Fig. 2B). The random forest model explained 95% of the variation in decomposition rates across study sites (Fig. 3A). Among the predictors, the initial litter C/N ratio emerged as the most critical factor. The k values showed a significant decreasing trend toward higher litter carbon content, while mean annual temperature (MAT) was positively correlated with k (Fig. 3B).

Fig. 3.

Variable importance for predicting terrestrial litter decomposition rates. (A) Increase in the root mean squared errors (IncMSE) from random forest for individual variables. Different colors show various aspects of variables in explaining litter decomposition rate. The insert shows observed and predicted decomposition rates with R² and RMS error (RMSE). (B) Partial dependence plots show the dependence of k on environmental predictors. LCN: litter initial carbon and nitrogen ratio, LC: litter initial carbon content, MAT: mean annual temperature, LN: litter initial nitrogen content, EVI: enhanced vegetation index, MAP: mean annual precipitation, SBD: soil biodiversity, CEC: cation exchange capacity, SOC: soil organic carbon, LAI: leaf area index, NPP: net primary production, SFP: soil fungal proportion, N: soil nitrogen content, BD: bulk density. Detailed information for the bioseries indicators is available in SI Appendix, Table S1.

To disentangle the individual effects of climate and litter substrate, we employed a controlled variable approach to develop models in cases without climate or litter substrate variables separately (SI Appendix, Fig. S14). When litter substrate was excluded, the model’s coefficient of determination (R²) decreased to 0.82, but excluding climate variables resulted in little change (R² of 0.96). Although climate indirectly influences litter substrate to certain extent, the predictive power of litter substrate during decomposition cannot be substituted by climate factors. Collectively, these combined findings suggest that the role of litter substrate in driving decomposition rate is far greater than previously recognized. There is a clear need going forward to change away from the climate-centric paradigm of litter decomposition at global scale.

Multiscale Patterns and Drivers of Litter Decompositions.

The geographical patterns and driving factors of litter decomposition rates are influenced by the complexity and nonlinearity of environmental factors, resulting in variations in decomposition mechanisms across different spatial and temporal scales. The role of microclimate has been found to be weak and cannot predict decomposition rates in the field (7), and only the averages of aggregated microenvironmental variables can support the hypothesis that climate is the primary control over decomposition rate (19). This suggests an overestimate of the influence of macroclimate on decomposition rate at global scale (Figs. 2 and 3). Although the dominant factors controlling litter decomposition rates vary slightly among climate zones, the climate effect declines when the scale is reduced to regional or continental levels (SI Appendix, Fig. S15). By contrast, litter substrate retains a strong explanatory power at regional or smaller scales due to its direct association with decomposer activity, allowing for direct prediction of litter decomposition rates (42, 43). Litter substrate is nonuniform during decomposition, with chemical compounds change over time, and the relative importance of dominant factors change with decomposition (17, 44). However, litter decomposition rates in early, middle, and late stages are primarily regulated by litter substrate globally (SI Appendix, Fig. S16).

We also evaluated the importance of litter substrate and climate across different litter types (SI Appendix, Fig. S17). Our findings revealed that litter substrate was the best predictor of decomposition rate for leaf litter. However, for root litter, climate variables were more important than litter substrate, inconsistent with previous studies (22, 45). We believe that this discrepancy is due to the difference in the optimal substrate predictors between leaf and root litter. Decomposition of root litter is primarily controlled by nonstructural carbohydrates, phenolics, or mycorrhizal types, rather than carbon and nitrogen contents (22, 45). For woody (i.e., stems and branches) and mixed litter, the best predictors of decomposition rates were MAT and litter C/N ratio, respectively (SI Appendix, Fig. S17).

Global Mapping of Leaf Litter Decomposition Rates.

Due to significant differences in the initial substrate chemistry among various litter types, predicting global decomposition rates for different litter types using a single model is challenging. Considering the limited sample sizes for roots, stems, branches, and woods (SI Appendix, Table S4), we focused solely on predicting the global decomposition rates of leaf litter. We found that leaf litter showed substantially lower decomposition rates in regions such as the Himalayas of Asia, the Rocky Mountains of North America, and the Andes of South America compared to other areas at similar latitudes (Fig. 4A). In these regions, lower temperature and precipitation further limit soil decomposer activity, reducing litter decomposition rates. In summary, these differences in distinctive climate should be considered in the improvement of parameters in ecosystem models. Notably, decomposition rates of leaf litter showed a symmetrical pattern, with peaks around 10 °N and 15 °S, but significantly declined around 30 °N and 30 °S (Fig. 4B). Previous studies have found that the decomposition rates of roots, stems, and wood of the same species are positively correlated with those of leaf litter (46, 47), suggesting a parallel pattern in the global distribution of other litter types.

Fig. 4.

Global distributions of leaf litter decomposition rates under current and future climate. (A) Global distribution of decomposition rates (k, transferred by logarithm) at 0.1°-pixel scale. (B) Latitude gradient of predicted decomposition rates of leaf litter under different climate scenarios. (C and D) Percentage changes in leaf litter decomposition rates under (C) SSP126 and (D) SSP585 scenarios during 2070 to 2100. The gray areas have no available data.

Future Climate Change Accelerates Litter Decomposition Rates.

Global decomposition rates of leaf litter are predicted to be accelerated by 10.42 and 21.56% (SI Appendix, Table S6), respectively, under the SSP126 and SSP585 scenarios during 2070 to 2100. The impact of future climate change on litter decomposition rates is not uniform, showing significant regional variability (SI Appendix, Fig. S18). Low-latitude regions between 0 to 15 °N will experience the smallest changes under both scenarios, and those between 0 to 15 °S will show only a marginal increase under SSP126 but exceed 10% under SSP585 (SI Appendix, Table S6). In contrast, most middle- to high-latitudes will experience either a significant acceleration or deceleration in decomposition rates. The regions experiencing significant changes due to climate change are primarily concentrated in areas where decomposition rates are either relatively high or low, such as in south-central Australia, southern Africa, central North America, and central Eurasia (Fig. 4). These regions currently have relatively low decomposition rates, but are predicted to increase significantly under future climate change scenarios. Conversely, regions with high decomposition rates in current climate, such as in Central Africa and Europe, are projected to experience greater declines under changing climate (Fig. 4). It is noteworthy that the various responses of vegetation, edaphic variables, and decomposer community to climate change are difficult to estimate precisely, so we quantified only the direct effect of climate change on litter decomposition rates (SI Appendix, Fig. S19), with an exclusion of the indirect effect.

Our study clarified the relative contributions of climate and litter substrate to decomposition rates, providing insights for improving the simulation of litter decomposition rates in ESMs. Climate and substrate are both critical drivers of decomposition rates at the global scale; however, the influence of climate diminishes substantially at regional and smaller scales, where substrate assumes an equally or even more dominant role (SI Appendix, Fig. S15). This scale dependency further underscores the necessity of integrating both climate and litter substrate into models to predict global litter decomposition rates. Since the predictive role of litter substrate cannot be substituted by climate (SI Appendix, Fig. S14), its representation should be prioritized for parameter optimization in ESMs. Simulations based on simplified climate-driven assumptions fail to capture the critical role of litter substrate in the dynamic decomposition rates over time, reducing the modeling accuracy. Therefore, integrating quantifiable parameters that reflect litter substrate (i.e., C/N ratio) into current ESMs can enable more precise estimates of terrestrial litter decomposition rates and enhance the applicability at both regional and global scales.

In summary, by integrating a global dataset spanning six continents, we elucidated litter decomposition rates across different climatic zones, ecosystems, and litter types in terrestrial ecosystems. We found that litter substrate and climate directly explained 36 and 30%, respectively, of the variations in decomposition rates, while climate also indirectly influenced decomposition through its effects on litter substrate, edaphic factors, and decomposer community. Meanwhile, litter decomposition rates do not consistently decline with increasing latitude but show significant regional differences at similar latitudes, closely linked to ecosystem types and elevation. Our findings provide a quantified relationship between litter substrate and climate with decomposition rates, offering a fundamental parameter for optimizing litter decomposition rates in ESMs.

Methods

Global Litter Decomposition Dataset.

We selected published articles using the following criteria:

• 1)Litter decomposition experiments should be conducted in situ under natural conditions. If there are other manipulations (such as warming, nitrogen addition, CO₂ enrichment, increased or decreased precipitation, etc.) included in the litter decomposition experiments, we only collected the decomposition data at the control plots.
• 2)Decomposition experiments of above-ground litter (such as leaves, twigs, woods, etc.) should be carried out at soil surface, but root litter should be studied at various soil depths.
• 3)Most studies used the litterbag methods (SI Appendix, Fig. S2), and the mesh sizes should be greater than or equal to 1 mm to allow soil fauna, microorganisms, and fine roots entry to litterbags. However, the mesh sizes used in the root litter decomposition experiments are usually smaller than 1 mm to ensure that fine roots are not lost from litterbags. Decomposition rates of leaf litter and roots are generally not affected by the methods used (SI Appendix, Fig. S2 B and C). While there are significant differences in total litter decomposition rates across different research methods, we believe this is due to the fact that direct methods are typically used to study woody litter (SI Appendix, Fig. S2E), which decomposes more slowly, leading to the observed differences among methods.
• 4)If experiments were conducted at different sites with variations in elevation, climate, ecosystem, or vegetation, but only averaged results are available, the literature was not included in our dataset. Averaged data from diverse sites cannot separate the differences between these observations, which are critical for our analysis.
• 5)Not all data are presented in terms of litter decomposition rates. If the articles report litter mass remaining, we used a first-order exponential equation (33) to estimate the decomposition rates. If litter mass loss were reported, we converted these data into mass remaining for calculation. We did not limit the experimental duration of litter decomposition and instead used all available values to calculate k. Most studies show that litter decomposition is rapid in tropical and subtropical regions, usually decomposed within 1 y, so many studies reported only 1 y decomposition data in these regions (48–52). If we only consider decomposition experiments lasting more than 1 y or several years, it would not adequately evaluate the actual conditions in these regions. To ensure the comparability of decomposition rate data, we classified the raw dataset into three stages based on the three-phase paradigm of litter decomposition in previous studies (53), including early (mass remaining of 80 to 100%), middle (mass remaining of 30 to 80%) and late decomposition stage (mass remaining of 0 to 30%). The data points for different decomposition stages are relatively evenly distributed globally (SI Appendix, Fig. S3), meeting our requirements for exploring the global patterns.
• 6)Data was also extracted from figures in the main text or SI Appendix. We used Web Plot Digitizer 4.6 (https://automeris.io/WebPlotDigitizer, Pacifica, CA) to obtain the values presented in figures.

We obtained 6,733 independent experiments from 1,204 articles (SI Appendix, Supporting text) conducted at 1,432 terrestrial sites across six continents (SI Appendix, Fig. S1). We also extracted the site location (latitude and longitude), elevation, ecosystem, litter type, and initial litter substrate (including litter carbon and nitrogen contents and C/N ratio). For literature that does not provide initial litter substrate data, we obtained data from the authors or retrieved them by searching the relevant publications for experiments conducted at the same sites. Moreover, we also categorized the vegetation types (broadleaved trees, coniferous trees, grass, shrub, bamboo, etc.) of the dominant species used for litter decomposition. For vegetation types that are not described in articles, we conducted searches by Wikipedia or World Flora Online (https://www.worldfloraonline.org/).

Variables Controlling Decomposition Rates.

We retrieved climate, edaphic factors, vegetation, soil biota, topography, and land cover data at each site to evaluate the drivers over litter decomposition rate. The selection of variables is based on two main criteria: First, the conclusions proposed by previous studies (7, 19–21, 23), which identified important and representative variables; and second, the availability and reliability of data sources for these variables at global scale.

Climate factors.

We extracted climate data from the CHELSA database (https://chelsa-climate.org/downloads) during 1981 to 2010, including MAT, MAP, and bioclimatic variables (SI Appendix, Table S1). The long-term average data could circumvent the fluctuations due to extreme climates. However, the recorded durations of MAT and MAP data in some literature vary greatly; for example, some data cover 5 y, others for 10 or 20 y, and some do not specify the time span. Therefore, to ensure the comparability of climate data among different sites, we only used the data extracted from global databases and not these provided in publications.

Edaphic factors.

We obtained edaphic factors from the SoilGrids database (54), including soil bulk density, pH, clay content, CEC, soil organic carbon, and nitrogen contents. Given that decomposition of aboveground litter (such as leaves, branches, and woods) is much different with belowground root litter, we used edaphic data at 0 to 5 and 5 to 15 cm depths, consistent with the aboveground and belowground environmental conditions, respectively. These extracted soil data serve to characterize the soil environment at each sampling site, but data availability is very limited in the articles. For example, most of the publications only provided edaphic variables at one depth, which cannot assess the different conditions in surface soils and subsoils. Additionally, most publications provided only pH or soil texture, and the limited data source reduces data availability. Thus, we exclusively employ edaphic data that are extracted from the global dataset and not the publications.

Vegetation factors.

Vegetation factors were retrieved from the MODIS dataset (https://developers.google.com/earth-engine/datasets/catalog/modis), encompassing net primary productivity (NPP, MOD17A3HGF.061), enhanced vegetation index (EVI, MOD13A1.061), and leaf area index (LAI, MOD15A2H.061). We obtained data spanning from 2012 to 2021 to calculate the annual averages during this decade.

Soil biota factors.

Soil biota factors include soil fungal proportion (10) and soil fauna diversity (55). It is noteworthy that data for soil fungal proportion are only available at 0 to 15 cm depth.

Terrain factors.

We utilized the ETOPO global terrain model (version 2022) from the National Oceanic and Atmospheric Administration USA (https://www.ncei.noaa.gov) to extract terrain data, including elevation, aspect, and slope.

Land cover.

The land cover (MCD12Q1.061) data was extracted from the MODIS dataset. Following the bands of LC_Type1 class table, we categorized the land cover types into corresponding ecosystems (SI Appendix, Table S2). Other types, such as urban, water bodies, and permanent snow or ice, were set as NA and not included in subsequent analyses.

Considering that different datasets have different resolutions, we uniformly resample all variables to the minimum resolution (0.1–degree) to ensure data consistency. During data resampling, the nearest neighbor interpolation method is used for categorical variables, while the bilinear interpolation method is applied to numerical variables. This approach ensures spatial comparability across all datasets. The primary packages are raster (56), tiff (57), rgdal (58), ncdf4 (59).

Data Analysis and Model Prediction.

Given that some environmental variables are missing, a total of 4,257 entries are included for correlation analysis, random forest analysis, SEM, and multimodel selection analysis after matching all independent variables with k. Spearman’s p was used to compute correlations between these variables and k (SI Appendix, Fig. S4).

Variable importance.

To elucidate the effects of the five categories of variables (climate, soil, vegetation, soil biota, and litter substrate) on k, we employed a structural equation model to assess both the direct and indirect effects of each variable. A piecewise SEM model, which overcomes the nonnormality or nonlinearity of variables (60), was employed as our data fit normality after logarithmic transformation (SI Appendix, Fig. S5). We utilized Fisher’s C-test (0.05 < P < 1.00) to determine the goodness of the models. Subsequently, modifications were made to the model based on significance (P < 0.05) and model fit indices. These analyses were conducted by using piecewiseSEM (61), nlme, and lme4 (62) packages. The relative effects of various factors on k are determined through multimodel selection. To mitigate the potential effect of high multicollinearity among variables on model performance, the variables with variance inflation factors exceeding 5 are removed. The best model predicting k was selected based on the Akaike information criterion (AIC; ∆AIC < 2) by using the MuMIn (63) package. Finally, we selected the best 7 models (SI Appendix, Table S3) and averaged the models to rule out the instability of a single model. All data underwent standardization before analyses. Random forest algorithm was performed to assess the influence of 38 variables on k (transferred by logarithm) by the h2o (64) package. Model parameters, the optimal number of trees, and the number of variables were sampled randomly as candidates at each split and calculated by using the randomForest (65) package. The model’s performance metrics showed RMSE of 0.21, R² of 0.95, and MAE of 0.15 (Fig. 3A).

Model predictions.

Due to significant differences in the initial substrate of different litter types, it is difficult to use a single model to predict the global k from different litter types. Considering the sample sizes were limited for root, twig, stem, branch, wood, and other litter (SI Appendix, Table S4), we only predicted the global pattern of k values for leaf litter. Moreover, due to the absence of globally available data sources for litter substrate, we first predicted the global patterns of litter substrate based on the dataset, and then derived the global k from these predictions.

To predict the global patterns of leaf litter k, we ran a series of random forest machine-learning models by utilizing the R package h2o. We first used the random forest models to predict global leaf litter substrate, incorporating 33 predictor variables (SI Appendix, Figs. S6A and S7A) into the models. We used grid-search procedure to iteratively explore the random forest model results to train covariates. We set the maximum number of models to 130 (the exact number depends on the time required for training each model and other runtime performance factors), each with hyperparameters including i) the number of trees (50, 100, or 150), ii) number of variables sampled at each split (4 to 14) and iii) minimum observations per leaf litter (2 to 5). We evaluated the performance of each model iteratively by using grid searches and randomized ten-fold cross-validation (66) and calculated the means and SD of the cross-validated models by using the coefficient of determination (R²) for each fold. We selected the best 10 models for predicting the k values of leaf litter based on the highest R² values and averaged their predicted results to create the final maps. The final R² values (based on the set of cross-validated predictions) for initial litter C (SI Appendix, Fig. S6) and N contents (SI Appendix, Fig. S7) were 0.59 and 0.51, respectively.

In predicting the global patterns of leaf litter k, we selected only the top six most important climate variables (bio1, bio6, bio8, bio10, bio11, and bio12, as shown in Fig. 2) due to the high redundancy in the influence of climate variables on k. After including the initial litter carbon and nitrogen contents and C/N ratio, the number of covariates used for predicting leaf litter k was increased to 25 (SI Appendix, Fig. S8). The random forest model remained the same as previously described, with the training parameters kept consistent. Ultimately, the R² value for leaf litter k was 0.82. The model uncertainty was performed based on the predictions of the top 10 models, which were aggregated to calculate the SD.

Predictions under future climate projections.

We predicted the potential changes in leaf litter k for the 2070 to 2100 period under the SSP126 and SSP585 scenarios (SI Appendix, Fig. S9). The future climate data were retrieved from the GFDL-ESM4, IPSL-CM6A-LR, and MPI-ESM1-2-HR models, and we averaged the climate data from these different sources to minimize the potential variability and random errors relying on a single dataset. The data were provided by CHELSA. It is important to note that our study focuses exclusively on how future changes in temperature and precipitation will alter the global patterns of leaf litter k, while keeping other predictor variables constant. We did not account for potential changes in vegetation, soil, or soil biota variables under climate change scenarios. Therefore, our results only capture the direct effect of climate on litter decomposition rates, implying that we may underestimate the overall impact of climate change on decomposition rates.

We employed boosted regression tree (BRT) models to compare the predictive performance between the two algorithms. We set the maximum number of models to 150, each with hyperparameters including i) the learning rate (0.01 or 0.05), ii) maximum depth (3, 4, 5, or 6), iii) sample rate (0.7 or 0.8) and iv) column sample rate (0.3, 0.5, or 0.8). The overall results of the two algorithms did not show significant differences based on the final R² values (0.76) from the BRT model (SI Appendix, Fig. S9), showing a consistent and robust global pattern (SI Appendix, Figs. S10–S12). All analyses were conducted in R version 4.4.0 (R Core Team 2024) (67).

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (CSV)

Code S01 (R)

Code S02 (R)

Data, Materials, and Software Availability

The global gridded dataset of environmental covariates used in this study is available in the public domain (see SI Appendix, Table S1 for the complete list). All other data are included in the manuscript, SI Appendix, and are also uploaded to Figshare (https://doi.org/10.6084/m9.figshare.28268372.v1) (68).