Global Spatial Projections of Forest Soil Respiration and Associated Uncertainties

Abstract

The accurate prediction of global forest soil respiration (Rs) is critical for climate change research. Rs consists of autotrophic (Ra) and heterotrophic (Rh) respiration, which respond differently to environmental factors. Predicting Rs as a single flux can be biased; therefore, Ra and Rh should be predicted separately to improve prediction accuracy. In this study, we used the SRDB_V5 database and the random forest model to analyze the uncertainty in predicting Rs using a single global model (SGM) and Ra/Rh using a specific categorical model (SCM) and predicted the spatial dynamics of the distribution pattern of forest Ra, Rh, and Rs in the future under the two different climate patterns. The results show that Rs is higher under tropical and inland climatic conditions, while Rh fluctuates less than Ra and Rs. In addition, the SCM predictions better capture key environmental factors and are more consistent with actual data. In the SSP585 (high emissions) scenario, Rs is projected to increase by 19.59 percent, while in the SSP126 (low emissions) scenario, Rs increases by only 3.76 percent over 80 years, which underlines the need for SCM in future projections.

1. Introduction

Forests, as a major component of terrestrial ecosystems, maintain 86% of the global vegetation carbon pool and 73% of the soil carbon pool, accounting for 45.8% of global soil respiration [1,2,3,4,5,6]. The accurate prediction of forest soil respiration (Rs) on a global scale is, therefore, critical for understanding terrestrial ecosystem feedback in future climate change. However, because global Rs cannot be measured directly, most studies have built Rs prediction models by combining environmental factors (e.g., temperature and precipitation) using machine learning methods, providing the possibility of Rs on a global scale [7,8,9,10,11,12,13]. However, the estimation of global Rs suffers from a large variability. There is currently a difference of 40.6 Pg C year⁻¹ (range: 68–108.6 Pg C year⁻¹) between the highest estimate of soil respiration and the lowest estimate [14]. This significantly limits our ability to balance the global carbon budget and predict climate change [15,16].

There are many reasons for bias in the predicted values of Rs, which can be categorized into two main areas. The first is the reliability of the data. The accuracy of prediction models relies on reliable Rs distribution datasets, which are, however, reduced by the accuracy and range of applicability of measurement instruments and methods (e.g., IRGA vs. the alkaline trap method [7,17,18]) and by the concentration of measurement timescales (models developed based on growing season measurements predict winter Rs) [19,20]. In addition, climate, soil type, vegetation cover, and land use patterns vary across the globe, and differences in these environmental conditions can significantly affect the predicted values of Rs [8,11,21]. Another aspect of the sources of error is model structure and uncertainty. For example, different models differ in their structure and assumptions. Some models, such as the study by Hashimoto et al. [8], may focus on the effects of temperature and moisture, which predict Rs primarily through temperature- and precipitation-driven models. Other models may consider soil organic matter content and microbial activity. For example, the model used by Bond-Lamberty et al. [22] integrates more soil and plant biological parameters for prediction. Some recent studies have also noted that data uncertainty and complexity, as well as the structure and focus of different models, lead to the formation of large errors in global Rs estimates. Jian et al. [23] suggested that the time scale of global Rs prediction models should be changed from annual to monthly to capture seasonal fluctuations. Their study showed that this improvement in the time scale could better reflect the effects of seasonal variation on Rs, thereby improving the dynamic responsiveness of the model and reducing the predictive uncertainty on the time scale. Warner et al. [16] used a quantile regression forest algorithm to predict global Rs at a high spatial resolution. This model can deal with uncertainty and complexity in the data using a large number of input variables and complex algorithms to improve the accuracy of the predictions. This approach considers not only temperature and precipitation but also a variety of factors, such as soil physical properties, vegetation type, and land use. Stell et al. [24] highlighted how spatial biases in information can affect global Rs estimates and suggested ways to improve global projections. Their model adjusts the prediction algorithm to reduce spatial uncertainty by considering the information density and data quality in different regions. By improving the model structure, integrating multiple data sources, and using advanced algorithms, these studies could lead to a more comprehensive understanding and reduction of uncertainty in global Rs projections and significantly improve the accuracy of global Rs projections.

However, current research improvements are only based on integrating the modeling of global Rs observations and analyzing the global distribution pattern of Rs. Because autotrophic respiration (Ra) and heterotrophic respiration (Rh) involve different carbon sources, biological processes, and metabolic pathways, they respond differently to environmental change [3,16,25,26]. Such differences lead to varying sensitivity of ecosystem components to environmental change. Ra is mainly influenced by plant photosynthesis and the growing season and is very sensitive to changes in light, temperature, and moisture, whereas Rh is produced by the microbial oxidation of soil organic carbon (SOC) stored in the soil for long periods of time [27]. If forest photosynthesis and Ra increase in response to climate change, this will accelerate the carbon cycle but will not necessarily lead to climate feedback. Instead, increased heterotrophic activity may release long-stored SOC, resulting in a positive feedback [28]. Therefore, categorical modeling can more accurately capture and quantify the impacts of these different drivers, adjust the model parameters for each respiration type, and reduce errors caused by inaccurate parameter settings. In contrast, integrated modeling based on the same variables amplifies the bias generated by the model estimates, which is not conducive to the understanding of the carbon cycle in terrestrial ecosystems and their response to climate change. Jian et al. [3] used the Global Soil Respiration Database (SRDB) to study the variation in Ra/Rs in different ecosystem types, using different measurement methods, and under different climatic conditions and found that there was no significant trend in the average Ra/Rs across ecosystem types and measurement methods. However, no study has yet analyzed the dynamic distribution patterns of global forest Rh or Ra in response to multiple environmental factors based on these data or further explored the results of categorical versus single models.

Therefore, this study aimed to predict the values and dynamic distribution patterns of global forest Ra, Rh, and Rs based on the latest version of SRDB-V5 observations. The specific objectives were as follows: (1) to analyze the main factors affecting global Ra, Rh, and Rs; (2) to project the spatial distribution of global forest Ra, Rh, and Rs and explore the associated uncertainty between the Rs predicted by a single global model (SGM) and the sum of Ra and Rh (referred to as Rah) predicted by a specific categorical model (SCM); and (3) to project the global data at key nodes for the period between 2020 and 2100 to determine the patterns of spatial dynamics of the relative contributions of the components of forest Ra, Rh, and Rs in two future temperature scenarios to improve the ability to assess the global carbon cycle and how to respond to climate change.

2. Materials and Methods

2.1. Research Data

The forest soil respiration data used for model training in this study were obtained from the latest version of the Soil Respiration Database (SRDB, version 5.0) [29]; compared with the previous version, 4.1 (10,429 observations), 1906 observations were added, and the observation data time covered 1961–2017. We further extracted soil autotrophic respiration data (983 observations) and soil heterotrophic respiration data (1330 observations) in the SRDB dataset, from which a few Ra or Rh missing data were obtained using Rs subtraction. From Figure 1, it can be seen that the observed data cover the five major climate zones of the world [30], which are mainly concentrated in the mid-latitudes, and the distribution locations of Ra and Rh are basically the same. We also counted the distribution of Ra, Rh, and Rs of forests in the SRDB, which had a mean Ra value of 524.64 g C m⁻² year⁻¹, a mean Rh value of 364.43 g C m⁻² year⁻¹, and a mean Rs value of 872.03 g C m⁻² year⁻¹. We also obtained environmental variables relevant to the study of forest Rs (Table S1), and in this study, we extracted the values of the corresponding pixel points for various environmental variables using latitude and longitude data with a spatial resolution of 0.5° latitude × 0.5° longitude. The time of all variables was studied on an annual scale. The global climate zoning data were obtained from the Köppen climate classification [30], which is specifically divided into five climate zones, A (tropical), B (arid), C (temperate), D (continental), and E (polar). The climate data used (annual temperature, annual precipitation, mean annual temperature, and mean annual precipitation) are from the WorldClim2 database; the future climate data use the CMIP6 latest climate model (SSP-RCP) that integrates the shared socio-economic pathways, of which we selected SSP1 as the sustainable development pathway and SSP5 as the fossil fuel development pathway, i.e., SSP126, SSP585. Data from the EC-Earth3-Veg model can be downloaded from WorldClim. Parameters related to the proportions of different soil components (clay, sand, silt), soil organic carbon content, soil bulk weight, and soil C/N ratio were obtained from the SoilGRIDS database (Table S1). Detailed information on global atmospheric nitrogen deposition data can be found in Table S1. Aboveground biomass (AGB) and belowground biomass (BGB) were obtained from Spawn et al. [31]. Global elevation data were collected from the Hole-filled Shuttle Radar Topography Mission [32]. Global leaf area index (LAI) data were obtained from the reprocessed MODIS Global Leaf Area Index dataset [33]. To better explore Ra, we also included Mycorrhizal data from Soudzilovskaia et al. [34]. Finally, we processed each pixel according to the International Geosphere-Biosphere Program (IGBP) to classify global surface vegetation types [35], and based on the classification results, we censored the global forest cover areas and delineated the study area for global forest prediction data.

2.2. Global Forest Rs Prediction Model

We categorized and analyzed the 20 predictors mentioned above into five categories: climate factors, geographic information, soil parameters, ecosystems, and other factors. Based on the relationship that Rs consists of Ra and Rh, we also constructed a random forest model to study the global forest Ra, Rh, and total forest Rs in 2019 to explore the main influencing factors of different respirations and the distribution pattern of forest Rs. The random forest model is a bootstrap nonparametric model based on regression trees, which is now heavily used in the field of ecological research. The main advantages of random forest include (1) the ability to deal with input samples with high-dimensional features and without the need for dimensionality reduction; (2) the ability to assess the importance of each feature in the research problem; (3) the ability to obtain an unbiased estimation of the internal generating error (OOB) in the generating process; and (4) the ability to obtain good results for the default value problem. The specific modeling process is divided into the following steps: First, this study will assess the important factors affecting Rs based on the second advantage of the random forest algorithm, so that a more concise and effective model can be built subsequently. We first selected 20 study-relevant environmental variables from the global forest Rs dataset for model construction (Table S1). We performed stratified random sampling based on different Rs values and used 80% of the data in the dataset as the model training set and 20% as the most model prediction set. To test the stability of the model, we used a 10-fold cross-validation to evaluate the model, in which the training dataset was randomly divided into 10 equal-sized subsets. In each model fitting process, one of the subsets is kept for validation, and the other nine subsets are used for training. The above process was repeated 10 times, and the average of each model validation result (R² and RSME) was taken to characterize the stability of the model. Then, the prediction performance of the Ra, Rh, and Rs models was judged using the prediction set. Finally, we used the constructed random forest model to predict global forest Rs in 2019. In this prediction, we used Monte Carlo methods to bootstrap the model for 200 resamplings and predict global forest Rs, propagating the model error into the global estimates. Model uncertainty was represented by calculating the mean, standard deviation (SD), coefficient of variation (CV), and 95% CI of the predicted Rs values for all corresponding pixel points of the global forest. In addition, we also summed the Ra and Rh predictions of the corresponding pixel points each time to obtain Rah and constructed the SCM of forest Rs, which was investigated in a controlled study with the Rs prediction results (SGM) in order to analyze the prediction performance of the SCM and SGM.

Finally, we extracted the mean annual temperature (MAT) and mean annual precipitation (MAP) and combined them with relatively stable variables, such as climate type, soil bulk weight, and elevation data, and the proportion of different soil components (clay, sand, and silt) were used to estimate the future global distribution of forest Ra, Rh, and Rs in the same steps described above to characterize the dynamics of forest Rs.

All the abovementioned operations were performed in the R program. We completed the parameterization of the random forest model by calling the “RandomForest” package in R (version 4.3.0), with the number of trees set equal to 200 and Mtry set to 5 (Figures S1 and S2).

2.3. Analysis of the Results of Global Forest Rs Projections

The total values of global forest Ra, Rh, and Rs studied in this paper were obtained using area-weighted summation. The total values of global forest Ra, Rh, and Rs were calculated by multiplying the area of each pixel by the value of the corresponding pixel in different climatic zones [36] and summed separately by different climatic zones.

To more clearly investigate the variability of Rah and Rs values in forests on a global scale, we assessed the uncertainty associated with predictions from SGM and SCM, namely, the potential biases in the estimates from different models. First, we obtained Rah based on the Ra and Rh values of each pixel point of the global forest for each prediction and summed the Ra and Rh of the corresponding location (Equation (1)). In order to observe the uncertainty of Rah more clearly, we further calculated the mean, SD, and CV of Rah for each pixel point. Subsequently, the relationship between Rah and Rs was examined to derive the spatial distribution and discrepancies on a global scale, aiming to explore the potential biases and their spatial distribution when different models were used, that is, the associated uncertainty.

$$\mathit{R}\mathit{a}\mathit{h}=\mathit{R}\mathit{a}+\mathit{R}\mathit{h}$$

(1)

$$Difference\_ratio={\displaystyle \frac{\mathit{R}\mathit{a}\mathit{h}-\mathit{R}\mathit{s}}{(\mathit{R}\mathit{a}\mathit{h}+\mathit{R}\mathit{s})/2}}$$

(2)

Here, Rah represents the sum of Ra and Rh estimated using SCM for global forests, whereas Rs represents the Rs values predicted by SGM parameterized with global data. Meanwhile, we used a nonparametric test (Kruskal–Wallis H-test) to determine whether the difference between global forest Rah and Rs is significant.

3. Results

3.1. Main Factors Influencing Rs

A constructed random forest model was used to assess the major influences on global soil Ra, Rh, and Rs. Based on the mean squared error of the model and changes in node purity, the top seven significant variables affecting soil Ra were found to be Soil_sand, LAI, Soil_silt, Soil_BD, MAT, IGBP, and Soil_CN; the top seven significant variables for soil Rh were Soil_silt, MAT, latitude IGBP, B_bio, Soil_sand, and SOC; and the first seven significant variables for Rs were MAT, IGBP, latitude, LAI, Soil_sand, Soil_silt, and MAP. The variables used in the text have been explained in Table S1. Partial dependency diagrams (Figure 2) were also drawn based on the top three most important variables affecting Ra, Rh, and Rs. From the interaction plots in Figure 2a–c, we found that the interactivity of the different variables of Ra, Rh, and Rs reflected the key drivers of the different respiratory processes and that the differences in the interactions of these variables (Vint) indicated the unique characteristics of each of the respiratory processes, as well as the influence of the importance of the different variables (Vimp). According to the results of the interaction, it can be found that sandy soil has a significant effect on Ra (Vimp > 75), and, in addition, according to the relationship between LAI and Ra (Vimp > 70), also reasonably explains the source of Ra. Whereas in the data of Rh, it was clearly observed that soil organic matter content (SOC, B_bio, Soil_silt) was an important factor influencing Rh, which also proves that there are different mechanisms driving Ra and Rh. In addition, MAT, surface cover, and soil properties (Vimp > 80) had more significant effects on Rs (Figure 2c). From the bias dependence diagrams, we can also clearly find how each variable affects Ra, Rh, and Rs. (1) Ra is mainly carried out by plant roots, which is jointly influenced by plant physiology and soil physicochemical properties, and thus shows the same positive correlation trend with Soil_sand, LAI. This is because LAI represents the plant’s photosynthesis potential, which affects the growth and respiration of the root system. Soil_sand influences the soil’s permeability and drainage, which directly affects the root system’s oxygen supply, so Soil_silt showed a negative correlation (Figure 2d–f). (2) Rh is mainly carried out by soil microorganisms and is strongly influenced by the activity of the microorganisms. Soil_sand, Soil_clay, and Soil_silt are variables that together affect the physical structure of the soil and its ability to retain water, thus influencing the microorganisms’ living environment, while MAT and Latitude influence the microorganisms’ activity and the rate of decomposition of organic matter (Figure 2g–i). (3) Rs includes both Ra and Rh, so its drivers are a combination of the two. MAT and IGBP together influence respiratory activity throughout the ecosystem. Vegetation type influences organic matter inputs and soil structure, while temperature influences plant and microbial metabolic activity. Latitude and MAP are factors that together influence climatic conditions and thus plant and microbial activity (Figure 2j–l).

3.2. Distribution Patterns of Forest Rs

Based on 20 variables parameterizing different random forest prediction models (Ra, Rh, and Rs), we predicted global Ra, Rh, and Rs values with a spatial resolution of 0.5, in which the models had better prediction performance (Tables S2 and S3): Ra model test results: R² = 0.94, RSME = 82.44, SD = 46.73, and CV = 11.52%; Rh model test results: R² = 0.79, RSME = 118.34, SD = 72.63, and CV = 12.87%; and Rs model test results: R² = 0.86, RSME = 120.28, SD = 113.99, and CV = 12.29% (Figure S3).

Comparing the predicted global distributions of forest Ra, Rh, and Rs in Figure 3, we find that they show a clear and consistent spatial pattern, with higher values in the Amazon region of South America, South Africa, and the coastal region of Australia and relatively lower values in other regions. To obtain a clearer view of the fluctuations in forest Ra, Rh, and Rs on a global scale we normalized the data. Globally, forest Rh fluctuates less than Ra and Rs, and the values are more evenly distributed (Figure S4). A comparison of different climatic zones revealed that soil Ra, Rh, and Rs per unit area (in the order of A > B > C > D > E) were higher in tropical forests than in other climatic regions; however, when compared in terms of total CO₂ release, A > D > C > B > E. Specific analyses showed (Figure 4) that global forest Ra values were mainly distributed between 39.65 g C m⁻² year⁻¹ and 1574.47 g C m⁻² year⁻¹, with a mean value of 405.68 g C m⁻² year⁻¹, which is higher than the mean value of Ra documented by the SRDB (364.43 g C m⁻² year⁻¹). The values of Rh were mainly distributed between 104.33 g C m⁻² year⁻¹ and 2108.19 g C m⁻² year⁻¹, with a mean value of 564.36 g C m⁻² year⁻¹, which is also higher than the mean value of Rh recorded by SRDB (524.64 g C m⁻² year⁻¹). The values of Rs were mainly distributed between 184.24 g C m⁻² year⁻¹ and 2875.28 g C m⁻² year⁻¹, with a mean value of 927.55 g C m⁻² year⁻¹, which is 55.52 g C m⁻² year⁻¹ higher than the mean value of total Rs recorded by the SRDB (872.03 g C m⁻² year⁻¹). However, it is worth noting that the mean values of Ra and Rh together are larger than the estimated value of Rs (with a difference of 42.49 g C m⁻² year⁻¹), while we found a significant difference between Rah and Rs through the results of difference analysis (Figure 4). This triggered us to think further about what kind of pattern the deviation between SCM and SGM shows globally, which will be analyzed in detail in the next section.

3.3. Limitations and Uncertainties of SGM Predictions

We obtained a global total forest respiration Rah for SCM showing a similar spatial distribution pattern as Rs (Figure 3 and Figure S5). By comparing the uncertainty in the predicted values of Rah (mean CV of 10.51%) with that of Rs (mean CV of 12.29%) (Figure 5a,b), we can find that the SCM (Rah) outperforms the SGM (Rs) as a whole. Moreover, in Figure 5c, we can clearly observe the uncertainty associated between SCM predictions and SGM. Compared to the Rs predictions, Rah is overall larger than Rs but has smaller values than Rs in the tropical rainforest, as well as in most of the temperate monsoon climate (D) regions (Difference_ratio < 0, high CV). It is noteworthy that the annual precipitation in the regions where Rah is smaller than Rs is larger than that of other climate types in the same climate zone. This is because Rs magnitude is strongly correlated with precipitation (MAP) (Figure 2), which is much more relevant than Ra and Rh. As temperature and precipitation decrease with increasing dimensions, the magnitude of Ra and Rh values is mainly controlled by different environmental factors, and thus there is a preference factor. This leads to the fact that at high latitudes, SCM is more favorable in capturing the response of different types of respiration to environmental factors, thus improving the accuracy of predictions. SGM, on the other hand, ignores extreme “hot spots” and “hot moments” in space and time and focuses more on averaging effects. As a result, SCM estimates are generally larger than SGM estimates globally (Difference_ratio > 0), and the CVs of SCM are relatively evenly distributed.

3.4. Future Climate Impacts on Forest Soil Respiration

The models were parameterized using relatively stable external factors to investigate the dynamic distribution patterns of global forest Rs in two future climate scenarios. Weighted sums were then calculated based on the forest area in different climate zones. The results are shown in Figure 6. It was found that in the SSP585 climate scenario, soil CO₂ emissions were significantly higher than in the SSP126 climate scenario, with Rh showing the most substantial increase (Figure 6), followed by Rs and Ra, showing the smallest increase. Compared with current emission levels, the total Ra in the SSP585 climate scenario is expected to increase by 10.45% by 2100, Rh is expected to increase by 23.71%, and Rs is expected to increase by 19.59%. In the SSP126 climate scenario, the total Ra is expected to increase by 3.76%, Rh is expected to increase by 6.54%, and Rs is expected to increase by 1.67% by 2100 (Figure 7). Comparing forest soils in different climate zones, it was found that future climate scenarios show different responses for Ra and Rh, particularly in polar climates, where Ra shows negative growth and Rh shows positive growth. Additionally, in arid and inland regions, the growth rate of Rh was lower than that of Ra, whereas in tropical and temperate regions, the growth rate of Rh exceeded that of Ra.

4. Discussion

4.1. Criticality of Global Forest Rs Prediction

Forest Rs plays a key role in global climate change, and it is particularly important to distinguish between Ra and Rh to study future climate impacts. In this study, temperature and ecotype (IGBP, LAI) were found to be highly influential for forest Rs (Figure 2c), which is consistent with previous findings [37,38,39]. However, the role of soil properties (soil constituents, etc.) will be much higher for forest Ra and Rh, but both do not have the same response to external variables (Figure 2a,b), in which forest Ra is closely related to LAI, with a single-peaked response relationship (Figure 2e) [3], and forest Rh has a positive correlation with mean annual temperature (MAT) (Figure 2h). Because forest soil autotrophs and heterotrophs are based on different generative processes, this also explains the different response results for Ra and Rh, but it means that the feedback effects on climate change may not be the same [14]. Globally, in situ measurements of Ra action in forests are difficult compared to measurements of Rh, which are relatively easier but still challenging [28,40,41]. Although total soil respiration is currently the most well-studied for data-driven global estimates [14,38,42,43], most of the earth system models only output Rh. To improve the estimation of global Rs, more data-driven global estimates of total Rs and its individual components are needed [44,45]. The SRDB contains both Ra and Rh, which contributes to the global estimation of each type of respiration [46,47]. Jian et al. [3] investigated a single, easily measurable flux of Rs through SRDB, estimating the proportion of Ra. However, there is still a need to further clarify the relative relationships between heterotrophic, autotrophic, and total respiration in forest soils globally. Hashimoto et al. [14], in a recent review study, also stated that a response model (SCM) for each soil respiration component is desirable and should be developed.

In addition, for the impact of the SCM model prediction results, it is also critical to select the appropriate characteristic variables of the model [48]. Compared to traditional statistical or process-based modeling approaches, researchers are often required to first formulate correlation hypotheses and select important factors [24,38]. The relationships between global forest Rs and various factors are extremely complex and sometimes unknown, which may mask the “true” relationships. Machine learning-based approaches can effectively deal with the complex relationships between global forest Rs and various factors, especially when combined with interpretable machine learning techniques that can capture the effects of variable interactions on soil respiration and give clear trends [49]. Therefore, future studies could attempt to use interpretable machine learning techniques to take advantage of their benefits to reveal differences in the distribution patterns of Ra and Rh on a global scale, further justifying the need for classification modeling.

4.2. Uncertainty Associated with the Model

In this study, the random forest algorithm was used to estimate global forest Ra, Rh, and Rs, and there are several scenarios that may be highly susceptible to bias in model predictions, mainly because of the uneven spatial and temporal distribution of measurements [48] that may lead to bias in global model predictions.

Firstly, as soil respiration data are a spatial type of data, spatial data may affect model parameterization and interpretation due to their inherent spatial autocorrelation [24,42,50,51]. Because most of the recorded data can be found in Figure 1a to be concentrated in the mid-latitude region, the model results are more biased toward the environmental characteristics of the region [48]. According to the SRDB database, Stell et al. [24] predicted global Rs and found that global estimates were closer to temperate values because temperate data are much more abundant than those from other climate zones. Most measurements used to parameterize the random forest model in this study also came from mid-latitude regions of the Northern Hemisphere (Figure 1). This means that our results may be biased toward data from these regions, which is a common issue that is currently difficult to resolve. Because economically developed countries are mostly located in the mid-latitude regions of the Northern Hemisphere, these areas have more funding and experts supporting field measurements. We urge researchers to increase the measurements in the Southern Hemisphere, which is crucial for future research.

Secondly, the uneven temporal distribution of measurements can lead to prediction biases, as it is particularly difficult to conduct measurements in the Arctic and under harsh conditions during long, cold winters. Most Arctic experiments are conducted only during the growing season [52]. Jian et al. [36] addressed this issue using finer temporal scales, such as daily and monthly, and collected data only at measurement sites during the growing season.

Thirdly, uniform predictions from a single model may lead to biased estimates, as confirmed in this study. The mean values of global forest Ra, Rh, and Rs estimates were 405.68 g C m⁻² year⁻¹, 564.36 g C m⁻² year⁻¹, and 927.55 g C m⁻² year⁻¹, respectively, with values slightly larger than the mean values recorded by the SRDB and in the range of 550–1179 g C m⁻² year⁻¹, as estimated by Zhao et al. [42]. The mean Rs in forest soils is 52.55 g C m⁻² year⁻¹ higher than the estimate of 875 g C m⁻² year⁻¹ reported by Zhao et al. [39]. In addition, we compared the forest Ra, Rh, and Rs and found that there were significant differences in their distribution patterns on a global scale. This also indirectly shows that Ra and Rh respond differently to external factors, indicating the need for predictive analyses using classification models (SCM). Theoretically, the total respiration (Rah), which is derived by summing Ra and Rh at the corresponding locations, should approximate Rs. However, the results showed that Rah was greater than Rs (Figure 4 and Figure 5). This bias is mainly due to the fact that Ra and Rh are based on different processes [28]. Ra is more influenced by plant physiological properties and soil physicochemical properties, whereas Rh is mainly influenced by soil organic matter and microbial activity (Figure 2) [14]. Rs is a combination of the two, and the global prediction of Rs as just a single composite flux hampers the constraining power of the SGM (which is also the reason for the higher CV, Figure 5). Therefore, on a global scale, SCM estimates are usually better than SGM, while SCM is less likely to ignore extreme “hot spots” and “hot moments” in space and time, and its model estimates are usually closer to the actual situation than SGM.

4.3. Future Trends in Global Forest Rs

This study compared future changes in forest Rs in two different climate models. Unlike previous studies, we included precipitation, climate type, elevation data, and soil physical parameters for prediction in our future model [7,11,12,36,37]. As extremely arid soil environments are more sensitive to climate change, there is already evidence that precipitation stimulation tends to lead to more intense CO₂ release following extreme droughts [53,54,55].

The results showed that in the SSP126 climate model, total forest Rs would increase by less than 3.76 percent of the release over the next 80 years, whereas in the SSP585 climate model, the increase would reach 19.59 percent, with Rh rising the most significantly (Figure 6 and Figure 7). This implies that if human CO₂ emissions in the fossil fuel development pathway are not mitigated and the total radiative forcing point is stabilized at 8.5 W m² by 2100, the Rs growth rate at that time will accelerate significantly. The data from this study suggest that total forest Rs will increase at a rate of 0.139 Pg C year⁻¹, which is similar to the global Rs growth rate estimated in previous studies (averaging 0.04 to 0.13 Pg C year⁻¹ [8,22,39]). In contrast, the global growth rates of forest Ra and Rh were 0.03 Pg C year⁻¹ and 0.10 Pg C year⁻¹, respectively. Compared to SSP126, if future humanity adheres to a sustainable development pathway, the total radiative forcing peak will be contained within 3 W m² by 2100, with the global forest Ra release increasing by only 1.67% (average annual growth rate of 0.005 Pg C year⁻¹), Rh release increasing by 6.54% (average annual growth rate of 0.028 Pg C year⁻¹), and Rs release increasing by 3.76% (average annual growth rate of 0.027 Pg C year⁻¹), which is much lower than the values in the SSP585 climate model. This phenomenon confirms the necessity of adhering to a sustainable future development path. In addition, the response of Ra and Rh to future climate also suggests that warming may lead to positive feedback by increasing heterotrophic activity, which in turn releases long-stored SOC [28]. Currently, terrestrial ecosystems remove approximately 1.7 Pg C of CO₂ from the atmosphere annually, which contributes to slowing global warming [56]. However, the sustainability of this capacity as a carbon sink depends on how soil carbon responds to future global warming.

5. Conclusions

Forest Rs plays an important role in global climate change. Machine learning provides a powerful tool for forest Rs prediction; however, due to the different carbon sources, biological processes, and metabolic pathways involved in the Rs component, the feedbacks of the Ra and Rh components to environmental changes are different, which also leads to significant uncertainty in the Rs model. In this study, by using more than 900 forest observations of Ra, Rh, and Rs in the Rs database, we established SGM and SCM to predict global forest Rs separately and analyzed their spatial patterns. There were differences in the distribution patterns of the three at the global scale. The sum of the mean values of forest Ra and Rh exceeded the estimate of Rs (the difference was 42.49 g C m⁻² year⁻¹). These results suggest that compared to SGM, SCM is more favorable for capturing environmental factors that have significant impacts on itself, and its prediction results are closer to the actual situation. To further explore the feedback between the forest carbon cycle and climate change in future climates, we modeled the response of forest Ra, Rh, and Rs in two different future climate models. At the end of the 21st century, according to the SSP585 climate model, the total forest Ra will increase by 10.45%, the total forest Rh will increase by 23.71%, and the increased release from forest Rs will reach 19.59%. This means that not only now but also in the future environment of carbon emission growth, we need to use SCM to better understand the Rs process and constrain the global Rs to reduce the uncertainty of the prediction results brought by the SGM model and improve the accuracy of the climate change prediction.