Soil Organic Carbon Stock Prediction: Fate under 2050 Climate Scenarios, the Case of Eastern Ethiopia

Abstract

Soil Organic carbon (SOC) is vital to the soil’s ecosystem functioning as well as improving soil fertility. Slight variation in C in the soil has significant potential to be either a source of CO₂ in the atmosphere or a sink to be stored in the form of soil organic matter. However, modeling SOC spatiotemporal changes was challenging due to lack of data to represent the high spatial heterogeneity in soil properties. Less expensive techniques, digital soil mapping (DSM) combined with space-for-time substitution (SFTS), were applied to predict the present and projected SOC stock for temperature and rainfall projections under different climate scenarios represented by the four Representative Concentration Pathways (RCPs): RCP2.6, RCP4.5, RCP6, and RCP8.5). The relationship between environmental covariates (n = 16) and measured SOC stock (148 samples) was developed using a random forest model. Then, the temporal changes in SOC stock over the baseline were developed for the top 30 cm soil depth of the selected districts (Chiro Zuria, Kuni, Gemechis and Mieso) of West Hararghe Zone at 30 m resolution. The model validation using the random sample of 20% of the data showed that the model explained 44% of the variance (R²) with a root mean square error (RMSE) of 8.96, a mean error (ME) of 0.16, and a Lin’s concordance correlation coefficient (CCC) of 0.88. Temperature was the most important predictor factor influencing the spatial distribution of SOC stock. An overall net gain of SOC stock over the present C stock was expected in the study area by 2050. The gain in areas with the lower baseline SOC stock counterbalanced the loss in areas with the higher baseline stock. The changes in the SOC stock depended on land use land cover (LULC), soil type, and agro-ecological zones. By 2050, cropland is supposed to lose its SOC stock under all RCPs; therefore, appropriate decisions are crucial to compensate for the loss of C.

1. Introduction

Soil holds the largest pool of carbon among terrestrial ecosystems [1]; therefore, it has great potential to be either a source or sink of CO₂, an important greenhouse gas in the atmosphere. In addition, due to its active interactions with the biosphere and atmosphere, and long residence time, SOC was proposed to have the potential to achieve the goal of four per 1000 (launched at the COP21) by increasing SOC stock by 0.4% per year from agriculture for compensation of the global emissions of greenhouse gases by anthropogenic activities [2,3,4]. In addition to climate change mitigation, soil C is vital to other soil ecosystem functions as well as optimizing soil fertility [4,5,6,7,8]. Soil organic carbon enhances soil aggregation and the ability of the soil to hold water. This enhances the adaptive capacity of agriculture to withstand potential impacts of climate change such as drought. The maintenance of SOC stocks in croplands and grasslands of the world therefore creates potent synergies between climate change mitigation and adaptation, as well as enhancing soil biological, physical, and chemical properties to sustain food production [9,10,11]. Therefore, improving SOC maintains soil’s ecosystem services such as providing nutrients and water for plant growth, and boosting agricultural productivity sustainably, ultimately contributing to sustainable development.

The storage of organic carbon in the soil results from the net balance between input of C into soil from the atmosphere and output of carbon in the form of CO₂ to the atmosphere. This process is affected by the complex interactions between soil intrinsic properties, environments factors, and anthropogenic activities. Studies have indicated that biomass production, climate, topography, soil physico-chemical properties, and land use and management influence the processes of SOC gains and losses [12,13,14]. Studies revealed that climate is one of the key drivers of SOC spatial distribution [12,15,16]. Rainfall and temperature play an important role in spatial as well as temporal variations of SOC owing to their effects on both the quantity of organic residue inputs and the rates of decomposition [3,17,18,19]. Therefore, climate change, predominantly, changes in temperature and rainfall, are the key drivers of change in temporal as well as spatial variation of soil C.

Whether SOC stocks increase or decrease under climate change depends upon the balance between the SOC input through an increment in net primary production (NPP) enhanced by increasing rainfall and SOC output facilitated by increasing temperature. Wiesmeier et al. [10] suggested that SOM decomposition rates modify faster as a function of temperature than rainfall. However, Gottschalk et al. [2] stated that the SOC change due to climate change is complicated and depends on the location-specific interactions between soil inherent properties and environmental conditions, as well as land use and management.

Climate model projection showed that an increment in average temperatures is expected under both low and high emissions scenarios in Africa [20]. Projections of rainfall are less certain than projections of temperature in East Africa due to the variation in the small-scale factors controlling rainfall in different seasons; they are difficult to represent well by Global Climate Models (GCMs). Regardless of the uncertainty, higher annual rainfall is also expected in eastern Africa [20]. Ethiopia is already experiencing an increase in temperature and there is a rapid change in observed temperature. A review showed an average rate of increase of 0.28 °C in mean annual temperature per decade between 1960 and 2006 in Ethiopia [21].

In Ethiopia, the agricultural sector is not merely the sector most vulnerable to climate change [22], but it also contributes large amounts of greenhouse gas emissions. The agricultural sector contributed 50% to the total amount of Ethiopia’s annual greenhouse gas (GHG) emissions in 2010 (150 Mt CO₂e); out of this, soil contributed the major share [23,24]. The rapidly increasing population has resulted in the fast expansion of agricultural land to sloppy areas and caused the soil to be susceptible to erosion and loss of carbon. The loss of SOC in the country was related to soil degradation by erosion [25,26]. In addition to releasing a considerable amount of CO₂ into the atmosphere, improper land use triggers the effects of climate change, making SOC sensitive to the potential impacts of climate change.

Through proper management, however, agricultural soil can be a sink for CO₂. Studies demonstrated that improved soil management can reduce emissions and sequester CO₂ from the atmosphere as C in SOM [9,23]. The estimated upper limit mitigation potential of agriculture (if appropriate best management practices are implemented) is estimated to be 5.5–6 Gt of CO₂e per year by 2030 and the major portion (about 89%) of this potential could be attained through sequestration of C in the soil [27]. Agricultural sectors can be resilient to the threats of climate change by adopting a CSA (Climate Smart Agriculture) approach, in which improving SOC plays the greatest role [23].

Prediction of SOC stocks has become a central issue over recent years to understand both the implications in the atmosphere and how the soil’s capacity to mitigate climate change and produce food is affected [13]. In spite of the differential influence of climate change on SOC [28], studies on the issue are rare at fine scales. In addition, modeling SOC spatiotemporal changes is scarce and limited in Africa. This is because it requires a lot of temporal data generated by experiments and indirect methods using simulation models also require several calibrated parameters [29,30]. The high cost associated with obtaining representative SOC data, particularly for a heterogeneous environment such as the study area, means that the measured SOC stock data are rare and fragmented. To solve this, an empirical approach has been proposed that can be used to estimate temporal SOC changes with lower costs [15]. However, this has never been attempted in the country as a whole.

The digital soil mapping (DSM) approach [31] uses easily accessible environmental covariates and different advanced statistical (machine learning) methods to model and map soil properties from a small number of observations. The most common machine learning algorithms applied in DSM are artificial neural networks, support vector machines, classification and regression trees, Cubist, and random forests (RF) [32,33]. Among the machine learning algorithms, the random forest approach had the best performance in predicting soil properties for DSM, which has been employed to generate accurate predictions of SOC in intricate soil-forming environments and with a limited number of soil survey points [32,34,35]. Random forest is a decision tree-based machine learning technique and develops multiple decision trees; therefore, it can handle overfitting [36,37,38]. In addition, it can model non-linear relationships and outperforms the linear regression algorithm [32,36,37,38,39].

The digital soil mapping (DSM) approach [31] that was employed to map soil carbon status at a particular time, can nowadays be applied to predict and map the likely carbon dynamics in the soil by employing the space-for-time substitution (SFTS) approach [15,40]. Space-for-time substitution is a technique used to learn the future path of natural systems from modeling the present spatial pattern and has been applied to study the effects of anthropogenic factors on soil properties [15,28,40]. The temporal SOC prediction can be produced by replacing one or more of the time variable factors with past or future time variables in the SCORPAN model [31]. That is, DSM techniques combined with SFTS processes can map SOC change triggered by climate change. The approach has also been successfully applied to estimate the variations in SOC stock and other soil properties in response to future climate change scenarios in different parts of the world [7,13,28,41]. In this study, DSM techniques combined with SFTS and a random forest model were applied to map the current SOC stock spatial distribution and project SOC stock in the future to analyze the change in SOC stock due to climate change in the selected districts of West Hararghe Zone.

Location-specific data on SOC helps to identify potential C sources and sinks and is thus crucial to identify hot spot sites for informed decision making to adopt appropriate measures that enhance SOC. Prediction of SOC change under future climate projection is essential to discover if the study area soils will act as a net sink or net source of C by 2050. The study also produced information on whether there were variations in the changes in SOC stock by soils, LULC, and agroecological zones or not. These play an enormous role in informed climate smart policymaking on improving SOC to boost agricultural productivity, to mitigate climate change as well as to sustain other ecosystem services from the soil. The objectives of this study were, therefore, to: (i) Predict and map the present/baseline soil organic carbon stocks, (ii) Project SOC stock by 2050 for different climate scenarios represented by the four Representative Concentration Pathways (RCPs): RCP2.6, RCP4.5, RCP6, and RCP8.5, (iii) Quantify the changes in SOC stock over the baseline and map, and (iv) Quantify the variations in the changes of the SOC stock by LULC, soil types, and agro-ecologies.

2. Materials and Methods

2.1. Study Area

The study area falls in West Hararghe, the Eastern Part of Ethiopia (Figure 1). The area spans total area of 4597.39 km² and lies between 40°8′ 57″ and 41°17′40″ E and 8°26′26″ and 9°19′43″ N. The area was selected to encompass the three traditionally classified agro-ecological zones: lowlands (500–1500 m asl), midlands (1500–2300 m asl), and highlands (2300–3200 m asl). Average annual temperature and rainfall of the area range between 20–38.5 °C and 500–1800 mm, respectively. Average annual temperature and rainfall range between 20–38.5 °C and 500–1800 mm, respectively.

The topography of the highland and midland areas in the study area is highly varied and undulating. Mountain forest is one of the remaining patches of forest in the area [42]. The lowland area is predominantly flat and the vegetation is acacia-dominated with some undergrowth of grasses. The dominant LULC in the study area is cropland and rangeland with a tiny percent of forest cover. The farming systems encompass mixed crop-livestock, agro-pastoral, and pastoral practices.

2.2. Methods

The technique of SOC prediction involves different steps. They are: target variable and predictors (environmental covariates) (raster stack) data preparation, overlaying and subsetting target variable and raster stack, fitting spatial prediction models, and finally prediction of the target variable to unsampled places [34,43].

Environmental covariate preparation involves activities such as downscaling or upscaling raster layers to the target resolution (30 m) for preparing a stack, and filtering out missing pixels) [43]. The covariate map (raster) is presented at Appendix A Figure A1. An open-source GIS software SAGA GIS [44], and R packages in R (3.6.3) [45], were used for processing the covariates. The target variable was measured point SOC stock. The workflow of the general methodology is presented in Figure 2 and each data preparation activity is briefly presented in the following subsections.

2.2.1. Soil Organic Carbon Stock Point Data

Considering the factors influencing SOC distribution in soil sampling helps to make efficient use of the amount of samples and helps to reflect the spatial distribution of SOC using a small number of samples [46]. Accordingly, soil sample locations were identified purposively by considering the variations in elevation and slope to capture the existing environmental variability. Similar procedures were suggested to better represent the heterogeneity in environments with a smaller number of soil samples [39,47].

Although prediction accuracy rises with an increasing density of observations, there is no general rule for soil sample density in DSM [15]. In this study, a total of 148 composite (0.5–1 kg) soil samples were collected during October, November, December, and January 2020–2021. Depending on the heterogeneity of the farm, two to three soil sub-samples were collected to make one composite sample. Before taking soil samples, the surface litter was removed from each sampling point. Each soil sample was taken from the surface (0–30 cm) in a zigzag manner. A similar procedure was followed in DSM [48,49,50,51]. A depth of 30 cm was preferred for the reason that it is critical for crop yield, plays an important role in mitigating atmospheric carbon dioxide (CO2), and is the most vulnerable to anthropogenic activities [52,53,54,55].

Disturbed and undisturbed soil samples were taken near each other for SOC analysis and bulk density (BD) estimation, respectively. The former was collected using an auger sampler and the later using a core sampler having a 5 cm diameter and 5 cm height. Site data were recorded from the sampling sites. Geographic coordinates were recorded with a hand-held GPS (global positioning system). Soil samples were prepared following the procedure described in the soil laboratory analysis manual [56]. The air-dried soil samples’ weight was taken before (wt) and after (wf) the separation of fine soil (less than 2 mm) for computation of the course fraction (Equation 2). Finally, the fraction that passes through 0.5 mm was used for SOC analysis. The titration/wet oxidation method was employed to determine SOC content in percent [57], and soil BD was estimated by the core method [58].

The soil organic carbon stock was determined using Equation (1) [59,60,61]

SOCS = d × BD × (1 − CF/100) × SOC (%)

(1)

where SOCS represents SOC stock (t ha⁻¹), d is the thickness in centimeters of the soil layer, BD is the bulk density (g cm⁻³) of the soil layer, SOC (%) is soil organic carbon content in percent, and CF is a coarse fragment in percent.

The coarse fraction was determined using Equation (2) [56]:

$$\mathrm{CF}=\left(\frac{\mathrm{w}\mathrm{t}-\mathrm{w}\mathrm{f}}{\mathrm{w}\mathrm{t}}\right)\times 10,$$

(2)

where w_t represents the total weight of the soil samples and w_f is the weight after the coarse fragment was separated. BD was calculated using Equation (3) [58]:

$$\mathrm{BD}(\mathrm{g}/{\mathrm{cm}}^3)=\frac{\mathrm{m}\mathrm{s}\left(\mathrm{g}\right)}{\mathrm{V}\left({\mathrm{c}\mathrm{m}}^3\right)},$$

(3)

where ms is the mass of oven-dry soil, and v is the total volume of the core sampler calculated as (Equation (4)) [58]:

$$\mathrm{v}=\mathrm{A}\mathrm{h}$$

(4)

where A represents the area of the core sampler, and h is the height.

2.2.2. Environmental Covariate (Predictors)

Soil organic carbon storage or loss is affected by the interaction of several environmental covariates or predictors. Based on a literature review [13,31,62,63,64,65,66,67], 16 (sixteen) predictors which are based on climate, terrain attributes, LULC, soil texture, and soil type were prepared to represent predictor variables (Table 1). The covariate (raster layer) maps are presented in Appendix A.

Digital Elevation Model (DEM)-based terrain parameters are commonly used predictors for SOC stock prediction [31,39,47,62,63,67]. The DEM was accessed from STRM (Shuttle Radar Topography Mission) [68] and downloaded using EarthExplorer (EE) (http://earthexplorer.usgs.gov (accessed on 20 February 2020)) from the archives of the U.S. Geological Survey (USGS) and processed using ArcGIS 10.5 [69]. Finally, the derived terrain parameters were processed from the DEM using geoprocessing tools in the SAGA GIS terrain analysis toolbox [44]. The derived terrain attributes include: aspect (ASP), plan curvature (PLCUR), profile curvature (PRCUR), convergence index (CI), midslope position (MSP), relative slope position (RSP), valley depth (VD), channel network distance (CND), and channel network base level (CNBL), which were used for SOC stock prediction. Therefore, a total of ten (elevation or DEM and the nine derived terrain attributes) were employed to represent terrain attributes for SOC stock spatial prediction. The definition, application, and method of derivation of each of the terrain attributes were described in [70,71,72,73].

Temperature and precipitation are important drivers of SOC storage, affecting both carbon input and its removal through decomposition [11,19]. The gridded rainfall and temperature data (4 km resolution), which were based on ENACTS program [74], were obtained from the Ethiopian National Meteorological Agency (ENMA) to represent the present climate. The data were produced by combining station data with satellite rainfall and reanalysis temperature data products [74]. The average annual precipitation and mean temperature (1990 to 2019) were then employed for prediction of the present SOC stock.

The future climate used for projection of SOC stock included projected annual precipitation and mean temperature (2020–2049). The data were accessed from the Global Climate Model (GCM) climate projections that were used in the IPCC Fifth Assessment report [20]. The use of multiple scenarios and multiple models (or ensemble models) allows for addressing uncertainties in climate projection. Multiple scenarios represent a wide range of possible future evolution of the Earth’s climate and multiple models reduce the effect of model errors [10,75]. A comparison of climate change projections with observed climate data for Ethiopia revealed that Fifth Coupled Model Inter-comparison Project (CMIP5) multiple-model [20,76] mean climate is closer to the observed climate than any individual model [77]. Therefore, future annual precipitation and mean temperature (2020–2049) data projected using CMIP5 multiple-model for the four RCPs (RCP2.6, RCP4.5, RCP6, and RCP8.5) [78] were accessed from CMIP5 multiple-model and processed to predict future SOC stock.

Table 1. Environmental covariates (base model and projection).

	SCORPAN Factor	Covariate	Abbreviatons	Unit	Data Type	Reference /Source/
Terrain parameter	r	DEM/elevatin	ELEV	m	continuous	earthexplorer.usgs.gov (accessed on 20 February 2020)
Climate (current) b	c	average mean annual temperature	TEMPb	°C	continuous	[74]
		average annual rainfall data	RFb	mm	continuous
Climate (2050) p	c	Projected Temperature (RCP2.6, RCP4.5, RCP6 and RCP8.5)	TEMP2.6, TEMP4.5, TEMP6 and TEMP8	°C	continuous	[79]
		Projected rainfall (RCP2.6, RCP4.5, RCP6 and RCP8.5)	RF2.6, RFP4.5, RF6 and RF8	mm	continuous	[79]
Land use land cover	o		LULC	-	categorical	www.mapserver-ethiopia (accessed on 9 February 2020)
Soil texture	s	clay and sand content(%)	CL, SND		continuous	[32]
Soil type	s		SOT	-	categorical	www.mapserver-ethiopia (accessed on 9 February 2020)

b represents the base model; p is the projection model.

Table 1. Environmental covariates (base model and projection).

	SCORPAN Factor	Covariate	Abbreviatons	Unit	Data Type	Reference /Source/
Terrain parameter	r	DEM/elevatin	ELEV	m	continuous	earthexplorer.usgs.gov (accessed on 20 February 2020)
Climate (current) b	c	average mean annual temperature	TEMPb	°C	continuous	[74]
		average annual rainfall data	RFb	mm	continuous
Climate (2050) p	c	Projected Temperature (RCP2.6, RCP4.5, RCP6 and RCP8.5)	TEMP2.6, TEMP4.5, TEMP6 and TEMP8	°C	continuous	[79]
		Projected rainfall (RCP2.6, RCP4.5, RCP6 and RCP8.5)	RF2.6, RFP4.5, RF6 and RF8	mm	continuous	[79]
Land use land cover	o		LULC	-	categorical	www.mapserver-ethiopia (accessed on 9 February 2020)
Soil texture	s	clay and sand content(%)	CL, SND		continuous	[32]
Soil type	s		SOT	-	categorical	www.mapserver-ethiopia (accessed on 9 February 2020)

b represents the base model; p is the projection model.

The projected climate data were accessed from the CCAFS-Climate data portal http://ccafs-climate.org (accessed on 20 March 2020) for 2050 [79]. CCAFS-Climate is a collection of high-resolution (1 km) climate data produced by bias-correcting of GCM outputs [79]. Therefore, the projected (2020–2049 (2030s)) annual mean temperature and annual rainfall of all the existing Global Circulation Models (GCMs) were averaged for each of the four emission scenarios (RCP2.6, RCP4.5, RCP6, and RCP8.5) and used as prediction factor to project SOC stock forward to 2050. Different GCMs were used to project climate under different RCPs [79].

The effect of climate change on SOC stock is different in different LULCs [3,7,32,64]. Therefore, the 2016 LULC map (20 m resolution) was accessed on 9 February 2020 from MapServer Ethiopia (www.mapserver-ethiopia (accessed on 9 February 2020)) to represent the present LULC for prediction of the current SOC stock. The LULC categories were regrouped according to IPCC’s land use categories [52] into forestland, cropland, grassland, and wetlands as defined in [80] and presented in

Table 2. LULC categories.

LULC a	LULC b
tree cover areas	Forestland
shrub cover and grassland	Grassland
cropland	Cropland
vegetation aquatic or flooded and open water	wetland
barelands	other lands

‘^a’ represents land cover classes accessed from MapServer Ethiopia. ‘^b’ represents land cover regrouped according to IPCC definition.

. The LULC map was used to model the current spatial distribution of SOC stock as well as to consider the effect of climate change on agricultural land (cropland and grassland) if the existing LULC continues unchanged in 2050. This was again used to compare the sensitivity of the SOC stock to projected climate in different LULC.

A study conducted in East Africa showed that SOC is strongly affected by soil-inherent properties such as soil texture and soil type, and they have been used to predict SOC spatial distribution [19,47]. A soil type map (20 m resolution) was also accessed from MapServer Ethiopia. According to Hengl et al. [34], predicted soil properties can be used as covariate layers for predictions of other soil properties. The revised prediction map (250 m resolution) of soil properties for Africa [32], which is available for download via http://www.isric.org/data/AfSoilGrids250m (accessed on 15 February 2020), was used to download predicted maps of soil clay and sand content.

2.2.3. Prediction Model

The “randomForest” package (version 4.6–14) [36] in R (3.6.3) was used to establish a relationship between SOC stock and the predictor variables. Random forest (RF) is a machine-learning technique based on the classification and regression tree [37]. In RF, several trees are grown using random samples of predictands and predictors, and then these tree models are aggregated into a comprehensive classifier or repressor [37].

The RF algorithm depends only on three user-defined parameters (tuning parameters), which determine the goodness of the model fitting. They are: the number of trees (ntree) in the forest, the minimum number of data points in each terminal node (nd), and a random subset of prediction factors (mtry) that are chosen randomly on each tree to determine the split at each node [36]. The default values of ntree, nodesize, and mtry are 500, five, and one-third of the total number of predictors, respectively, for regression problems [36].

A study showed that tuning ntree and nodesize did not reduce out-of-bag RMSE significantly [81]. However, the choice of mtry influenced the predictive performance of the RF model [82,83,84,85]. In this study, therefore, the default values of ntree and nodesize, the mtry that was optimized using e tuneRF command found within the random Forest package in R, were employed to fit the RF model to the data. During the generation process of RF, each tree is grown using approximately two-thirds of the training (bootstrap) data, and one-third (called out-of-bag (OOB) data) is used for internal validation [36,38,84]. Then, mtry was tuned depending on the smallest out-of-bag (OOB) error using the internal validation mechanisms, and the mtry associated with the smallest OOB error was used to fit the RF model to the data.

2.2.4. Model Evaluation

According to Liaw and Wiener [36], the RF model has an internal validation mechanism using the OOB error (as it has also been presented in Section 2.2.3) and the OOB mean of squared residuals and R², a statistical measure of how well an out-of-bag prediction approximates the real data points, were calculated automatically. Moreover, data splitting is the most frequently used validation approach used to evaluate map accuracy in DSM [86]. K-fold cross-validation (in which k is 5 or 10) is the most commonly used model validation technique in machine learning [35,87,88]. In this study, the RF model’s performance in predicting the SOC stock was evaluated using 20% of the measured SOC stock point data, based on 10-fold cross-validation.

The model evaluation was conducted using the four most common performance indicators used in model evaluation. They were: coefficient of determination (R²), root mean square error (RMSE), mean error (ME), and Lin’s concordance correlation coefficient (CCC) [86]. R² is the percentage of the explained variance of the response variable, ME measures the average bias in prediction, and Lin’s concordance correlation coefficient (CCC) measures the degree of agreement between predicted and observed or the distance to the 1:1 line [49,89,90]. The first three model evaluation indices (R², RMSE, and ME) were calculated using Equations (5)–(7), respectively. CCC was calculated using Equation (8) [90]. Similar indices were used in similar studies [49,63].

$${\mathbf{R}}^2=\frac{\sum _{\mathbf{i}=1}^{\mathbf{n}}{\left({\widehat{\mathbf{z}}}_{\mathbf{i}}-\stackrel{-}{\mathbf{z}}\right)}^2}{{\sum _{\mathbf{i}=1}^{\mathbf{n}}({\mathbf{z}}_{\mathbf{i}}-\stackrel{-}{\mathbf{z}})}^2},$$

(5)

$$\mathbf{R}\mathbf{M}\mathbf{S}\mathbf{E}=\sqrt{\frac{1}{\mathbf{n}}}{\sum }_{\mathbf{i}=1}^{\mathbf{n}}{({\mathbf{z}}_{\mathbf{i}}-{\widehat{\mathbf{z}}}_{\mathbf{i}})}^2,$$

(6)

$$\mathrm{ME}=1/\mathbf{n}{\sum }_{\mathbf{i}=1}^{\mathbf{n}}({\widehat{\mathbf{z}}}_{\mathbf{i}}-{\mathbf{z}}_{\mathbf{i}}).$$

(7)

where R² is the coefficient of determination; RMSE is root mean square error; ME is the mean error; z_i corresponds to the measured/observed SOC stock of the depth i; $\stackrel{-}{\mathrm{z}}$ is the mean of observed SOC stock, and ${\widehat{\mathrm{z}}}_{\mathrm{i}}$ is the predicted value; n is 0.8 × N and 0.2 × N for training and validation data sets, respectively; N is the total number of samples.

$$\mathbf{CCC}=\frac{2\mathsf{\rho }{\mathsf{\sigma }}_{\mathrm{o}}{\mathsf{\sigma }}_{\mathrm{p}}}{{\mathsf{\sigma }}_{\mathrm{o}}^2+{\mathsf{\sigma }}_{\mathrm{p}}^2+({{\mathsf{\mu }}_{\mathrm{o}}+{\mathsf{\mu }}_{\mathrm{p}})}^2}$$

(8)

CCC is the Concordance Correlation Coefficient, $\mathsf{\rho }$ is the Pearson correlation coefficient between observed and predicted values; ${\mathsf{\sigma }}_{\mathrm{o}}^2$ and ${\mathsf{\sigma }}_{\mathrm{p}}^2$ are observed and predicted variance, respectively; and ${\mathsf{\mu }}_{\mathrm{o}}$ and ${\mathsf{\mu }}_{\mathrm{p}}$ are observed and predicted mean, respectively. All the calculations were done using Microsoft Excel and R software.

2.2.5. SOC Stock Projection

The SCORPAN approach [31] can be used to project SOC stock in the future climate scenarios based on a calibrated SCORPAN model. The empirical models built between measured soil properties and the spatially varying driving factors at the present time can be extrapolated to future time by substituting the time-varying factors [15,40]. In this study, the change in climate over space is substituted for the change over time, and other factors such as topography, bulk density, and soil inherent properties are relatively stable [7,28,59]. In this study, the base model (Section 2.2.3) was rerun by replacing the baseline climate with the projected climate (2050) to project SOC stock to 2050 using the space-for-time substitution approach [15,40].

Projected climate grids (2050) using multi-model ensemble GCM for the four emission scenarios each were then substituted for the baseline climate data in the base model to prepare predictive digital maps of projected SOC stock for the RCPs. The assumptions are that the drivers of the spatial patterns also drive temporal changes [40], the complex relationship between soil organic carbon and its drivers is time-independent, and the C level is at a steady-state condition (amount of input is constant) [15].

2.2.6. Changes in Soil Organic Carbon Stock

In this study, we used DSM techniques [31] combined with a space-for-time substitution approach [15,40] to examine the potential change in SOC stock due to climate change for 2050 in the selected districts of West Hararghe. The spatial extent of SOC stock changes due to climate change from near current (2020) to near future (2050) was calculated as the difference between projected and baseline SOC stock maps for each 30 m pixel using the Map algebra tool found in Arc Toolbox of ArcGIS 10.5. The difference between predicted baseline and projected SOC stock (projected minus baseline) indicated if there was a negative (loss) or a positive (gain) change in SOC stock which estimated emission or sequestration of carbon, respectively. The changes in SOC stock were stratified into different agro-ecologies, LULC, and soil types to assess the variation in SOC stock changes in the strata.

3. Results

3.1. Descriptive Statistics of Measured SOC Stock

The measured SOC stock in the study area ranged between 7.97 t ha⁻¹ and 79.49 t ha⁻¹ and showed high variability (CV = 47.40%) [91]. The mean of SOC stock was 36.67 tonnes per hectare (t ha⁻¹⁾ in rangeland, 41.54 t ha⁻¹ in cropland, and 52.56 t ha⁻¹ in forestland. The result showed the mean SOC stock among the LULC varied in the order forest > crop land > rangeland. The data have a skewness coefficient (calculated using skew and function in Excel) of 0.4 t/ha, respectively.

3.2. SOC Stock Spatial Distribution and Drivers

The predicted SOC stock values were in the range of 17.35 and 70.67 t/ha, with a mean of 38.76 t/ha, a standard deviation of 10.63 t/ha, and a CV of 27 percent in the whole area. The overall SOC stock within 0–30 cm of depth was 240.17TgC. The prediction map of baseline SOC stock was produced at a resolution of 30 m × 30 m for 0–30 cm depth (Figure 3). The level of influence of predictors on SOC stock spatial distribution was measured using a Gini-based variable importance measure. The variable importance ranking result (Figure 4) showed the variations in the influencing level of the covariates in the spatial distribution of SOC stock. Among the covariates, baseline temperature (TEMPb) was the most influencing factor, followed by elevation (ELV) and soil type (SOT). The baseline rainfall (RFb) was the next important factor influencing SOC stock distribution. Derived terrain attributes such as VD (valley depth), CNBL (channel network base level), RSP (relative slope position), CND (cannel network distance), PLCUR (plane curvature), and PRCUR (profile curvature) were found at medium-level of the variable importance ranking. These were linked to soil re-distribution through erosion and deposition that affects SOC spatial distribution [92]. Aspect (ASP), CI (convergence index), and Mid-slope position (MSP) are still important but at a lower level of importance. Lastly, land use land cover (LULC) is at the lowest rank in variable importance ranking, showing little influence in SOC stock prediction.

3.3. Model Performance

At the initial stage, the model’s internal OOB evaluation was used for parameter tuning. Parameter optimization (mtry selection) was performed for the best fitting of the model. For the best prediction accuracy, the best mtry was selected in terms of the lowest OOB error. The result showed that OBB error was lowest at mtry = 10. Then, the model was fitted using the default ntree, nodesize, and the tuned mtry (10). The RF model internal validation result showed OOB prediction mean square error and the OOB variance explained (R²OOB) were 215.74 and 43 %, respectively. The base model performance evaluation was again conducted using 10-fold cross-validation (

Table 3. Random forest model performance, calibration, and validation results by 10-fold cross-validation.

Indices	Sample Number (n)	R2 (%)	RMSE(t C)	ME	Lins CCC
Training dataset	118	56	6.88	1.6	0.91
Validation data set	30	44	8.96	0.16	0.88

). The data were randomly split into training and validation data sets (80% and 20%, respectively) using the sample function in R. The result showed (Table 3) an R² of 56 and 44% for training and validation datasets, respectively. The RMSE and Lin’s CCC are also presented in Table 3 for training and validation data sets.

The scattered plots of predicted versus measured SOC stock for training and validation datasets are presented in Figure 4a,b, respectively. CCC quantifies how far the measured data deviated from the 1:1 line [90] (Figure 5).

3.4. Baseline SOC Stock by Soil and Land Cover Types

Predicted SOC stocks for LULC classes (IPCC-based) and soil groups (based on MapServer Ethiopia) were calculated using zonal statistics of the spatial analyst toolbox in ArcMap (

Table 4. SOC stock baseline and changes by LULC and soils.

	Baseline SOC Stock			SOC Stock Change in Climate Change Scenarios
				RCP 2.6		RCP 4.5		RCP 6		RCP 8.5
LULC	Mean (t/ha)	Area (%)	Total (Tg)	Mean (t/ha)	Total (Tg)	Mean (t/ha)	Total (Tg)	Mean	total	Mean	total
forestland	36.87	2.57	5.87	3.09	0.49	2.12	0.34	2.82	0.45	3.45	0.55
frassland	36.22	48.09	107.93	3.40	10.12	1.80	5.35	2.66	7.92	3.71	11.05
cropland	41.35	49.15	125.94	−0.31	−0.93	−0.66	−2.00	−0.71	−2.15	0.04	0.14
wetland	28.95	0.0003	0.0005	0.80	1.29 × 105	1.26	2.02 × 105	1.58	2.52 × 105	1.97	0.00
other lands	39.56	0.028	0.068	3.03	0.005	1.55	0.003	2.10	0.004	3.46	0.01
settlements	38.19	0.15	0.36	−0.76	−0.01	−0.52	−0.01	−0.50	−0.01	−0.70	−0.01
total		100	240.17		9.68		3.69		6.22		11.73
soil groups
Vertisols (VR)	52.48	12.01	39.04	0.78	0.53	−1.12	−0.76	−1.87	−1.27	2.07	1.41
Calcisol (CL)	55.87	1.01	3.51	−2.39	−0.14	−4.55	−0.26	−5.19	−0.30	−1.59	−0.09
Cambisols (CM)	30.30	27.25	51.17	1.47	2.27	1.19	1.84	1.92	2.95	1.27	1.96
Regosol (RG)	28.06	6.64	11.54	3.45	1.29	1.96	0.74	3.54	1.33	2.81	1.06
Leptosol (LP)	40.61	37.95	95.48	2.28	4.90	1.09	2.35	1.66	3.56	2.72	5.84
Fluvisols (FL)	32.19	2.98	5.94	6.00	1.01	3.56	0.60	4.88	0.82	7.39	1.24
Luvisols (LV)	45.14	11.13	31.13	−2.04	−1.29	−2.23	−1.41	−2.61	−1.65	−1.59	−1.00
Nitosols (NT)	33.12	0.70	1.44	6.00	0.24	6.42	0.26	5.40	0.21	6.77	0.27
N/A	44.56	0.33	0.92	1.36	0.03	0.89	0.02	0.68	0.01	1.32	0.02
Total		100	240.18		8.84		3.37		5.68		10.72
Districts	38.25	11.56	27.39
Chiro Zuria	50.03	17.42	54.03	0.88	0.63	1.70	1.22	1.55	1.11	1.11	0.79
Gemechis	47.22	24.73	72.33	−2.41	−2.60	−3.54	−3.82	−3.58	−3.86	−2.21	−2.38
Kuni	30.13	46.29	86.40	0.08	0.13	−1.34	−2.06	−1.40	−2.14	0.74	1.13
Mieso	38.25	11.56	27.39	4.02	11.53	2.91	8.35	3.87	11.11	4.25	12.19
Total		100	240.15		9.68		3.68		6.21		11.72

). The predicted SOC stock for the LULC classes showed that cropland contained the highest mean SOC stock (41.35 t/ha), followed by forestland (36.87 t/ha) and grassland with the least (36.22 t/ha). Ninety-seven percent (97%) of the total SOC stock was stored on agricultural land (cropland and grassland). The mean SOC stock for dominant soil groups ranged between 28.06 t/ha (Regosol (RG)) and 55.87 t/ha (Calcisol(CL)), which cover smaller areas of 6.64 and 1.01%, respectively (Table 4). Vertisols (VR) and Calcisol (CL) are soil groups that have relatively higher SOC stock. The most dominant soil types (which relatively cover higher areas) in the study area, Leptosols(LP) and Cambisols (CM) which are dominantly found in Mieso district, contained 40% and 21% of the total SOC stock, respectively. The distribution of the mean predicted SOC stock across the districts showed a maximum value of 50.03 t/ha in Gemechis, followed by Kuni. Although the mean of SOC stock in Mieso was the lowest (30.13 t/ha) (Table 4), due to covering the largest area, it contained the highest total SOC stock (36%) of the total stock in the whole study area.

3.5. SOC Stock Projection

A prediction map of baseline SOC stock was produced at a 30 × 30 m resolution for 0–30 cm soil depth (Figure 3). Then, the predicted baseline SOC stock was transformed forward (based on the projected climate) to produce a projected SOC stock prediction map. The baseline predicted SOC stock values were summarized in Section 3.2. A projected prediction map of SOC stock was produced for 2050 climate scenarios (under the four RCPs: RCP 2.6, 4.5, 6, 8.5) at a resolution of 30 m × 30 m for 0–30 cm depth (Figure 6).

The general patterns of the projected SOC stock maps (Figure 6) are similar to that of the baseline stock map (Figure 3). However, areas with lower SOC stock in the baseline map (such as the Northwestern and Southeastern parts of the area) have relatively higher stocks of C in the projected SOC stock maps. The descriptive statistics of predicted baseline and projected SOC stock values were presented in

Table 5. Summary of baseline and projected SOC stock.

	Minimum (t/ha)	Maximum (t/ha)	Mean (t/ha)	STD	CV	Total (Tg)
Baseline	17.35	70.67	38.76	10.63	27	240.15
RCP 2.6	17.08	70.36	40.32	9.20	23	249.84
RCP 4.5	17.85	71.34	39.36	8.90	23	243.84
RCP 6	17.75	70.29	39.76	8.47	21	246.38
RCP 8.5	17.08	72.19	40.65	9.46	23	251.89

. The mean and total of the projected SOC stock (t/ha) in each RCPs are higher than that of the baseline SOC stock (Table 5).

3.6. Change in SOC Stock under Future Climate

The result of the comparison between the projected and baseline climate showed an increment in temperature and rainfall in the near future (by 2050) for the four emission pathways. Although the atmospheric CO₂ concentrations are lower by 2050 than by 2100 [20], the result of the analysis showed that there is still a warming trend in temperature. The projected climate in 2050 showed a mean increase in average annual temperature by 1.25, 1.5.1.85, and 1.85 °C, and a mean increase in annual rainfall of 250.94, 257.33, 249.31, and 256.601 in mm for RCPs 2.6, 4.5, 6, and 8.5, respectively, in the study area. The spatial distribution of changes in SOC stock as a result of changes in temperature and rainfall for the four RCPs was presented in a map (Figure 7a–d).

The difference between projected and baseline SOC stock prediction showed an overall additional mean gain of 1.56, 0.59, 1, and 1.89 t/ha of SOC stock under RCP 2.6, 4.5, 6 and 8.5, respectively, over the baseline (38.76 t/ha) in soils of the study area by the end of 2050 (

Table 6. Summary of overall changes in SOC stock by RCP.

RCPs	Mean Change (t/ha)	Total Change (Tg)
2.6	1.56	9.68
4.5	0.59	3.68
6	1	6.21
8.5	1.89	11.73

). From the result, it can be observed that the effect of projected climate on SOC stock was weak (less than 5% of the mean baseline stock). That is, the maximum change observed was 11.73 TgC under RCP 8.5, which is 5% of the baseline SOC stock.

SOC Stock Changes by LULC, Soil Type, and Agro-Ecologies

Predicted SOC stock change for 2050 varies by soil groups, LULC, and agro-ecological zones. The gain (positive values) and loss (negative values) of SOC stock in each LULC, soil group, and agro-ecology are presented in Table 4. Between the agricultural lands, grassland gains SOC stock under all RCPs. However, cropland loses its SOC stock under all RCPs, except RCP 8.5 by a maximum mean rate of 0.71 t/ha under RCP 6 (Table 4). However, the change in SOC stock in the LULC classes varies by agro-ecological zones (highland and lowlands) (

Table 7. SOC stock changes by agro-ecological zones.

LULC	Mieso (RCP 2.6)		Gemechis (RCP 2.6)		Mieso (RCP 4.5)		Gemechis (RCP 4.5)
	Mean (t/ha)	Total (Tg)	Mean (t/ha)	Total (Tg)	Mean (t/ha)	Total (Tg)	Mean (t/ha)	Total (Tg)
forestland	4.95	0.50	−0.37	−0.0072	3.77	0.38	−1.19	−0.023
grassland	4.83	8.82	−1.83	−0.7524	3.22	5.89	−3.30	−1.356
cropland	2.35	2.21	−2.83	−1.8396	2.22	2.08	−3.77	−2.445
wetland	0.84	1.17 × 105	−2.76	−0.0003	1.33	1.86 × 105	−3.26	−0.00034
other lands	1.51	0.0009	−1.78	−0.0006	0.34	0.000216	−1.61	−0.00054
settlements	1.17	0.0035	−0.37	−0.0072	0.54	0.0016	−1.19	−0.023

). For example, all LULC classes lose SOC under RCP 2.6, and 4.5 in Gemechis (highland), and cropland loses the largest mean stock (3.77 t/ha under RCP 4.5) (Table 7). However, in Mieso (lowland), all the LULC classes gain SOC stock, although cropland gains the least (of forestland and grassland).

Among the dominant soil groups in the area, Calcisol (CL) and Luvisols (LV) are expected to lose SOC stock by 2050 under all emission pathways (Table 4). Calcisols lose the largest SOC stock (−5.19 t/ha) under RCP 6. Vertisols (VR) is expected to lose a mean of −1.12 and −1.87 t/ha of C under RCPs 4.5 and 6, respectively. The rest soil groups are expected to accumulate SOC stock. The largest carbon gain was expected in Fluvisols (FL) at a rate of 7.39 t/ha by 2050.

4. Discussion

4.1. SOC Stock Spatial Prediction

Temperature was the most important factor influencing the spatial distribution of SOC stock followed by elevation (ELV) and soil type (SOT). LULC was the predictor least influencing SOC stock spatial distribution. The reason for this may be LULC combined with different factors such as terrain parameters, soil types, and climate variables, which might have covered its influence [59,93,94]. The spatial distribution of SOC stock in the prediction map (Figure 5) showed that the SOC stock density was highest in Gemechis, southern Chiro, and the northwestern part of Kuni. These areas are located in higher elevations and cooler climates. The lowest predicted SOC stock density values were predominantly found in Mieso district, which is a moisture-limited area. The performance (validation dataset) of the RF model fitted using the soil samples and the environmental covariates with respect to the explained variance (44%) was comparable to several DSM studies of similar soil depth [7,49,59,63,90]. Therefore, it can give an insight into the expected changes in SOC stock due to climate change.

4.2. Predicted Change in SOC Stock in a Changing Climate

The change in SOC due to the change in climate showed an overall net increase (gain) in SOC stock above the baseline prediction by 2050 in the area; the loss of C in one area is compensated by the gain in other areas. However, the effect of the changes in climate on SOC stock was weak (less than 5%) under all the RCPs. The soil in areas with lower baseline SOC stock (such as the Mieso district) became accumulation areas, while the soil in the areas of higher stock (Gemechis and Kuni districts) is expected to lose carbon. Soil carbon sequestration depends on the initial C stock and higher SOC sequestration occurs in soils with low initial SOC content [11,95]. This is due to the fact that soils with high initial SOC content make a good environment for microbial activities and speed up SOM decomposition rate [93].

There were variations in SOC stock change among LULC categories, soil types, and agro-ecological zones in the area. Cropland loses while grassland gains SOC stock by 2050 under all RCPs in the whole area. This may be due to the difference in the quality of litter added to the soil in cropland and grassland [96]. In addition, the change in SOC stock in the LULC classes depends on the agro-ecology zones. A possible reason may be that an increase in temperature favors greenhouse gas emission from the soil with higher elevation as stated by Alani et al. [97]. Since temperature was the strongest influencing factor, the effect of its change on decomposition may be stronger than the effect of changes in rainfall on NPP.

Among the soil groups, soils with higher baseline SOC stock (Calcisol (CL), Luvisols (LV), and Vertisol (VR)) are expected to lose their carbon by 2050 under all the RCPs. Among the soil groups, however, Vertisol (VR) gains SOC under RCP 2.6 and 8.5. The reason for this could be that the high clay content in the soil slows down the SOM decomposition since the clay content was also one of the important covariates in the SOC stock prediction. Studies showed that, under the same environmental conditions, high clay content stabilizes SOM in soil by trapping soil organic matter in its small space and creating organo-mineral complexes on its large area [11,98,99]. The rest of the soil groups (having lower baseline SOC) are expected to accumulate SOC stock.

In general, the model projection showed a gain of SOC in areas with lower baseline SOC stock in the study area. The reason may be that the soil is far from saturation; therefore, it has higher potential to sequester C. Top soil with low initial SOC stock is a potential sink for C to contribute to achieving the COP21 ambition of the four per 1000 goal to compensate for the global emissions of greenhouse gases by anthropogenic sources [4].

The study found that loss of C is expected in cropland soil by 2050 in all climate scenarios. Therefore, unless the predicted loss of C in cropland is balanced with an increase in C input, the soil will be a source of emission and at the same time cause a decrease in the productivity of crops, which again decreases C input to the soil. This will also be exacerbated if the existing land mismanagement (such as cultivation of sloppy areas for lack of land) continues making the soil more sensitive to impacts of climate change such as increasing soil erosion. Therefore, by implementing best management [4] practices along with appropriate land use, the expected loss of C in cropland can be compensated. Finally, the method can be applied for the whole country to monitor the change in SOC stock and report the emission and sequestration of carbon in agricultural soils under the framework of UNFCCC.

4.3. Limitations of the Study

One limitation of the study was that CO₂‘s fertilization effect was not considered in predicting the effect of climate change on SOC stock. However, the increase in CO₂ in the atmosphere, in addition to changing the climate system, may also increase the NPP due to the CO₂ fertilization effect. Therefore, future studies are suggested to consider CO₂ fertilization effect using various crop modeling approaches and compare the change in SOC stock in response to climate change of the area with and without CO₂ fertilization effect. In addition, projection to the future of SOC stock lacks validation due to lack of observed data in future time.

5. Conclusions

This study applied a space-for-time substitution approach along with random forest model for analysis of the combined effects of increased temperature and rainfall by 2050 (as projected by climate models) on SOC stock. The model performance was comparable with similar studies; therefore, it can give insight into the expected changes of SOC stock due to climate change by 2050. The following can be concluded from the study:

• An overall net increase (gain) in SOC stock of less than 5% above the baseline is expected by 2050 in the whole study area under all the RCPs (of the multi-model);
• Areas with lower baseline stock become accumulation areas, while the areas with higher initial stock lose SOC stock;
• The SOC stock change in response to climate change in the area varies by LULC;
• Cropland soil is expected to lose C while that of grasslands is expected to gain C in the study area under all RCPs;
• The gain and loss of carbon in the LULC classes depend on agro-ecological zones;
• SOC change in the study area depends on soil types, and soil groups having higher clay content will lose less SOC stock by 2050.