Access the full text.
Sign up today, get DeepDyve free for 14 days.
Loading next page...
/lp/taylor-francis/modeling-and-mapping-spatial-distribution-of-baseline-soil-organic-CrpcarI8R0
References
References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.
- Publisher
- Taylor & Francis
- Copyright
- © 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group on behalf of the International Water, Air & Soil Conservation Society(INWASCON).
- ISSN
- 2474-9508
- DOI
- 10.1080/24749508.2023.2167632
- Publisher site
- See Article on Publisher Site
Abstract
GEOLOGY, ECOLOGY, AND LANDSCAPES INWASCON https://doi.org/10.1080/24749508.2023.2167632 RESEARCH ARTICLE Modeling and mapping spatial distribution of baseline soil organic carbon stock, a case of West Hararghe, Oromia Regional State, Eastern Ethiopia a,b c d e f Martha Kidemu Negassa , Mitiku Haile , Gudina Legesse Feyisa , Lemma Wogi and Feyera Merga African Center of Excellence for Climate Smart Agriculture and Biodiversity Conservation, Haramaya University, Dire Dawa, Ethiopia; b c Climate Change Research Directorate, Ethiopian Environment and Forest Research Institute, Addis Ababa, Ethiopia; Department of Land Resource Management and Environmental Protection, Mekelle University, Mekelle, Tigray, Ethiopia; Center for Environmental Science, Addis Ababa University, Addis Ababa, Ethiopia; School of Natural Resources Management and Environmental Sciences, Haramaya University, Dire Dawa, Ethiopia; Alliance of Bioversity International and CIAT, Improving crops theme, Addis Ababa, Ethiopia ABSTRACT ARTICLE HISTORY Received 14 September 2022 Location-specific information on soil organic carbon (SOC) helps to identify potential sources Accepted 8 January 2023 and sinks of carbon. The objective of this study was to produce a 30m-resolution digital map of SOC stock for selected districts in West Hararghe Zone of Ethiopia. 148 soil samples(0-30cm) KEYWORDS were collected for SOC analysis and BD estimation. Eighteen environmental covariates were The soil; random forest; SOC acquired from satellite sources, digital elevation model (DEM), and maps. A random forest stock prediction; model was fitted to the data. The accuracy of the prediction was tested using the 10-fold cross- environmental covariates; validation method. The model explained 36% of the variance (R2) with a root mean square digital soil mapping error of 9.51. The most relevant predictors were normalized difference vegetation index (NDVI), temperature, and elevation. The predicted total SOC stock value was 240.5TgC and ranged between 17.91 and 70.16 t/ha, with a mean of 38.82 t/ha. The soils under agricultural land had a higher amount of carbon than forestland. The mean SOC stock was relatively higher in fine- textured soils and lower in soils, which are dominantly found in drier areas. This study could help provide information on a less expensive method of SOC stock estimation. In addition, the study produced an updatable baseline SOC stock map. 1. Introduction studies on SOC stock assessment using a digital soil mapping (DSM) approach that uses easily available Soil has a significant influence on the concentration of environmental covariates and advanced statistical CO in the atmosphere (Lal, 2004). Improving and methods such as machine learning algorithms are protecting soil organic carbon (SOC) is a global prior- common around the world (Adhikari et al., 2014, ity for climate change mitigation and a local concern 2019; Bangelesa et al., 2020; Bian et al., 2019; for food security (Cowie et al., 2011; Lal, 2004; Lal McBratney et al., 2003). However, studies conducted et al., 2013; Nayak et al., 2019). Organic carbon (OC) using these methods are rare in Ethiopia and have sequestration in soil is thus one of the most cost- never been attempted in the study area. effective options for climate change mitigation, with A few studies on the DSM have started to appear in the additional co-benefit of improving soil fertility, different parts of the country (Birhanu & Chalchisa, particularly in low-input subsistence farming systems 2019; Cerretelli et al., 2021; Temesgen et al., 2022). such as those found in sub-Saharan Africa (Lal et al., Cerretelli et al. (2021) identified the limitations of 2013; Pouladi et al., 2019). Soil with an optimal OC using global organic carbon stock data for local-scale content can absorb and store water and make it avail- interventions in the case of Ethiopia. Birhanu and able for crops under drought conditions (FAO, 2017), Chalchisa (2019) classified and mapped agricultural which improves the adaptation capacity of agriculture land into small, homogenous management units as to the impacts of climate change and subsequently a remedy for a uniform rate of application of nutrient builds its resilience. and chemical amendments in the West Wollega zone Location-specific information on SOC stocks, of Ethiopia. In addition, Temesgen et al. (2022) con- therefore, helps to understand potential sources and ducted a study on mapping the spatial distribution of sinks of carbon (C). However, the extremely high SOC in southern Ethiopia using ordinary kriging (OK) spatial variability of SOC and the high cost associated geostatistical tools. This was an indication of the with conventional monitoring of soil properties are recent inception of the DSM study in the country. among a number of challenges to getting representa- Common machine learning algorithms applied to tive information for a larger area than plots. Recent DSM are artificial neural networks (ANNs), support CONTACT Martha Kidemu Negassa marthanegasa@gmail.com African Center of Excellence for Climate Smart Agriculture and Biodiversity Conservation, Haramaya University, Dire Dawa P.O. Box 138, Ethiopia © 2023 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group on behalf of the International Water, Air & Soil Conservation Society(INWASCON). This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. 2 M. K. NEGASSA ET AL. vector machines (SVMs), classification and regression SOC for climate-smart agriculture as well as other trees, Cubist, and random forests (RF) (Hastie et al., ecosystem services. 2009; Hengl et al., 2015). Among the machine learning algorithms, the random forest approach had the best 2. Materials and methods performance in predicting soil properties for DSM (Gomes et al., 2019). RF (Breiman, 2001) is 2.1. Study area a statistical tool crucial for understanding the complex The study area fell in West Hararghe, the eastern part relations between soil attributes and their site-specific of Ethiopia (Figure 1). The area spans a total area of environmental factors, which have proven to be effi- 4,597.39 km and lies between 40° 8“57”“and 41° 17” cient in mapping soil properties across a wide range of 40”‘E and 8° 26’ 26”“and 9° 19” 43’’ N. The area was data scenarios and scales of soil variability (Hengl selected to include the three traditionally classified et al., 2015; Mayer et al., 2019). It has been stated agro-ecological zones: lowlands (500–1,500 m asl), that the RF algorithm outperforms the linear regres- midlands (1,500–2,300 m asl), and highlands (2,300– sion algorithm, with average decreases of 15–75% in 3,200 m asl). The average annual temperature and root mean squared error (RMSE) across soil properties rainfall range between 20°C and 38.5°C and 500 and and depths (Hengl et al., 2015). This method has been 1800 mm, respectively. used to generate accurate predictions of SOC from the The topography of the highland and midland areas plot to the global scale and also on a country-specific is highly variable and undulating. Mountain forest is basis in complex soil-forming environments and with one of the few remaining patches of forest in the area a limited number of soil survey points (Hengl et al., (Sudi et al., 2018). The lowland area is predominantly 2015; Hengl et al., 2017; Liu et al., 2022; Poggio et al., flat, and the vegetation is acacia dominated by some 2021). undergrowth of grass. The dominant LULC in the The ability of RF to model non-linear relationships, study area is cropland and rangeland, with a small handle categorical and continuous predictors, and percentage of forest cover. Farming systems in the have robustness for SOC stock prediction with low study area include mixed crop-livestock, agropastoral, sampling density has made it promising for DSM and pastoral. (Breiman, 2001; Grimm et al., 2008; Liaw & Wiener, 2002; Yang et al., 2016). In addition, the rapid advancement of remote sensing (RS) technologies in 2.2. Flowchart of digital soil mapping terms of spatial and spectral resolution, eased accessi- The process of SOC stock prediction involves the bility of covariates and open-source software such as R target variable (point SOC stock) and predictors’ and SAGA GIS helped easily application of RF for data preparation, regression matrix preparation, fit- DSM. In DSM, SOC predictor variables (covariates) ting spatial prediction models, and spatial prediction are represented by a combination of one or more and mapping of target variables (Figure 2) (Hengl SCORPAN factors (McBratney et al., 2003), and the et al., 2017). Each step was briefly presented in the interaction of these drivers determines final SOC following sections. levels (Mayes et al., 2014; Xiong et al., 2014). Digital elevation model (DEM) and RS imagery data are common environmental covariates used to predict 2.3. Data preparation SOC (Adhikari et al., 2014; McBratney et al., 2003). The main aim of this study was, therefore, to map 2.3.1. Soil organic carbon stock point data the current spatial distribution of SOC stock, which Taking into account the factors influencing the SOC serves as the baseline for future monitoring of the distribution in soil sampling helps to be effective in the changes. The objectives were to derive a predictive number of samples and improve the samples’ accuracy model, predict and develop a fine-scale map of SOC to reflect the distribution of SOC (Hobley & stock for the study area, and estimate SOC stock by Willgoose, 2010). Accordingly, soil sample locations land use land cover (LULC) and soils of the study area. were identified purposively by considering factors The generated information on the relationships influencing the SOC distribution-like variations in between SOC stock and the covariates helped to pre- elevation to capture existing variability. Similar proce- dict and map the SOC stock’s spatial distribution. The dures were suggested by Hobley and Willgoose (2010); study contributes to knowledge and information on Yang et al. (2016); and Wang et al. (2018) to better methods for rapid estimation of SOC stock using represent heterogeneity with a smaller number of soil easily available covariates without the need for samples. observed SOC stock. Finally, the study produced an Although prediction accuracy increases with an updatable spatial distribution of SOC stock for spa- increasing density of observations, there is no general tially explicit assessments of SOC stock, which can be rule for soil sample density in DSM (Minasny et al., used for informed decision-making on improving 2013). In this study, a total of 148 composite (0.5–1 kg) GEOLOGY, ECOLOGY, AND LANDSCAPES 3 Figure 1. Study area (according to GADM(https://gadm.org/)) DEM superimposed on the locations of sampling sites. Covariates SOC Overlay Stock Regressiom matrix Y X X .. X 1 2 p Model Fit spatial prediction parameters model (random forest) Prediction Predict map Figure 2. Flowchart of DSM for spatial prediction. soil samples were collected during October, crop yield (Sun et al., 2010), plays an important role in November, December, and January, 2020–2021. mitigating atmospheric carbon dioxide (CO ), and is Depending on the heterogeneity of the farm, two to most vulnerable to anthropogenic activities (Ipcc three soil subsamples were collected to make one (intergevernmental Panel on Climate Change) et al., composite sample. Before sampling, surface litter was 2006; Gao et al., 2008; Wynn et al., 2006). removed from the sampling point. Each soil sample Disturbed and undisturbed soil samples were taken was taken from the surface (0–30 cm) in a zigzag near each other for SOC analysis and bulk density manner similar to the procedures described by (BD) estimation using an auger sampler and a core Vågen and Winowiecki (2013); Dharumarajan et al. sampler (having a 5-cm diameter and 5-cm height), (2017); Zhang et al. (2019); and Arunrat et al. (2020). respectively. Site data were recorded from the sam- A depth of 30 cm was preferred because it is crucial for pling sites. Geographic coordinates were recorded 4 M. K. NEGASSA ET AL. with a hand-held global positioning system (GPS). Soil 2019; Malone et al., 2009; McBratney et al., 2003; samples were prepared following the procedure Wang et al., 2018; Li et al., 2019; Yang et al., 2016; described in the soil laboratory analysis manual Yigini & Panagos, 2016), 18 predictor variables (which (Reeuwijk, 2002). The air-dried soil samples’ weight include climate, terrain attributes, Normalized differ- was taken before (wt) and after (wf) the separation of ence vegetation index (NDVI), land cover/use, soil fine soil (less than 2 mm) for estimation of the course texture, and soil type) were prepared to represent fraction (equation 2). Finally, the fraction that passes predictor variables. The list of predictors and their through 0.5 mm was used for SOC analysis. The titra- raster maps are presented in Table 1 and Appendix tion/wet oxidation method was used to determine the Figure A1, respectively. SOC content in percent (Walkley & Black, 1934), and the soil BD was estimated by the core method (Blake & 2.3.2.1. Topography factor. DEM-based terrain Hartge, 1986). parameters are one of the most commonly used factors Soil organic carbon stock (SOCS) was determined for SOCS prediction (Adhikari et al., 2014; using equation 1 (Y. -H. Lin et al., 2021; Huang et al., Lamichhane et al., 2019; Malone et al., 2009; 2019; Tessema & Kibebew, 2019): McBratney et al., 2003; Wang et al., 2018; Yang et al., 2016). All terrain parameters in this study were SOCS ¼ d� BD�ð1 CF=100Þ� SOCð%Þ (1) derived from the DEM based on Shuttle Radar −1 where SOCS represents SOC stock (t ha ), d is the Topography Mission (STRM) (Farr et al., 2007), thickness in centimeters of the soil layer, BD is the downloaded using EarthExplorer (EE) (http://earthex −3 bulk density (g cm ) of the soil layer, SOC (%) is SOC plorer.usgs.gov) from the archives of the U.S. content in percent, and CF is coarse fragment Geological Survey (USGS). The DEM was processed in percent. using ArcGIS 10.5 (ESRI, 2012). Finally, the terrain The coarse fraction was determined using equa- parameters were derived from the DEM using geopro- tion (2): cessing tools in the SAGA GIS terrain analysis toolbox � � (Conrad et al., 2015). In addition to DEM (elevation), wt wf CF ¼ � 100 (2) nine terrain attributes (aspects (ASP), plane curvature wt (PLCUR), profile curvature (PRCUR), convergence where w represents the total weight of the soil samples index (CI), midslope position (MSP), relative slope and w is the weight after the course fragment was position (RSP), valley depth (VD), channel network separated. distance (CND), and channel network base level (CNBL) were derived and used in addition to DEM ðmsðgÞ BD g=cm ¼ (3) (elevation) for SOCS prediction. Each of the terrain Vðcm Þ attributes, including the method of derivation, was where ms is the mass of oven-dry soil and v is the total described in Zevenbergen and Thorne (1987); volume of the core sampler, calculated as: Wilson et al. (2000); Böhner and Selige (2006); and Horst-Heinen et al. (2021). v ¼ Ah (4) where A represents the area of the core sampler, and 2.3.2.2. Landsat 8 based NDVI. Remotely sensed h is the height. vegetation parameters like the NDVI (Rouse et al., 1974) are good indicators of primary ecological pro- 2.3.2. Preparing environmental covariate ductivity, and they can be used for SOC prediction (predictors) (Minasny et al., 2013; Wang et al., 2018). The NDVI SOC storage or losses are affected by the interaction of was accessed upon request from Landsat Spectral several environmental covariates or predictors. Based Indices products of the U.S. Geological Survey Earth on a literature review (Adhikari et al., 2014; D. Chen Resources Observation and Science (EROS) Center et al., 2019; L. Deng et al., 2016; Lamichhane et al., (Vermote et al., 2016). A mean of 8 years (2013 to Table 1. Covariates preparation. Scorpan factor Covriate Abbreviaton unit Data type Reference(sourse) Terrain parameter r DEM/elevatin ELEV m Continuous earthexplorer.usgs.gov Climate data c Average annual temperature TEMP C Continuous Tufa et al. (2018) Average annual rainfall data RF mm Continuous Land use land cover o LACOV - Categorical www.mapserver-ethiopia vegetation indices o NDVI Continuous Rouse et al. (1974) Soil particle size fraction s Clay and sand content (%) CL and SND Continuous Hengl et al. (2015) respctively Soil type s SOT Categorical www.mapserver-ethiopia Predicted SOC s SOC t/ha Continuous Hengl et al. (2015) P GEOLOGY, ECOLOGY, AND LANDSCAPES 5 2020) of NDVI was used as one of the environmental 2019). Downscaling or upscaling (aggregating) raster covariates to predict SOC stock similar to layers to the target resolution (30 m) for preparing (Dharumarajan et al., 2017). Processing of NDVI was a stack, filtering out missing pixels, overlaying, and conducted using ArcGIS 10.5 (ESRI, 2012). sub-setting raster stacks and points were major activ- ities performed. Then, by overlaying point datasets 2.3.2.3. Climate. Among the climatic parameters, (SOCS sample points) and raster stacks (covariates), temperature and precipitation are the key drivers of a regression matrix was prepared for constructing SOC storage, affecting both carbon input into the soil a model. An open-source GIS software SAGA GIS and its decomposition (Wiesmeier et al., 2019). (Conrad et al., 2015) and R packages in R (R Core Gridded rainfall and temperature data (4-km resolu- Team, 2020), were used for processing the covariates. tion), which were based on ENACTS program (Tufa et al., 2018), were obtained from the Ethiopian 2.4. Model fitting National Meteorological Agency (ENMA) and used to represent climate covariates. The data were pro- The “randomForest” package (version 4.6–14) (Liaw duced by combining station data with satellite rainfall & Wiener, 2002) in R (3.6.3) was used to establish the and reanalysis temperature data products (Tufa et al., relationship between SOCS and the predictor vari- 2018). Then, the average annual near current (1990– ables. RF is a machine learning technique that is 2019) temperature and rainfall data were processed based on the classification and regression tree and used for SOC stock prediction. (Breiman, 2001). In RF, several trees are grown using random samples of predictands and predictors, and 2.3.2.4. Land use land cover and soil type. A review then these tree models are aggregated into on the effects of LULC changes on SOC showed that a comprehensive classifier or repressor (Breiman, land use conversions have significantly reduced soil 2001). The RF algorithm depends only on three user- C stocks globally (L. Deng et al., 2016). An LULC map defined parameters (tuning parameters), which deter- of 20-m resolution, which was available at http://www. mine the goodness of the model fitting: the number of mapserver-ethiopia, was accessed from MapServer trees (ntree) in the forest, the minimum number of Ethiopia. The mapping services are based on data data points in each terminal node (nd), and a random from WLRC and open-source sources like NASA subset of prediction factors (mtry) that are randomly and are developed by experts in Addis Ababa, selected on each tree to determine the split at each Ethiopia. The LULC types were regrouped according node (Liaw & Wiener, 2002). The default values of to the IPCC’s land use categories (Ipcc(intergevern- ntree, nodesize, and mtry are 500, 5, and one-third of mental Panel on Climate Change) et al., 2006) as the total number of predictors, respectively, for regres- defined in IPCC (2003). According to a study con- sion problems (Liaw & Wiener, 2002). ducted in East Africa, soil inherent properties, such as A study showed that tuning ntree and nodesize did soil texture and soil type, have a strong influence on not reduce out-of-bag RMSE significantly (Nussbaum SOC (Vågen & Winowiecki, 2013; Wiesmeier et al., et al., 2017). However, mtry must be optimized since it 2019). A 20-m resolution soil-type map, which was can influence the predictive performance of the RF based on the FAO-UNESCO Soil Classification model (Breiman & Cutler, 2004; Hastie et al., 2009; System (Driessen et al., 2001), was also accessed from Lu et al., 2019). In this study, therefore, the default MapServer Ethiopia. values of ntree and nodesize were used, and mtry was tuned using the command tuneRF found within the 2.3.2.5. Predicted soil organic carbon and soil parti- randomForest package in R. During the generation cle size fraction. According to Hengl et al. (2017), soil process of RF, each tree is grown using approximately predictions can be used as covariate layers for down- two-thirds of the training (bootstrap) data, and one- scaling and producing higher-resolution local predic- third (called out-of-bag (OOB) data) is used for inter- tions of soil properties. A revised prediction map of nal validation (Grimm et al., 2008; Hastie et al., 2009; SOC stock for Africa at 250 m resolution (Hengl et al., Liaw & Wiener, 2002). Then, mtry was then tuned 2015), which is available for download via http://www. depending on the smallest OOB error, and the mtry isric.org/data/AfSoilGrids250m was used for predict- associated with the smallest OOB error was used to fit ing SOC stock. Predicted maps of soil clay and sand the RF model. content (as proxies for fine and coarse texture frac- tions, respectively) were also accessed from the same 2.5. Model evaluation site. According to Liaw and Wiener (2002), the RF model 2.3.3. Preparing regression matrix has an internal validation mechanism using the OOB Covariate preparation for the SOC stock prediction error (as it has also been presented in Section 2.4) and includes different activities (Hengl & Macmillan, the OOB mean of squared residuals. R , which is 6 M. K. NEGASSA ET AL. a statistical measure of how well an OOB prediction N is total number of samples approximates the real data points, was calculated auto- 2ρσ σ o p CCC ¼ (8) matically. Data splitting is the most frequently used 2 2 σ þ σ þðμ þ μ Þ o p o p validation approach used to evaluate map accuracy in DSM (S. Chen et al., 2022). K-fold cross-validation (in CCC is concordance correlation coefficient which k is 5 or 10) is the most commonly used model ρ is Pearson’s correlation coefficient between validation technique in machine learning (John et al., observed and predicted values 2020; Liu et al., 2022; Rodriguez et al., 2009). In this 2 2 σ and σ are observed and predicted variance, o p study, the RF model’s performance in predicting SOC respectively stock was evaluated using 20% of the measured SOC μ and μ are observed and predicted mean, o p stock point data, based on 10-fold cross-validation. respectively. The model evaluation was conducted using the four All the calculations were done using Microsoft most common performance indicators used in model Excel and R software. evaluation. They were as follows: coefficient of deter- mination (R ), root mean square error (RMSE), mean 2.5.1. Descriptive statistics and correlation analysis error (ME), and Lin’s concordance correlation coeffi- Descriptive statistics like mean, standard deviation, cient (CCC) (S. Chen et al., 2022). R is the percentage and coefficient of variation (CV) were used to sum- of explained variance of the response variable, ME marize the numeric variables used in this study. The measures average bias in prediction, and Lin’s concor- CV values were categorized into three: CV < 15), CV dance correlation coefficient (CCC) measures the level = 15–35%, and CV > 35%, indicating least, moderate, of agreement between predicted and observed values, and high variability, respectively (Obi et al., 2010). which again measures the distance to the 1:1 line A correlation analysis was conducted to reveal the (Dharumarajan et al., 2017; Hengl et al., 2018; relationship between continuous predictor variables L. I. Lin, 1989). The first three model evaluation and SOCS using Pearson correlation (Mukaka, 2012). indices (R , RMSE, and ME) were calculated using All the statistics were calculated using Microsoft Excel equations 2, 3, and 4, respectively. The CCC was software. calculated using Equation 5 (Hengl et al., 2018). Similar indices were used in similar studies (Adhikari et al., 2014; Dharumarajan et al., 2017). 3. Results ðb z zÞ^2 i 3.1. Description statistics of SOCS 2 i¼1 R ¼ (5) ðz zÞ^2 i¼1 Descriptive statistics of the observed (measured) SOCS for the 0–30 cm soil depth are presented in rffiffiffi 1 n Table 2. The measured SOCS in the study area ranged RMSE ¼ ðz ^ zÞ (6) i i i¼1 between 7.97 t ha-1 and 79.49 t ha-1 with a CV of 47.40%, showing high variability. The highest mean SOCS was recorded for forestland, followed by crop- ME ¼ 1=n ðz zÞ (7) i i i¼1 land. The CV values of the SOCS in cropland and grassland showed high variability of SOCS within where R is the coefficient of determination each LULC. RMSE is root mean square error, n is the number of samples ME is mean error z corresponds to the measured/observed SOCS of the depth i, σ 3.1.1. Pearson correlation � z is the mean of observed SOCS and The Pearson correlation analysis result for numeric ^ z is the predicted value predictors showed that all the covariates have n is 0.8*N and 0.2*N for training and validation a significant linear correlation (at p < 0.05) with SOC datasets, respectively stock except the profile curvature (PRCUR), plane Table 2. Predictand description. SOCS (t/ha) of the land use/land cover Descriptive statistics Entire study area Cropland Rangeland Forest n 100 38 10 148 Mean 41.54 36.67 52.56 41.04 Median 36.04 28.46 51.86 36.04 Min 7.97 11.49 23.41 7.97 Max 79.49 76.86 78.88 79.49 STDV 19.02 19.94 18.31 19.45 CV% 45.78 54.37 34.84 47.40 GEOLOGY, ECOLOGY, AND LANDSCAPES 7 curvature(PLCUR), midslope position (MSP), clay the scattered plots of predicted versus measured content (CL), and curvature index (CI). Out of the SOCS for training and validation datasets (with a line significant covariates, temperature (TEMP), and valley where predicted and measured values are equal or depth (VD) have significant negative correlations with a 1:1 line added) were presented in Figure 3 a and b, p = −0.555 (p < 0.001) and p = −0.213 (p < 0.01), respectively. The CCC quantifies how far the mea- respectively. NDVI (p = 0.525), prediction map of sured data deviated from the 1:1 line (Hengl et al., SOCS for Africa (SOPp) (p = 0.469), elevation (Elev) 2018) (Figure 3). (p = 0.424), channel network distance (CND) (p = 0.288), rainfall (RF) (p = 0.401), relative slope position 3.2.2. Spatial prediction of soil organic carbon (RSP) (p = 0.316), and channel network base level stock in study area (CNBL) (p = 0.422) showed positive significant corre- A prediction map of SOCS at 0–30 cm depth was lation with SOCS at (p < 0.001). Temperature has the produced at a resolution of 30 m × 30 m (Figure 4). highest correlation coefficient, followed by NDVI, On the map, permanent bodies of water (wetlands), SOPp, and elevation, and VD has the least correlation bare rock (otherlands), and settlements were excluded. coefficient (p = 0.213), significant at p < 0.05. In this study, a LULC map from MapServer Ethiopia, which is available at www.mapserver-Ethiopia was used to define soil masks (areas of interest for the prediction of SOCS): forestland, grassland, and crop- 3.2. Digital mapping of soil organic carbon stock land. The predicted SOC stock values were in the 3.2.1. Model performance range of 17.91 to 70.16 t/ha, with a mean of 38.82 t/ At the initial stage, the model’s internal OOB evalua- ha, a standard deviation of 10.11 t/ha, and a CV of 26% tion was used for parameter tuning. For the best pre- in the whole area. The overall SOC stock within 0–30 diction accuracy, the best mtry was selected in terms of cm of depth was 240.5 Tg. the lowest OOB error. The result showed that OBB error was lowest at the default mtry (mtry = 6). Then, 3.2.3. Importance of factors affecting SOC stock the model was fitted using the default ntree, nodesize, The variable importance (Gini-based importance) and the tuned mtry. The OOB prediction mean square result (Figure 5) showed that there was a variation in error and the OOB variance explained (R OOB) were the importance of covariates in affecting the spatial 225.62% and 40%, respectively. The model perfor- distribution of SOCS. Among the covariates, NDVI, mance evaluation was conducted using 10-fold cross- TEMP, and elevation (ELV) were the most important validation (Table 3). The data were randomly divided predictors. VD and soil type (SOT) were grouped in into training and validation data sets (80% and 20% the second rank, followed by the derived terrain attri- respectively) using sample function in R. The result bute CNBL. This was again followed by sand content showed (Table 3) that R was 53% and 36% for train- (SND) and rainfall (RF). The next important variables ing and validation datasets, respectively. The RMSE were the derived terrain attributes: channel network and Lin’s CCC were also presented for training and distance (CND), plane curvature (PLCUR), relative validation data sets. In addition, slope position (RSP), profile curvature (PLCUR), Table 3. Random forest model calibration and validation results by 10-fold cross-validation. Indices Sample number(n) R (%) RMSE(t C) ME Lins CCC Training dataset 118 53 6.76 0.03 0.92 Validation dataset 30 36 9.52 0.36 0.82 Figure 3. Graph of predicted vs measured for training (a) and test set (b) datasets. 8 M. K. NEGASSA ET AL. Figure 4. Distribution of predicted SOC stock in the study area. Figure 5. IncNodepurity-based variable importance of random forest model for soil organic carbon stock predictions (abbrevia- tions are presented in section 2.3.2.1 and Table 1). GEOLOGY, ECOLOGY, AND LANDSCAPES 9 curvature index (CI), aspect (ASP), convergence index 4. Discussion (CI), and mid-slope position (MSP), having a medium 4.1. Predictand description influence on SOC stock distribution. Finally, predic- tion maps of SOC stock for Africa at 250 m resolution The observed SOC stock was higher in cropland than (SOCp) and soil clay (CL) content are still important, in grassland. This may be due to the existing land but they have less influence on SOC stock. The land management practices in cropland. Agricultural man- cover (LACOV) covariate was at the bottom of the agement practices such as erosion control and the variable importance ranking and showed little influ- addition of organic amendments may increase carbon ence on SOC stock prediction. levels in soil beyond the historic LULC (Ontl & Schulte, 2012). Adding manure to cropland and prac- ticing crop rotation have been commonly practiced in 3.2.4. Predicted soil organic carbon stock by soil the midland and highland areas of the study area. and land cover types A study showed that SOC sequestration was signifi- Predicted SOC stocks for LULC classes (IPCC-based) cantly increased by adding manure (Xia et al., 2017). and soil groups (based on MapServer Ethiopia) were In addition, the management of degraded agricultural calculated using zonal statistics of the spatial analyst lands has great potential to contribute to increased toolbox in ArcMap (Table 4). The predicted SOC stock for the LULC class showed that cropland con- SOC sequestration (DeLonge et al., 2013). In the tained the highest mean SOC stock (41.32 t/ha), fol- study area, soil and water conservation practices, lowed by forestland (40.33 t/ha) and grassland (36.19 such as soil and stone bund construction across the t/ha). Ninety-seven percent (97%) of the total SOC slope, were observed mostly in the highland and mid- stock was stored on agricultural land (cropland and land area of the study area. grassland). The mean SOC stock for dominant soil In addition, lower SOC stocks in grassland could be groups ranged between 29.92 t/ha (Fluvisols) and due to land degradation. It was observed that in most 37.69 t/ha (Calcisol), which cover smaller areas of of the grassland, the soil had a very shallow depth due 2.99% and 1.02%, respectively. to soil erosion by water. However, studies have shown Vertisols, like calcisols, have a large (46.76 t/ha) SOC that improving carbon sequestration in grasslands is stock. The most dominant soil types in the study area a promising and cost-effective strategy to mitigate are leptosols and cambisols (mainly found in Mieso), climate change (Lal, 2003; Tessema & Kibebew, which contained 38% and 20% of the total SOC stock, 2019). Tessema and Kibebew (2019) showed that respectively. The distribution of the mean predicted grassland management practices, such as grassland SOC stock across the districts showed a maximum regeneration and restoration of degraded land store value in Gemechis (50.76 t/ha), followed by Kuni. SOC stock, which varies between 0.1 and 93 Mg C ha- Although the mean predicted SOC stock in Mieso was 1. Grassland covers the largest area of the dry area in the lowest (30.31 t/ha) (Table 4), due to covering the the study area (as in the case of Mieso district). largest area, it contained the highest total SOC stock Therefore, if it is managed, it will be a promising (38% of the stock in the whole study area) (Figure 6). area for climate change mitigation. Table 4. Soil organic carbon stocks by land cover types and soils. Categories a b Mean (t/ha) Area (%) Total (Tg) LULC LULC tree cover areas Forestland 40.33 2.37 6.01 shrub cover and grassland Grassland 36.19 48.22 107.74 cropland Cropland 41.32 49.41 126.11 Total 100 239.86 Soil groups Vertisols(VR) 46.76 12.01 31.74 Calcisol (CL) 51.27 1.02 2.94 Cambisols(CM) 31.16 27.24 48.04 Regosol (RG) 28.37 6.63 10.65 Leptosol (LP) 42.21 37.92 90.64 Fluvisols (FL) 29.92 2.99 5.04 Luvisols (LV) 45.06 11.16 28.38 Nitosols (NT) 34.45 0.71 1.37 N/A 43.41 0.33 0.82 Total 100 219.62 Districts Chiro Zuria 40.34 11.56 28.89 Gemechis 50.76 17.42 54.80 Kuni 45.63 24.73 69.89 Mieso 30.31 46.29 86.91 Total 100 240.50 land cover classes from MapServer Ethiopia Land cover regrouped according to IPCC definition 10 M. K. NEGASSA ET AL. Figure 6. Predicted SOC stock stratified by districts (Percentage values represent the fraction of the total SOC stock (240.50 Tg)). 4.2. Digital soil organic carbon stock mapping of SOC spatial distribution in topsoil due to the varia- tions in small-scale soil management practices. The 4.2.1. Model performance other reasons that affected the performance of the The mtry tuning to improve prediction results model could be sampling density, error in covariates revealed that the default mtry performed best. and the target variable, and prediction scale (Adhikari A similar result was reported by Grimm et al. (2008). et al., 2014; Hengl & Macmillan, 2019; Mondal et al., The performance of RF model fitting using soil sam- 2017). For those reasons, the amount of variance ples and environmental covariates was evaluated using explained by soil spatial regression models may be four evaluation indices and the scatter plot of mea- less than the residual variance (Hengl & Macmillan, sured vs. predicted SOC stock. The model perfor- 2019). mance was better for the calibration data set than the validation dataset (Table 3). The positive value of ME 4.2.2. Variable importance showed that the model overestimated SOC stock. The The variable importance tests of the environmental scatter plot showed that the predicted points were near covariates used in this study showed each had the 1:1 line. This showed that the predicted values a varying level of influence on the SOC stock distribu- were near the measured ones. tion. Among those covariates, normalized difference The model performance validation datasets with vegetation index (NDVI), temperature (TEMP), and respect to the explained variance (36% of the total elevation (ELV) had the largest influence on the spatial variance) was comparable to several DSM studies of distribution of SOC stock. Several studies have stated similar soil depth (Dharumarajan et al. (2017) (R = that NDVI is a strong predictor of SOC stock (Lei 20%); Adhikari et al. (2014) (R = 0.41); Adhikari et al. et al., 2019; Mondal et al., 2017; Wiesmeier et al., 2 2 (2019) (R = 38%); Gomes et al. (2019) (R = 0.23 to 2019). The influence of NDVI is through the input of 0.32 for 0–30 cm depth of soil); Huang et al. (2019) organic matter into the soil, and that of temperature is 2 2 (R = 48%); Bui et al. (2009) (R = 49%); and Hengl through the microbial decomposition of SOM 2 2 et al. (2018) (R = 41%)). It can be observed that R (Conant et al., 2011). It was also shown that among values of less than 0.5 are common in DSM studies. terrain attributes, elevation is the primary terrain attri- The reason for the lower performance of soil property bute that best relates to soil properties (John et al., prediction in DSM may be due to the high variability 2021). Elevation affects plant productivity indirectly GEOLOGY, ECOLOGY, AND LANDSCAPES 11 by affecting temperature and rainfall. It affects the may lower SOC stock. For the same reasons as mea- microclimate, which affects plant growth. sured SOC stock, predicted SOC stock was lower in Other important covariates, such as channel net- grassland than in cropland. Agricultural land (crop- work distance (CND) and channel network base- land and grassland) accounts for 98% of the total land level (CNBL), were linked to soil re-distribution area. Therefore, management practices that improve through erosion and deposition that affects vegeta- soil carbon could have a significant effect on carbon tion as well as SOC decomposition (Piri Sahragard sequestration for climate change mitigation, which & Pahlavan Rad, 2020). Land cover (LACOV) had can significantly contribute to achieving the Intended almost no influence on SOC stock. The reason for Nationally Determined Contributions (INDC) of the that may be that it combined different soil proper- country. According to the IPCC estimates, soil carbon ties, like terrain parameters, and climate variables, sequestration can contribute to 89% of the total tech- which might have covered its influence (Huang nical mitigation potential for agriculture globally et al., 2019). A study has also found that the effects (Smith et al., 2007). This potential is particularly useful of land use on SOC storage may be affected by soil in areas where large areas of land are under agricul- type (Mayes et al., 2014). In addition, soil proper- tural use, as in the case of the study area. ties those vary greatly depending on elevation play In the case of soil type, there was a substantial differ- crucial roles in determining SOC (Arunrat et al., ence in the amount of SOC stock stored as well as the 2020). Another reason could be that existing man- area of land covered. Vertisols, which are very important agement practices rather than LULC might have soils in Ethiopian agriculture, have a large SOC stock a larger influence on the SOC stock distribution. next to calcisols. This may be due to the high clay content in vertisols (Eyasu, 2016), which restricts the movement of air and water that are important for the microbial 4.2.3. Soil organic carbon spatial prediction decomposition of organic matter. Several studies have The spatial distribution of SOC stock in the prediction also demonstrated the strong relationship between silt map (Figure 4) showed that the SOC stock density was and clay content and SOC due to the formation of highest in Gemechis, southern Chiro, and the north- stabilized clay-humic complexes against microbial western part of Kuni. These areas are located in decomposition of SOM (Gonçalves et al., 2017; Six greener, higher-elevation, and cooler climate areas. et al., 2002). This helps long-term carbon storage in the The variable importance plot also showed that soil and contributes a lot to climate change mitigation. NDVI, temperature, and elevation were in the first The reason for the lower SOC stock in Cambisol rank of all the covariates influencing the SOC stock could be due to its availability in dry areas where the distribution. This was again revealed by the Pearson biological oxidation of OM is limited by low soil moist- correlation test, showing the strongest relationship ure. Although Cambisol has a smaller SOC stock, due between those covariates and SOC stock (sec- to its largest area coverage, it could have a considerable tion 3.2.2). role in storing a large amount of SOC stock. Cambisol The lowest predicted SOC stock density values were was dominant in Mieso district, where moisture is dominantly found in Mieso. This might be due to the a limiting factor for agricultural productivity. Eyasu location of the district; because the entire area of et al. (2016) also stated that one of the conditions for Mieso is located in a moisture-limited area that limits the development of Cambisol is low moisture. SOM input. Although Mieso had a relatively lower Application of appropriate management practices that SOC stock per hectare, the largest total predicted conserve limited soil moisture and improve SOM (for SOC stock was observed there (Figure 6). The areas example, by using a conservation agriculture approach) with the highest density of predicted SOC stock could could improve the organic matter content of Cambisol. be potential sources of carbon emissions, while the This not only increases agricultural productivity but areas with the lowest density could be potential also helps to mitigate climate change. sinks. Lower SOC stock density means higher seques- tration potential through appropriate land manage- ment. Moreover, the lower decomposition rate of 5. Summary and conclusions SOM in dry land soils (as in the case of Mieso district) would be an advantage for long-term SOC storage and The Random Forest (RF) method was applied to fit climate change mitigation. a model to SOC stock and environmental covariates. Variable tuning proved the default mtry to be the best 4.2.4. Predicted soil organic carbon by land cover choice to improve prediction accuracy. The prediction and soil type of SOC stock using the developed model allowed the Predicted SOC stock by LULC showed that cropland production of a SOC stock map. This aided in reveal- had the highest stock, followed by forestland. Similar ing small-scale spatial variations in SOC stock across result was reported by Adhikari et al. (2014). This may the area. The map showed the variability in soil units, be due to the inclusion of bare land in the forest, which unlike traditional approaches in which the SOC stock 12 M. K. NEGASSA ET AL. mean is spatially linked to soil units. The findings denmark. PLoS One, 9(8), e105519. https://doi.org/10. 1371/journal.pone.0105519 showed that SOC storage levels in any soil are affected Adhikari, K., Owens, P. R., Libohova, Z., Miller, D. M., by the combined effect of the predictor factors, and Wills, S. A., & Nemecek, J. (2019). Assessing soil organic each factor has a different level of influence. carbon stock of Wisconsin, USA and its fate under future Information on the relationship between the SOC land use and climate change. The Science of the Total stock and the covariates was drawn from both variable Environment, 667, 833–845. https://doi.org/10.1016/j.sci importance measures, the map of the predicted SOC totenv.2019.02.420 Arunrat, N., Pumijumnong, N., Sereenonchai, S., & stock and the Pearson correlation test. Chareonwong, U. (2020). Factors controlling soil organic The findings of the study showed that the predicted carbon sequestration of highland agricultural areas in the SOC stock values were in the range of 17.91 to 70.16 t/ mae chaem basin, northern thailand. Agronomy, 10(2), ha, with a mean of 38.82 t/ha. The overall SOC stock 305. https://doi.org/10.3390/agronomy10020305 within the 0–30 cm soil depth was 240.50 Tg. The area Bangelesa, F., Adam, E., Knight, J., Dhau, I., Ramudzuli, M., & Mokotjomela, T. M. (2020). Predicting soil organic with the highest NDVI, highest elevation, and lowest carbon content using hyperspectral remote sensing in temperature has the highest SOC stock. The most a degraded mountain landscape in lesotho. Applied and important covariates that influenced SOC stock in Environmental Soil Science, 2020, 1–11. https://doi.org/ the study area were NDVI, temperature, and elevation. 10.1155/2020/2158573 Correlation analysis again revealed NDVI, tempera- Bian, Z., Guo, X., Wang, S., Zhuang, Q., Jin, X., Wang, Q., & ture, and elevation to have had a strong relationship Jia, S. (2019). Applying statistical methods to map soil organic carbon of agricultural lands in northeastern with SOC stock. coastal areas of China. Archives of Agronomy and Soil The study could help provide information on Science, 66(4), 532–544. https://doi.org/10.1080/ improving soil mapping at a lower scale using 03650340.2019.1626983 a cheaper method. The estimation of SOC stock Birhanu, I., & Chalchisa, T. (2019). Digital soil mapping for based on the available data can be used as a baseline site-specific management of soils. Geoderma, 351, 85–91. https://doi.org/10.1016/j.geoderma.2019.05.026 SOC stock. The method could support national car- Blake, G. R., & Hartge, K. H. (1986). Bulk density 1. methods bon monitoring by updating the prediction using soil anal. part 1—physical mineral. Methods, 9, 363–375. environmental covariates without the need to measure Böhner, J., & Selige, T. (2006). Spatial prediction of soil SOC stock. Finally, the study identified locations of attributes using terrain analysis and climate regionalisa- potential sources and sinks of carbon so that informed tion. In Gottinger G. (Ed.), SAGA-analyses and modelling location-specific decisions can be made on SOC man- applications (pp. 13-120). Göttingen. Breiman, L. (2001). Random forests. Machine Learning, 45 agement activities that reduce loss of carbon and (1), 5–32. https://doi.org/10.1023/A:1010933404324 improve carbon input, respectively, to ensure climate- Breiman, L., & Cutler, A., 2004. Random Forest — manual, smart agriculture. on-line: http://www.stat.berkeley.edu/~breiman/ RandomForests/cc_manual.htm . Bui, E. N., Henderson, B. L., & Viergever, K. (2009). Using Acknowledgments knowledge discovery with data mining from the Australian Soil Resource Information System database We are grateful to the African Center of Excellence for to inform soil carbon mapping in Australia. Global bio- Climate Smart Agriculture and Biodiversity Conservation geochemical cycles, 23(4), 12–15. https://doi.org/10.1029/ (ACE Climate SABC) at Haramaya University for their 2009GB003506 academic and financial support in conducting this research. Cerretelli, S., Poggio, L., Yakob, G., Boke, S., Habte, M., Our gratitude goes to the reviewers of the original manu- Coull, M., Gimona, A. . . . Gimona, A. (2021). The advan- script, without whose help this would not have looked as it tages and limitations of global datasets to assess carbon does. stocks as proxy for land degradation in an Ethiopian case study. Geoderma, 399, 115117. https://doi.org/10.1016/j. geoderma.2021.115117 Chen, S., Arrouays, D., Mulder, V., Poggio, L., Minasny, B., Disclosure statement Roudier, P., Libohova, Z., Lagacherie, P., Shi, Z., No potential conflict of interest was reported by the Hannam, J., Meersmans, J., Richer de Forges, A. C., & author(s). Walter, C. (2022). Digital mapping of GlobalSoilmap soil properties at a broad scale: A review. Geoderma, 409, 115567. https://doi.org/10.1016/j.geoderma.2021.115567 Chen, D., Chang, N., Xiao, J., Zhou, Q., & Wu, W. (2019). Funding Mapping dynamics of soil organic matter in croplands with MODIS data and machine learning algorithms. The The work was supported by the The World Bank . Science of the Total Environment, 669, 844–855. https:// doi.org/10.1016/j.scitotenv.2019.03.151 Conant, R. T., Ryan, M. G., Ågren, G. I., Birge, H. E., References Davidson, E. A., Eliasson, P. E., Bradford, M. A. . . . Adhikari, K., Hartemink, A. E., Minasny, B., Bou Kheir, R., Bradford, M. A. (2011). Temperature and soil organic mat- Greve, M. B., Greve, M. H., & Hui, D. (2014). Digital ter decomposition rates - synthesis of current knowledge mapping of soil organic carbon contents and stocks in and a way forward. Global Change Biology, 17(11), GEOLOGY, ECOLOGY, AND LANDSCAPES 13 3392–3404. https://doi.org/10.1111/j.1365-2486.2011. Forests analysis. Geoderma, 146(1–2), 102–113. https:// 02496.x doi.org/10.1016/j.geoderma.2008.05.008 Conrad, O., Bechtel, B., Bock, M., Dietrich, H., Fischer, E., Hastie, T. J., Tibshirani, R. J., & Friedman, J. J. H. Gerlitz, L., Böhner, J. . . . Böhner, J. (2015). System for (2009). The elements of statistical learning: Data Automated Geoscientific Analyses (SAGA) v. 2.1.4. mining, inference, and prediction. In T. Hastie, R. Geoscientific Model Development, 8(7), 1991–2007. Tibshirani, J. Friedman (Eds.), Springer series in sta- https://doi.org/10.5194/gmd-8-1991-2015 tistics. Springer-Verlag New York.And Prediction. Cowie, A. L., Penman, T. D., Gorissen, L., Winslow, M. D., Springer series in statistics (Vol. 2009, pp. 764). Lehmann, J., Tyrrell, T. D., Akhtar-Schuster, M. . . . Springer-Verlag New York. Akhtar-Schuster, M. (2011). Towards sustainable land Hengl, T., Heuvelink, G. B. M., Kempen, B., management in the drylands: Scientific connections in Leenaars, J. G. B., Walsh, M. G., Shepherd, K. D., monitoring and assessing dryland degradation, climate Tondoh, J. E. . . . Tondoh, J. E. (2015). Mapping soil change and biodiversity. Land Degradation & properties of Africa at 250 m resolution: Random forests Development, 22(2), 248–260. https://doi.org/10.1002/ significantly improve current predictions. PLoS One, 10(6), ldr.1086 e0125814. https://doi.org/10.1371/journal.pone.0125814 Delonge, M. S., Ryals, R., & Silver, W. L. (2013). A Life Cycle Hengl, T., & Macmillan, R. A. (2019). Predictive soil map- Model to Evaluate Carbon Sequestration Potential and ping with R. OpenGeoHub foundation. Greenhouse Gas Dynamics of Managed Grasslands. Hengl, T., Mendes de Jesus, J., Heuvelink, J. B. M., Ruiperez Ecosystems (New York, N.Y.), 16(6), 962–979. https:// Gonzalez, M., Kilibarda, M., Blagotic, A., & Kemepen, B. doi.org/10.1007/s10021-013-9660-5 (2017). SoilGrids250m: Global gridded soil information Deng, L., Zhu, G., Tang, Z., & Shangguan, Z. (2016). Global based on machine learning. Plos One, 12(2), e0169748. patterns of the effects of land-use changes on soil carbon https://doi.org/10.1371/journal.pone stocks. Global Ecology and Conservation, 5, 127–138. Hengl, T., Nussbaum, M., Wright, M. N., https://doi.org/10.1016/j.gecco.2015.12.004 Heuvelink, G. B. M., & Gräler, B. (2018). Random forest Dharumarajan, S., Hegde, R., & Singh, S. K. (2017). Spatial as a generic framework for predictive modeling of spatial prediction of major soil properties using Random Forest and spatio-temporal variables. PeerJ, 6, e5518. https://doi. techniques - a case study in semi-arid tropics of South org/10.7717/peerj.5518 India. Geoderma Regional, 10, 154–162. https://doi.org/ Hobley, E., & Willgoose, G. (2010). Measuring soil organic 10.1016/j.geodrs.2017.07.005 carbon stocks – issues and considerations. In 19th world Driessen, P. M., Deckers, J., & Spaargaren, O. (2001). congress of soil science (pp. 62-65). Soil Solutions for Lecture notes on the major soils of the world. In (World a Changing World. Soil Resources Reports : FAO (Vol. 94). Food and Horst-Heinen, T. Z., Dalmolin, R. S. D., ten Caten, A., Agriculture Organization of the United Nations (FAO). Moura-Bueno, J. M., Grunwald, S., de a Pedron, F., da ftp://ftp.fao.org/agl/agll/docs/wsrr94e.pdf Silva-Sangoi, D. V. . . . da Silva-Sangoi, D. V. (2021). Soil Elias, E. (2016). Soils of the Ethiopian Highlands: depth prediction by digital soil mapping and its impact in Geomorpholog and Properties (p. 385pp). CASCADE pine forestry productivity in South Brazil. Forest Ecology Project, ALTERA, Wageningen University and Research and Management, 488, 118983. https://doi.org/10.1016/j. Center (Wageningen UR). The Netherlands. foreco.2021.118983 Esri. (2012). What is GIS? Esri. Huang, J., Hartemink, A. E., & Zhang, Y. (2019). Climate FAO(Food and Agriculture Organization of United States). and land-use change effects on soil carbon stocks over (2017). SOC - the hidden potential.Food and agriculture 150 years in Wisconsin, USA. Remote Sensing, 11(12), organization of the united nations. FAO. 1504. https://doi.org/10.3390/rs11121504 Farr, T. G., Rosen, P. A., Caro, E., Crippen, R., Duren, R., IPCC(Intergevernmental Panel on Climate Change). (2003). Hensley, S., Alsdorf, D. . . . Alsdorf, D. (2007). The shuttle Good practice guidance for land use, land-use change and radar topography mission. Reviews of Geophysics, 45(2). forestry.IPCC national greenhouse gas inventories pro- https://doi.org/10.1029/2005rg000183 gramme, comptes rendus - biologies. Institute for Global Gao, J., Pan, G., Jiang, X., Pan, J., & Zhuang, D. (2008). Environmental Strategies (IGES) for the IPCC. Land-use induced changes in topsoil organic carbon IPCC(Intergevernmental Panel on Climate Change). 2006. stock of paddy fields using MODIS and TM/ETM analy- In Eggleston, H. S., Miwa, K., Srivastava, N. and sis: A case study of Wujiang County, China. Journal of Tanabe, K. I. (Eds.), IPCC guidelines for national green- Environmental Sciences, 20(7), 852–858. https://doi.org/ house inventories– aprimer, prepared by the national 10.1016/s1001-0742(08)62137-3 greenhouse gas inventories programme. Institute for Gomes, L. C., Faria, R. M., de Souza, E., Veloso, G. V., Global Environmental Strategies. Schaefer, C. E. G. R., & Filho, E. I. F. (2019). Modelling John, K., Abraham, I. I., Kebonye, N. M., Agyeman, P. C., and mapping soil organic carbon stocks in Brazil. Ayito, E. O., & Kudjo, A. S. (2021). Soil organic carbon Geoderma, 340, 337–350. https://doi.org/10.1016/j.geo prediction with terrain derivatives using geostatistics and derma.2019.01.007 sequential Gaussian simulation. Journal of the Saudi Gonçalves, D. R. P., de Moraes Sá, J. C., Mishra, U., Society of Agricultural Sciences, 20(6), 379–389. https:// Cerri, C. E. P., Ferreira, L. A., & Furlan, F. J. F. (2017). doi.org/10.1016/j.jssas.2021.04.005 Soil type and texture impacts on soil organic carbon John, K., Abraham, I. I., Michael, K. N., Okon, A. E., storage in a sub-tropical agro-ecosystem. Geoderma, Chapman, A. P., & Marcus, A. S. (2020). Using machine 286, 88–97. https://doi.org/10.1016/j.geoderma.2016.10. learning algorithms to estimate soil organic carbon varia- bility with environmental variables and soil nutrient indi- Grimm, R., Behrens, T., Märker, M., & Elsenbeer, H. (2008). cators in an alluvial soil. Land, 9(12), 487. https://doi.org/ Soil organic carbon concentrations and stocks on Barro 10.3390/land9120487 Colorado Island — Digital soil mapping using Random 14 M. K. NEGASSA ET AL. Lal, R. (2003). Soil erosion and the global carbon budget. Agronomy, 1–47. https://doi.org/10.1016/b978-0-12- 405942-9.00001-3 Environment International, 29(4), 437–450. https://doi. org/10.1016/s0160-4120(02)00192-7 Mondal, A., Khare, D., Kundu, S., Mondal, S., Mukherjee, S., Lal, R. (2004). Soil carbon sequestration to mitigate climate & Mukhopadhyay, A. (2017). Spatial soil organic carbon change. Geoderma, 123(1–2), 1–22. https://doi.org/10. (SOC) prediction by regression kriging using remote sensing data. The Egyptian Journal of Remote Sensing 1016/j.geoderma.2004.01.032 and Space Science, 20(1), 61–70. https://doi.org/10.1016/ Lal, R., Lorenz, K., Hüttl, R. F., Schneider, B. U., & Von Braun, J. (2013). Ecosystem services and carbon seques- j.ejrs.2016.06.004 Mukaka, M. M. (2012). Statistics Corner: A guide to appro- tration in the biosphere. Springer. https://doi.org/10. priate use of Correlation coefficient in medical research. 1007/978-94-007-6455-2 Malawi Medical Journa, 24(3), 69–71. Lamichhane, S., Kumar, L., & Wilson, B. (2019). Digital soil Nayak, A. K., Rahman, M. M., Naidu, R., Dhal, B., mapping algorithms and covariates for soil organic car- Swain, C. K., Nayak, A. D., Pathak, H. . . . Pathak, H. bon mapping and their implications: A review. (2019). Current and emerging methodologies for estimat- Geoderma, 352, 395–413. https://doi.org/10.1016/j.geo ing carbon sequestration in agricultural soils: A review. derma.2019.05.031 The Science of the Total Environment, 665, 890–912. Lei, Z., Yu, D., Zhou, F., Zhang, Y., Yu, D., Zhou, Y., & https://doi.org/10.1016/j.scitotenv.2019.02.125 Han, Y. (2019). Changes in soil organic carbon and its Nussbaum, M., Spiess, K., Baltensweiler, A., Grob, U., influencing factors in the growth of Pinus sylvestris var. Keller, A., Greiner, L., Schaepman, M. E., & Papritz, A. mongolica plantation in Horqin Sandy Land, Northeast (2017). Evaluation of digital soil mapping approaches with China. Scientific Reports, 9(1). https://doi.org/10.1038/ large sets of environmental covariates (pp. 1–32). SOIL s41598-019-52945-5 Discuss. Liaw, A., & Wiener, M. (2002). Classification and Obi, J. C., Awonuga, A. O., & Umeojiakor, A. O. (2010). Regression by randomForest. Researchgate. 3, 18–22. Spatial dependence of some physical properties of a typic Lin, L. I. (1989). A concordance correlation coefficient to plithaqaualf on the basement complex in southwestern evaluate reproducibility. Biometrics, 45(1), 255–268. nigeria. Journal of Tropical Agriculture, Food, https://doi.org/10.2307/2532051 Environment and Extension, 9(1), 38–46. https://doi.org/ Lin, Y. -H., Lee, P. -C., Menyailo, O. V., & Cheng, C. -H. 10.4314/as.v9i1.57456 (2021). Changes in soil organic carbon concentration and Ontl, T. A., & Schulte, L. A. (2012). Soil Carbon Storage. stock after forest regeneration of agricultural fields in Nature Education Knowledge, 3(10), 35. Taiwan. Forests, 12(9), 1222. https://doi.org/10.3390/ Piri Sahragard, H., & Pahlavan Rad, M. R. (2020). Prediction f12091222 of soil properties using random forest with sparse data in Liu, F., Wu, H., Zhao, Z., Li, D., Yang, J., Song, X., Shi, Z., a semi-active volcanic mountain. Eurasian Soil Science, 53 Zhu, A., & Zhang, G. (2022). Mapping high resolution (9), 1222–1233. https://doi.org/10.1134/ national soil information grids of China. Science Bulletin, s1064229320090136 67(3), 328–340. https://doi.org/10.1016/j.scib.2021.10.013 Poggio, L., de Sousa, L. M., Batjes, N. H., Heuvelink, G. B. M., Li, T., Zhang Wang, X., Cheng, S., Fang, H., Liu, W., & Kempen, B., Ribeiro, E., & Rossiter, D. (2021). SoilGrids Yuan, J. (2019). Soil erosion affects variations of soil 2.0: Producing soil information for the globe with quanti- organic carbon and soil respiration along a slope in fied spatial uncertainty. SOIL, 7(1), 217–240. https://doi. Northeast China. Ecological Processes, 8(1). https://doi. org/10.5194/soil-7-217-2021 org/10.1186/13717-019-0184-6 Pouladi, N., Møller, A. B., Tabatabai, S., & Greve, M. H. Lu, Y., Liu, F., Zhao, Y., Song, X., & Zhang, G. (2019). An (2019). Mapping soil organic matter contents at field level integrated method of selecting environmental covariates with Cubist, Random Forest and kriging. Geoderma, 342, for predictive soil depth mapping. Journal of Integrative 85–92. https://doi.org/10.1016/j.geoderma.2019.02.019 Agriculture, 18(2), 301–315. https://doi.org/10.1016/ R Core Team. (2020). R: A language and environment for s2095-3119(18)61936-7 statistical computing. R Foundation for Statistical Malone, B. P., McBratney, A. B., Minasny, B., & Computing. URL https://www.R-project.org/ Laslett, G. M. (2009). Mapping continuous depth func- Reeuwijk, L. P. V., 2002. PROCEDURES for SOIL tions of soil carbon storage and available water capacity. ANALYSIS: Technical paper 9, procedures for soil Geoderma, 154(1–2), 138–152. https://doi.org/10.1016/j. analysis. International SOil Reference and Information geoderma.2009.10.007 Center, Mayer, S., Kühnel, A., Burmeister, J., Kögel-Knabner, I., & Rodriguez, J. D., Perez, A., & Lozano, J. A. (2009). Sensitivity Wiesmeier, M. (2019). Controlling factors of organic analysis of k-fold cross validation in prediction error carbon stocks in agricultural topsoils and subsoils of estimation. IEEE Transactions on Pattern Analysis and Bavaria. Soil and Tillage Research, 192, 22–32. https:// Machine Intelligence, 32(3), 569–575. https://doi.org/10. doi.org/10.1016/j.still.2019.04.021 1109/TPAMI.2009.187 Mayes, M., Marin-Spiotta, E., Szymanski, L., Akif Rouse, J. W., Haas, R. H., Schell, J. A., & Deering, D. W. Erdoğan, M., Ozdoğan, M., & Clayton, M. (2014). Soil (1974). Monitoring vegetation systems in the Great Plains type mediates effects of land use on soil carbon and nitro- with ERTS. In S. C. Freden, E. P. Mercanti, & M. Becker gen in the Konya Basin, Turkey. Geoderma, 232-234, (Eds.), Third earth resources technology satellite–1 sympo- 517–527. https://doi.org/10.1016/j.geoderma.2014.06.002 sium. volume i: technical presentations, NASA SP-351 (pp. McBratney, A., Mendonça Santos, M., & Minasny, B. (2003). 309–317). NASA. On digital soil mapping. Geoderma, 117(1–2), 3–52. Six, J., Conant, R. T., Paul, E. A., & Paustian, K. (2002). https://doi.org/10.1016/s0016-7061(03)00223-4 Stabilization mechanisms of soil organic matter. Minasny, B., McBratney, A. B., Malone, B. P., & Wheeler, I. Implications for C-Saturation of Soils, 241(2), 155–176. (2013). Digital mapping of soil carbon. Advances in https://doi.org/10.1023/a:1016125726789 GEOLOGY, ECOLOGY, AND LANDSCAPES 15 Smith, P., Martino, D., Cai, Z., Gwary, D., Janzon, H., Wang, S., Adhikari, K., Wang, Q., Jin, X., & Li, H. (2018). Kumar, P., McCarl, B., Ogle, S., O’Mara, F., Rice, C., Role of environmental variables in the spatial distribution Scholes, B., & Sirotekno, O. (2007). Agriculture. of soil carbon (C), nitrogen (N), and C:N ratio from the CLimate Change 2007: Mitigation. Contribution of northeastern coastal agroecosystems in China. Ecological Working Group III to the Fourth Assessment Report of indicators, 84, 263–272. https://doi.org/10.1016/j.ecolind. the Intergovernmental Panel on Climate. 2017.08.046 Sudi, D., Meseret, C., & Sasikumar, J. (2018). Floristic com- Wiesmeier, M., Urbanski, L., Hobley, E., Lang, B., von position, structural analysis and regeneration status of Lützow, M., Marin-Spiotta, E., Kögel-Knabner, I. van woody species of natural forest in Gemechis District of Wesemael, B, Rabot E, Ließ M, Garcia-Franco N, West Hararghe Zone, Oromia, Ethiopia. Journal of Natural Wollschläger U. (2019). Soil organic carbon storage as Sciences Research, 8, 11–24. a key function of soils - a review of drivers and indicators Sun, W., Huang, Y., Zhang, W., & Y., Y. (2010). Carbon at various scales. Geoderma, 333, 149–162. https://doi. sequestration and its potential in agricultural soils of org/10.1016/j.geoderma.2018.07.026 China. Global Biogeochem, 24. https://doi.org/10.1029/ Wilson, J. P., & Gallant, J. C. (2000). Digital terrain analysis. 2009GB003484 In J. P. Wilson & J. C. Gallant (Eds.), Terrain analysis: Smith, P., D., Martino, Z., Cai, D., Gwary, H., Janzon, P., Principles and applications (pp. 1–27). John Wiley & Kumar, B., McCarl, S., Ogle, F., O’Mara, C., Rice, B., Sons. Scholes, O., Sirotekno. (2007). Agriculture. In CLimate Wynn, J. G., Bird, M. I., Vellen, L., Grand-Clement, E., Change 2007: Mitigation. Contribution of Working Carter, J., & Berry, S. L. (2006). Continental-scale mea- Group III to the fourth Assessment Report of the surement of the soil organic carbon pool with climatic, Intergovernmental Panel on Climate Change [B. Metz, edaphic, and biotic controls. Global biogeochemical cycles, O.R. Davidson, P.R. Bosch, R. Dave, L.A. Mayer(eds)], 20(1), n/a–n/a. https://doi.org/10.1029/2005gb002576 Cambridge University Press, Cambridge, United Xia, L., Lam, S. K., Yan, X., & Chen, D. (2017). How does Kingdom and New York ,NY, USA. recycling of livestock manure in agroecosystems affect Temesgen, A., Bobe, B., Alemayehu, R., Lemma, W., & crop productivity, reactive nitrogen losses, and soil car- Samuel, F. (2022). Spatial variability of soil organic car- bon balance? Environmental Science & Technology, 51 bon stock in Gurje Subwatershed, Hadiya Zone, Southern (13), 7450–7457. https://doi.org/10.1021/acs.est.6b06470 Ethiopia. Applied and Environmental Soil Science, 2022, Xiong, X., Grunwald, S., Myers, D. B., Ross, C. W., 12. https://doi.org/10.1155/2022/5274482 Harris, W. G., & Comerford, N. B. (2014). Interaction Tessema, T., & Kibebew, K. (2019). Carbon stock under effects of climate and land use/land cover change on soil major land use/land cover types of Hades sub-watershed, organic carbon sequestration. The Science of the Total eastern Ethiopia. Carbon Balance and Management, 14(1). Environment, 493, 974–982. https://doi.org/10.1016/j.sci https://doi.org/10.1186/s13021-019-0122-z totenv.2014.06.088 Tufa, D., Hansen, J., Rose, A., Damen, B., & Sheinkman, M. Yang, R. -M., Zhang, G. -L., Liu, F., Lu, Y. -Y., Yang, F., (2018). Enhancing National Climate Services (ENACTS) Yang, F., Li, D. -C. (2016). Comparison of boosted regres- approach to support climate resilience in agriculture sion tree and random forest models for mapping topsoil Improving availability, access, and use of historical cli- organic carbon concentration in an alpine ecosystem. mate information.Info Note. Ecological indicators, 60, 870–878. https://doi.org/10. Vågen, T. -G., & Winowiecki, L. A. (2013). Mapping of soil 1016/j.ecolind.2015.08.036 organic carbon stocks for spatially explicit assessments of Yigini, Y., & Panagos, P. (2016). Assessment of soil organic climate change mitigation potential. Environmental carbon stocks under future climate and land cover Research Letters, 8(1), 015011. https://doi.org/10.1088/ changes in Europe. The Science of the Total 1748-9326/8/1/015011 Environment, 557-558, 838–850. https://doi.org/10.1016/ Vermote, E., Justice, C., Claverie, M., & Franch, B. (2016). j.scitotenv.2016.03.085 Preliminary analysis of the performance of the Landsat 8/ Zevenbergen, L. W., & Thorne, C. R. (1987). Quantitative OLI land surface reflectance product. Remote Sensing of analysis of land surface topography. Earth Surface Environment, 185, 46–56. https://doi.org/10.1016/j.rse. Processes and Landforms, 12(1), 47–56. https://doi.org/ 2016.04.008 10.1002/esp.3290120107 Walkley, A., & Black, I. A. (1934). An examination of the Zhang, Y., Guo, L., Chen, Y., Shi, T., Luo, M., Ju, Q., Degtjareff method for determining soil organic matter Wang, S. (2019). Prediction of soil organic carbon based and a proposed modification of the chronic acid titration on Landsat 8 monthly NDVI data for the jianghan plain method. Soil science, 37(1), 29–38. https://doi.org/10. in Hubei Province, China. Remote Sensing, 11(14), 1683. 1097/00010694-193401000-00003 https://doi.org/10.3390/rs11141683 16 M. K. NEGASSA ET AL. Appendix Figure A1. Map of covariates.
Journal
Geology Ecology and Landscapes – Taylor & Francis
Published: Jan 20, 2023
Keywords: The soil; random forest; SOC stock prediction; environmental covariates; digital soil mapping