Integrating Landscape Pattern Metrics to Map Spatial Distribution of Farmland Soil Organic Carbon on Lower Liaohe Plain of Northeast China

Abstract

Accurate digital mapping of farmland soil organic carbon (SOC) contributes to sustainable agricultural development and climate change mitigation. Farmland landscape pattern has changed greatly under anthropogenic influence, which should be considered an environmental variable to characterize the impact of human activities on SOC. In this study, we verified the feasibility of integrating landscape patterns in SOC prediction on Lower Liaohe Plain. Specifically, ten variables (climate, topographic, and landscape pattern variables) were selected for prediction with Random Forest (RF) and Support Vector Machines (SVMs). The effectiveness of landscape metrics was verified by establishing different variable combinations: (1) natural variables, and (2) natural and landscape pattern variables. The results confirmed that landscape variables improved mapping accuracy compared with natural variables. R² of RF and SVM increased by 20.63% and 20.75%, respectively. RF performed better than SVM with smaller prediction error. Ranking of importance of variables showed that temperature and precipitation were the most important variables. The Aggregation Index (AI) contributed more than elevation, becoming the most important landscape variable. The Mean Contiguity Index (CONTIG-MN) and Landscape Contagion Index (CONTAG) also contributed more than other topographic variables. We conclude that landscape patterns can improve mapping accuracy and support SOC sequestration by optimizing farmland landscape management policies.

1. Introduction

Soil plays a key role in providing ecosystem services that support life on Earth [1,2,3], and soil carbon plays a central role in many soil functions, contributing to its use as a universal indicator of soil properties [4,5]. Agricultural land (4.86 × 10⁷ km²) accounts for 38.18% of the world’s land area (12.74 × 10⁷ km²), and 8–10% of the global soil carbon storage [6]. Farmland soil organic carbon (SOC) is a critical indicator of soil quality and fertility, as well as an important source of agricultural greenhouse gas emission reduction potential [7]. Therefore, accurately understanding the spatial distribution of farmland SOC is essential for mitigating climate change and ensuring food security [8].

Digital soil mapping (DSM) makes it possible to understand the spatial distribution of SOC on different scales [3,9,10], and has the advantages of high efficiency and low cost [9,11]. It has been performed by extracting the soil–environment relationships with covariates such as topography, climate, and soil types [12]. However, farmland is generally located on plains or in low-lying areas [13], where natural variables are less heterogeneous, causing the difficulty of characterizing spatial distribution of agricultural SOC using weakly correlated environmental factors [6]. Farmland is greatly affected by human activities. Some studies have reported that the impact of human activities on farmland SOC content may exceed that of natural factors [14,15]. Therefore, recent studies have focused on characterizing new variables to reflect the impact of human activities on farmland SOC, for example, cropping systems [16], agricultural land use history [17], population and socioeconomic variables [15]. However, there are difficulties in obtaining these data on a large scale. The lack of environmental variables that can characterize the impact of human activities on SOC limits the accuracy of SOC mapping in agricultural areas [11].

Land use types are regarded as important human activities [16]. Different spatial and temporal combinations of land use types will produce different landscape patterns [18]. Human-induced changes in landscape patterns often affect ecosystem functions and services [19,20]. Changing the composition and configuration of the landscape alters nutrient cycling and energy transformation [21]. Some studies have proved that the land use pattern has a “spillover effect” on the surrounding SOC. Pardon et al. confirmed that SOC and soil nutrient were significantly higher in the vicinity of trees in field boundaries, which affected the soil conditions at least 30 m away from the tree row [22]. Didham et al. observed that fertilization in farmland resulted in an increase in soil nutrients within adjacent forest [23]. Wu et al. found that a highly connected and water-sufficient landscape was conducive to the accumulation of farmland SOC [24]. The effect of farmland landscape patterns on SOC can be explained by landscape complementarity, that is, more heterogeneous landscapes allow for more efficient use of complementary resources within the landscape [25]. Single and homogeneous farmland landscapes are not conducive to the mutual complement and coordination of resources. For example, trees change conditions under the canopy through shading, root turnover, and litter input, which have significant effects on soil moisture, temperature, carbon substrate availability, and nutrient availability for the agricultural crop [22,26,27]. Therefore, the landscape pattern has great potential to be a good indicator of soil properties [24]. However, most studies focused on the relationship between land use type and SOC at different sampling sites, while ignoring the impact of the surrounding land use landscape pattern on SOC at the sampling sites [28]. In digital soil mapping, the lack of consideration of farmland landscape patterns is not conducive to improving the accuracy of SOC mapping.

Landscape pattern metrics can be used to quantitatively analyze the correlations between landscape patterns and ecological processes [20], for example, the relationships between landscape pattern metrics and soil arthropods [29], and the net primary productivity level [20]. Landscape pattern metrics support analysis at different scales. At present, many studies have quantitatively expressed the relationships between landscape pattern metrics and SOC in urban or forest ecosystems, and found that metrics representing patch shape and landscape aggregation, such as the Area-Weighted Patch Shape Index (SHAPE_AM), Aggregation Index (AI), and Landscape Shape Index (LSI) were significantly related to SOC [30,31,32]. However, the knowledge of farmland ecosystems and SOC is insufficient; in particular, the application of landscape pattern variables in farmland SOC digital mapping remains to be further verified. The number, size, shape, and spatial distribution of farmland patches can greatly alter ecological processes [33]. Furthermore, each landscape exhibits a unique pattern at each scale due to different ecological processes [34]. Wu et al. demonstrated a scale-dependent relationship between landscape metrics and SOC density [24]. Thus, it is necessary to understand the relationship between landscape patterns and SOC at different scales. Among these, landscape and regional scales are the most operational spatial scales for the study of sustainable processes and mechanisms [35]. Spatial modeling integrated with landscape metrics can not only quantify the impact of human activities on SOC at a large scale, but also have a positive impact on guiding sustainable landscape management practices.

This study was carried out on Lower Liaohe Plain of northeastern China, where the landscape pattern has undergone long-term human modification. The Northeast Plain has always been an ideal area to study the anthropogenic influence on SOC distribution in agricultural ecosystems [15]. Lower Liaohe Plain provides a good opportunity to quantify the extent to which the landscape pattern can improve the mapping accuracy of farmland SOC. In addition, there are important black soil resources in the Lower Liaohe Plain area, and providing a more accurate SOC spatial map plays an important role in sustainable utilization of black soil. Previous studies normally used one predictive model without comparing whether other models would improve the results [3]. A review by Stupariu et al. found that Random Forest (RF) and Support Vector Machines (SVMs) are commonly used machine learning algorithms in landscape ecological analysis [36]. Therefore, this study integrated landscape pattern metrics based on the use of these two models to map the spatial distribution of SOC. The main objectives of this study are to: (1) depict the new environmental variable, the landscape pattern, and define the contribution degree of landscape metrics in spatial mapping, (2) evaluate the performance of RF and SVM models, and (3) map the spatial distribution of SOC on Lower Liaohe Plain. This study can provide important information for improving the accuracy of farmland SOC spatial prediction in plain areas, which also provides a basis for realizing farmland SOC sequestration based on landscape pattern optimization.

2. Material and Methods

2.1. Study Area

Lower Liaohe Plain (120°42′–124°45′ E, 40°43′–43°27′ N) was selected as the study area (Figure 1). It is located in the center of Liaoning Province, and has an approximate area of 5.186 million hm². The study area is an alluvial plain, and the terrain has gentle slopes with most altitudes below 50 m. In general, the study area has the characteristics of a landscape pattern that is dominated by a man-made landscape and is in a relatively important position. Arable land is the main land use, and occupies more than 50% of the total land area [37]. The farmland management measures are similar, and chemical fertilizer is the main fertilizer. In recent years, the ratio of nitrogen, phosphorus, and potassium fertilizer has tended to be reasonable, and crop yields have tended to increase [38]. The study area has an obvious monsoon continental climate, with cold winters and hot summers, and dry winters and wet summers. The average annual temperature ranges from 7 to 11 °C, and annual precipitation ranges from 600 to 1100 mm [39]. Lower Liaohe Plain is an important commodity grain base and also the agglomeration area of central urban agglomeration and Shenyang Economic zone in Liaoning Province [40]. The main crop types are Zea mays and rice. The main soil types are brown soil, cinnamon soil, black soil, alluvial soil, aeolian sand soil, meadow soil, marsh soil, paddy soil, and red clay [40].

2.2. Soil Sampling

SOC data (0–20 cm) for this study came from the soil quality data set of a farmland fertility evaluation conducted in 2010. A total of 307 samples were included, where all samples were from dry land, and the crop type was Zea mays. Therefore, this study mainly predicted the spatial distribution of farmland SOC in dry land. Samples were collected after harvest and no crops were growing in the field during soil sample collection. At each sampling site, soil surface litter was removed first, then five random samples were collected and manually homogenized to form a composite sample, placed in self-sealing bags, numbered, and brought back to the laboratory. SOC was determined by Walkley–Black dichromate oxidation [41].

2.3. Environmental and Landscape Pattern Variables

Natural factors shape the basic spatial pattern of SOC, which is the basis of SOC mapping in farmland [17]. In this study, temperature, precipitation, four topographic variables, and nine landscape pattern variables were selected as prediction variables. All of the variables were collected and converted to raster data in ArcGIS 10.2 (ESRI Inc., Redlands, CA, USA) [42]. Since each data type has a different resolution, we used the nearest neighbor resampling method [11] to resample the raster data to 40 m resolution in order to meet data processing requirements and ensure the calculation accuracy of landscape pattern metrics.

2.3.1. Climate Variables

The annual mean temperature (MAT) and annual mean precipitation (MAP) were employed as climatic variables. The spatial interpolation data of MAT and MAP in the last 30 years (1982–2010) were downloaded from 673 meteorological stations of the National Meteorological Information Center of China Meteorological Administration (http://data.cma.cn/en (accessed on 30 January 2022)). The climate data originally had a resolution of 1000 m and were resampled to 40 m resolution in ArcGIS 10.2 for use in this study. The units of MAP and MAT are 0.1 mm and 0.1 °C, respectively.

2.3.2. Topographic Variables

We selected four terrain-related variables by referencing relevant literature, namely, elevation (ELE), slope gradient (SG), slope aspect (SA), and plane curvature (Planc) [10,15]. The topographic data were derived from the SRTM digital elevation model (DEM) of the Geographic Data Spatial Cloud (http://www.gscloud.cn (accessed on 30 January 2022)). Other related derivative factors were calculated using the Spatial Analyst module of ArcGIS 10.2 based on ELE. The resolution was 40 m.

2.3.3. Landscape Pattern Variables

In this study, one of the most important objectives was to verify the effectiveness of farmland landscape patterns in improving SOC mapping accuracy. Landscape metrics are useful methods for understanding the influences of human activities on the landscape [20]. Previous research has provided us with information about the correlation between landscape pattern metrics and SOC [24,43]. Therefore, we selected Patch Cohesion Index (COHESION), Landscape Contagion Index (CONTAG), Aggregation Index (AI), Mean Contiguity Index (CONTIG-MN), Landscape Shape Index (LSI), Area-weighted Mean Perimeter-Area Ratio (PARA-AM), Mean Fractal Dimension Index (FRAC-MN), Shannon’s Diversity Index (SHDI), and Shannon’s Evenness Index (SHEI) in this study (

Table 1. Descriptions of landscape pattern metrics.

Metrics/Units	Explanations	Formulas
COHESION (%)	Describe the physical connectivity between patches.	$\mathrm{COHESION}=1-\left(\frac{{\displaystyle \sum _{i=1}^m{\displaystyle \sum _{j=1}^n{{\displaystyle p}}_{ij}}}}{{\displaystyle \sum _{i=1}^m{\displaystyle \sum _{j=1}^n{{\displaystyle p}}_{ij}\sqrt{{{\displaystyle a}}_{ij}}}}}\right){\left(1-1/\sqrt{Z}\right)}^{-1}\times 100$ where pij is the perimeter in meters, aij is the area in square meters and Z is the number of cells.
CONTAG (%)	Reflect the degree of agglomeration or extended trend of different patch types. High value of CONTAG indicates that dominant patch types in the landscape are highly connected.	$\mathrm{CONTAG}=1+\frac{{\displaystyle \sum _{q=1}^{{{\displaystyle \mathrm{n}}}_a}{{\displaystyle p}}_q\ln \left({{\displaystyle p}}_q\right)}}{2\ln \left(t\right)}$ where pq is the adjacency table for all classes divided by the sum of that table and t is the number of classes in the landscape.
AI (%)	Describe the degree of patch aggregation in the landscape and reflects the dispersion of landscape elements in the landscape.	$\mathrm{AI}=\left[{\displaystyle \sum _{i=1}^m\left(\frac{{{\displaystyle g}}_{ii}}{\max -{{\displaystyle g}}_{ij}}\right){{\displaystyle p}}_i}\right]\times 100$ where gii is the number of like adjacencies based on the single-count method and max − gii is the classwise maximum number of like adjacencies of class i and Pi the proportion of landscape compromised of class i.
SHDI	Reflect the heterogeneity of landscape. The larger the value is, the greater the diversity of landscape and the richer the landscape types are.	$\mathrm{SHDI}=-{\displaystyle \sum _{i=1}^m\left({{\displaystyle p}}_i\ln {{\displaystyle p}}_i\right)}$ where Pi is the proportion of class i.
SHEI	Reflect the evenness of each patch distributed in the landscape. The smaller the value, the higher the landscape dominance, indicating that the landscape is dominated by one or a few types of patches.	$\mathrm{SHEI}=\frac{-{\displaystyle \sum _{i=1}^m\left({{\displaystyle p}}_i\ln {{\displaystyle p}}_i\right)}}{{{\displaystyle \ln }}_m}$ where Pi is the proportion of class i and m is the number of classes.
LSI	Describe the complexity of landscape shape.	$\mathrm{LSI}=\frac{E}{\min E}$ where E is the total edge length in cell surfaces and minE is the minimum total edge length in cell surfaces.
PARA-AM	Reflect the ratio of the weighted perimeter to area of each patch area in landscape, can reflect the edge effect of patches.	$\mathrm{PARA}-\mathrm{AM}=\frac{{{\displaystyle p}}_{ij}}{{{\displaystyle a}}_i}$ where pij is the perimeter in meters, ai is the area of patch i.
CONTIG-MN	Assesses the spatial connectedness, or contiguity between patches.	$\mathrm{CONTIG}-\mathrm{MN}=\frac{\left[\frac{{\displaystyle \sum _{r=1}^z{{\displaystyle c}}_{ijr}}}{{{\displaystyle a}}_{ij}}\right]}{v-1}-1$ where cijr is the contiguity value for pixel r in patch ij, aij the area of the respective patch (number of cells) and v is the size of the filter matrix.
FRAC-MN	Reflect the complexity of landscape spatial shape.	$\mathrm{FRAC}-\mathrm{MN}=2\ln \left(0.25{{\displaystyle p}}_{ij}\right)/\ln {{\displaystyle a}}_{ij}$ where pij is the perimeter in meters, aij is the area in square meters.

COHESION, Patch Cohesion Index; CONTAG, Landscape Contagion Index; AI, Aggregation Index; SHDI, Shannon’s Diversity Index; SHEI, Shannon’s Evenness Index; LSI, Landscape Shape Index; PARA-AM, Area-Weighted Mean Perimeter-Area Ratio; CONTIG-MN, Mean Contiguity Index; FRAC-MN, Mean Fractal Dimension Index.

). The vector data of land use in 2010 was converted into raster data and land use types were divided into farmland, woodland, grassland, wetland, water body, and artificial surface. We firstly determined the granularity of landscape analysis, which was usually the minimum resolution of the raster when transforming spatial vector data into raster data [44], and it was 40 m in this study. The optimal amplitude was selected based on the optimal granularity. In the calculation of the moving window, different window radius sizes would cause different degrees of variation in the landscape metric. We used the Moving Window function in Fragstats 4.2 software [45], with window radius set to 40 m as the starting point, 760 m as the end point, and 40 m as the change interval, and calculated the value of landscape metrics under different windows. Then, ArcGIS 10.2 was used to generate random points, and each random point was assigned numerical values and geographic coordinates. Finally, the attribute values of each point were extracted into GS+, and the block basis ratio was calculated at different moving window radii [29]. The feature scale was identified to be 600 m, and 600 m landscape pattern metrics were used in this study. The calculation steps were based on Guo et al. [29] and Su [44]. We calculated the landscape pattern metrics at the landscape level, which mainly reflects the overall spatial configuration of the landscape around sampling points [28]. Appendix A provides the spatial distribution map of each landscape metric.

2.4. Prediction Models

Four different models were established for farmland SOC prediction. The first model used natural variables only trained by RF (RF_1), the second model used natural variables and landscape pattern variables trained by RF (RF_2), the third model used natural variables only trained by SVM (SVM_1), and the fourth model used natural variables and landscape pattern variables trained by SVM (SVM_2).

2.4.1. Random Forest

Random Forest (RF) is a classification regression algorithm based on decision trees [46]. It uses the bootstrap resampling method to randomly extract several samples from the original data set, conducts decision tree modeling for each sample data, and then combines them into a prediction model of multiple decision trees. The final prediction result is the average of the predicted values for multiple trees. In RF modeling, the user must specify two training parameters: n_estimators and max_depth. N_estimators refers to the number of decision trees, that is, the number of times of re-sampling by bootstrap, and max_depth refers to the maximum depth of decision trees. The default value of max_depth is one-third of the number of predictors [17], which in this study is 3. The value of n_estimators is the minimum that generates an error within the stable range of the RF model [17], which in this study is 1000. Moreover, RF can measure the relative importance of each covariable by quantifying the increase in prediction error when a predictor is scrambled in the validation data [47]. The index%IncMSE (increase in mean squared error) was used to characterize the relative importance of the predictor variables [48]. RF model construction and prediction were implemented using the Random Forest Regressor package in the Python (3.9.0) Scikit-Learn library. Variable relative importance sorting was implemented with a call to the Feature_importances attribute. The ArcGIS10.2 mapping module was used to complete thematic mapping of SOC spatial distribution.

2.4.2. Support Vector Machine

Support Vector Machine (SVM) is another prominent supervised machine learning technique for classification and regression [3]. The principle of SVM is to map data to high-dimensional feature space through a kernel function, and then build an optimal hyperplane in the new space according to the principle of structural risk minimization [49]. The type of kernel function selected in this study is the radial basis function (RBF), which can handle multiple regression problems [50]. The hyperplane function of SVM is expressed as f (x) = ωx + b, where b is the bias and ω is a vector of the hyperplane [51]. SVM model construction and prediction were implemented using the Support Vector Machines package in Python (3.9.0) Scikit-Learn library. ArcGIS10.2 was used to complete thematic mapping of SOC spatial distribution.

2.5. Evaluation of Model Accuracy

The 10-fold cross-validation method was used to evaluate the predictive performance of the models [29]. The statistical indices used to characterize the model accuracy were determination coefficient (R²), root mean square error (RMSE), and mean absolute prediction error (MAE). All indices were calculated using the Scikit-Learn feature pack of Python V 3.9.0. The indices were computed as follows:

$${\mathrm{R}}^2=1-\frac{{\displaystyle \sum _{i=1}^n{\left(P_i-O_i\right)}^2}}{{\displaystyle \sum _{i=1}^n{\left(O_i-\overline{O}\right)}^2}}$$

(1)

$$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n(P_i-O_i}{)}^2}$$

(2)

$$\mathrm{MAE}=\frac{1}{n}{\displaystyle \sum _{i=1}^n\left|P_i-O_i\right|}$$

(3)

where P_i and O_i are the predicted and observed SOC values at the location i, respectively; n is the number of total samples; $\overline{P}$ and $\overline{O}$ are the mean values of the predicted and observed SOC, respectively. The smaller the values of MAE and RMSE, and the closer R² to 1, the higher the accuracy of the model.

3. Results

3.1. Descriptive Statistics

Statistical descriptions of SOC content and environmental variables are shown in

Table 2. Summary statistics of SOC, environmental, and landscape metric variables of 307 samples.

Property/Unit	Min.	Max.	Mean	SD	Skewness	Kurtosis
SOC (g kg−1)	0.928	24.360	9.917	3.763	0.522	0.605
MAP (0.1 mm)	2112.970	5439.010	2798.404	516.924	1.858	5.394
MAT (0.1 °C)	6.521	44.382	21.873	4.697	1.413	4.689
ELE (m)	−29.000	309.000	58.498	61.492	1.797	3.265
SG (degree)	0.000	48.559	5.939	5.022	2.826	17.035
SA (degree)	−1.000	355.601	181.704	102.585	−0.133	−1.086
Planc	−5.583	3.949	0.062	1.058	−0.218	3.257
AI	77.788	99.413	94.487	4.282	−1.507	2.159
CONTAG	22.403	99.772	60.979	16.451	0.218	−0.543
CONTIG-MN	0.244	0.951	0.610	0.196	0.058	−1.263
FRAC-MN	1.018	1.125	1.059	0.017	0.517	0.514

Min., minimum; Max., maximum; SD, standard deviation; MAP, mean annual precipitation; MAT, mean annual temperature; ELE, elevation; SG, slope gradient; SA, slope aspect; Planc, plane curvature; AI, Aggregation Index; CONTAG, Landscape Contagion Index; CONTIG-MN, Mean Contiguity Index; FRAC-MN, Mean Fractal Dimension Index.

. SOC content ranged from 0.928 g kg⁻¹ to 24.360 g kg⁻¹ and its average value was 9.917 g kg⁻¹. The Kolmogorov–Smirnov test confirmed that SOC showed no significance (p = 0.084 > 0.05), and was in line with a normal distribution.

In order to obtain more stable results for the relative importance of SOC predicted by the selected variables, only variables with a Pearson correlation coefficient less than 0.75 were used in the model [52]. Variables used for prediction were ELE, SG, SA, Planc, MAT, MAP, AI, CONTIG-MN, FRAC-MN, and CONTAG (Table 2). The Pearson correlation coefficients between all variables and SOC are shown in Figure 2, although we only describe the correlation between variables used for spatial mapping and SOC here. SOC was positively correlated with ELE (r = 0.174 **) (p < 0.01), and no significant correlations were found between SOC with other topographic variables. For the climate variables, MAP (r = 0.367 **) (p < 0.01) was positively correlated with SOC, while MAT (r = −0.241 **) (p < 0.01) was negatively correlated with SOC. For the landscape metric variables, AI (r = −0.357 **) (p < 0.001), CONTAG (r = −0.238 **), and CONTIG-MN (r = −0.194 **) (p < 0.001) were negatively correlated with SOC. There was no significant correlation between FRAC-MN and SOC (p < 0.005).

3.2. Model Performance of Different Variable Combinations

Table 3. Evaluation of prediction performances of four models using 10-fold cross-validation.

Models	Variable Combinations	R2	RMSE	MAE
RF_1	MAP, MAT, ELE, SG, SA, Planc	0.286	3.042	2.349
RF_2	MAP, MAT, ELE, SG, SA, Planc, AI, CONTAG, CONTIG-MN, FRAC-MN	0.345	2.961	2.282
SVM_1	MAP, MAT, ELE, SG, SA, Planc	0.212	3.159	2.453
SVM_2	MAP, MAT, ELE, SG, SA, Planc, AI, CONTAG, CONTIG-MN, FRAC-MN	0.256	3.139	2.400

MAP, mean annual precipitation; MAT, mean annual temperature; ELE, elevation; SG, slope gradient; SA, slope aspect; Planc, plane curvature; AI, Aggregation Index; CONTAG, Landscape Contagion Index; CONTIG-MN, Mean Contiguity Index; FRAC-MN, Mean Fractal Dimension Index; R², coefficient of determination; RMSE: root mean square error; MAE: mean absolute error.

shows the accuracy results of different variable combinations. Although the R² values of all combinations were not high, when landscape metric variables (AI, CONTAG, CONTIG-MN, and FRAC-MN) (RF_2 and SVM_2) were added to only natural variable combinations (RF_1 and SVM_1), R² increased by 20.63% for RF and 20.75% for SVM; RMSE decreased from 3.042 to 2.961 for RF, and 3.159 to 3.139 for SVM; and MAE decreased from 2.349 to 2.282 for RF, and from 2.453 to 2.400 for SVM, respectively. This indicates that, on the basis of natural variables, incorporating landscape pattern variables can improve the prediction accuracy of SOC. Appendix B provides a scatter plot of observed and predicted SOC values from the four models to further verify the importance of incorporating landscape variables in SOC prediction. In addition, RF showed better performance than SVM whether considering natural variables only or the combination of natural and landscape pattern variables, which indicates that RF was an effective predictive model for predicting farmland SOC in plain areas.

3.3. Relative Importance of the Variables

The relative importance of the variables used to predict SOC is shown in Figure 3. In the ranking that combined natural and landscape pattern variables, MAP and MAT were the most important variables in spatial mapping, and ranked first and second, respectively. The contributing rate of the landscape pattern metric AI was 19.553%, which exceeded ELE (8.287%), and became the most important landscape variable. CONTIG-MN (3.504%) and CONTAG (2.256%) also contributed more than other topographic variables. In the ranking of only considering natural variables, MAP (50.500%) and MAT (30.052%) contributed more than 80% to SOC variation, while terrain and other related factors played a weak role. These results suggested that landscape pattern variables selected in this study played an important role in spatial mapping.

3.4. The Predicted Map of SOC

Figure 4 shows the spatial distribution of SOC contents predicted using RF and SVM models in this study. The four maps showed a similar spatial distribution trend, but with different levels of details. Overall, the spatial pattern of SOC was higher in the east than in the west. The high-value regions were mainly concentrated in the southeast, while the low-value regions were in the north and west of the study area. The value ranges of SOC predicted by RF were smaller than those predicted by SVM. The low values of SOC predicted by RF were higher than those of the SVM model, while the high values were lower than those of SVM. When landscape pattern metrics were integrated for mapping, SOC could be displayed in more detail and accuracy, and the map of RF models was better than that of SVM. To evaluate the predicted SOC maps, we also provided a land use map of Lower Liaohe Plain and an SOC map based on ordinary Kriging interpolation, as referenced in Appendix C and Appendix D. We further analyzed the regions with significant differences. Box 1 and Box 2 are mainly located in Kangping county. According to RF models, the spatial distribution characteristics of SOC showed a trend of high in the south and low in the north. The northern area is adjacent to Horqin Sandy Land, resulting in low SOC content. Box 2 showed more spatial information than box 1. As the advancement of some ecosystem restoration projects promoted the restoration of surface vegetation, the RF_2 map combined with landscape variables showed the influence of woodland or grassland on farmland SOC; thus, the SOC content was higher in some sites. Box 3 and box 4 are mainly distributed in Liaoyang city, which is located in the transition zone between the low mountains of the Liaodong region and Liaohe Plain. Erosion gullies of sloping farmland in some areas may lead to the decrease in SOC content. The SOC content of SVM_1 was higher than that of SVM_2, which may be an overestimation of SOC.

4. Discussion

4.1. Potential of Integrating Landscape Metrics for SOC Spatial Mapping

In this study, we confirmed that climatic factors had the greatest influence on SOC spatial distribution in Lower Liaohe Plain and played an important role in spatial mapping. This was consistent with the existing findings that climatic conditions were the main driver of SOC at large geographic scales [53,54], because climate affects both the input of carbon into the soil and the decomposition of organic carbon [5]. Among the topographic factors, only ELE had a positive correlation with SOC, and the relationships between other topographic factors and SOC were not significant. Similar results were reported by Dong et al., who found that elevation was the only one factor that had a significant effect on spatial distribution of soil nutrients in ten topographic factors, which could be ascribed to the fact that land has been leveled in agriculture areas and the role of other factors has been weakened [9]. Nevertheless, it should be noted that ELE has both direct and indirect effects on the spatial distribution of SOC [55]. With the change in elevation gradient, vegetation, soil, and climate conditions have corresponding changes [56], thus affecting SOC content. Therefore, ELE could explain the spatial variation of SOC to a certain extent even when the terrain of the study area was gentle. However, when landscape pattern metrics were taken into account, we observed that the roles of topographic factors were weakened, while the variables representing human activities were more effective. Most of the landscape metric variables showed significant correlations with SOC (except FRAC-MN). In general, combined with the ecological implications of each landscape pattern metric, we found that highly spread and concentrated farmland landscapes had a negative impact on SOC, highly connected farmland patches had a negative impact on SOC, while diverse landscape types and complex patch shapes were conducive to SOC sequestration.

Lower Liaohe Plain is a typical landscape pattern formed by long-term cultivation [57], with an average cultivation history of 200 years [58]. Longer cultivation histories significantly affect SOC levels by influencing soil properties and biomass production [59], and continuous change in landscape patterns also indirectly affects SOC content. Farming practices over a long time have led to the massive transformation of structurally diverse and complex farmland landscapes into a homogeneous and intensive-use environment [60]. The gradual replacement of these non-cultivated landscapes with cultivated landscapes has negatively affected SOC for the following reasons. Semi-natural landscape elements, such as woodlands and grasslands, can protect soil structure, fix carbon, and promote carbon accumulation [24,61]. In addition, semi-natural landscapes also contribute to functional agrobiodiversity in agroforestry systems, as trees and their associated understory vegetation belts help to preserve arthropod biodiversity and enhance arthropod-associated ecosystem services in tree rows and arable land zones [62]. Trees act as “resource islands” for soil macrofauna in agricultural landscapes, and continuous cultivation activities will result in only a few sparse trees remaining, thus affecting the abundance of soil macrofauna and the key soil functions of soil-mediated ecosystem services supported by trees [27]. Soil fauna have a significant impact on soil formation and the flow of matter and energy. For example, earthworms are good evidence, as multi-species earthworm communities are most capable of transporting carbon and nitrogen from surface litter to the soil [63]. Compared with traditional agricultural systems, the abundance of earthworms in agroforestry systems was higher [26], which obviously has positive significance for SOC sequestration. Therefore, we concluded that farmland landscapes with high contagion limited biological activities within soil or on the surface to some extent, thus further affecting the carbon sequestration function, while the structurally diverse farmland could significantly promote ecosystem functions and services [60] and increase the potential of SOC sequestration in farmland.

The results also indicated that highly connected and simply shaped farmland patches had negative effects on SOC. Human activities can affect the connectivity between landscapes and the complexity of patch shapes [64], which further affect SOC. Visually isolated and discontinuous patches appear as highly connected farmland patches after external forces; as a result, farmland became the main landscape matrix. Zhou et al. found that poor landscape connectivity was conducive to the improvement in net primary productivity at the landscape scale; when landscape types were few and concentrated, net primary productivity values were lower [20]. Moreover, partial natural or semi-natural habitat loss leads to simplification and regularization of patch shapes, as natural landscapes are usually characterized by irregular shape units with less obvious boundaries [65]. Eliminating some natural and semi-natural types to increase farmland area is mainly undertaken to maximize the production function of farmland, but this itself has certain risks, which will lead to the reduction in ecological functions, including SOC loss and biodiversity decline. Li et al. highlighted the negative impacts of increased farmland area, such as road erosion and associated nutrient losses, while reducing the average farmland size may have a positive impact on biodiversity within farmland (i.e., the crop-grown portion of the landscape) [66], which was more strongly dependent on the presence of semi-natural farmland boundary habitats [67]. The farmland landscape pattern is largely controlled by humans and is susceptible to landscape management policies [68]. Thus, the results of this study demonstrated the potential to reduce the negative impact of agricultural intensification on SOC through rational allocation of farmland landscapes.

4.2. Model Performance

The spatial characteristics of the natural landscape in plain or flat terrain areas are similar, and the spatial variation of most traditional soil-forming factors is small, causing challenges in the accuracy of farmland SOC mapping in plain areas [6,68]. Thus, the R² values reported by many studies were generally not high. Malone et al. found that agricultural areas performed poorly, with R² between 0.20 and 0.27 [69]. Zhang et al. applied five models to predict SOC, and observed that R² was between 0.281 and 0.391 [68]. Dong et al. found that R² of SOM content in topsoil ranged from 0.25 to 0.33 [9]. Overall, our results were comparable to those reported in these studies, with R² between 0.212 and 0.345.

In this study, R² of RF_1 and SVM_1 models in which only natural variables were considered was 0.286 and 0.212, respectively, and R² of the improved RF_2 and SVM_2 models, which considered natural and landscape pattern variables, was 0.345 and 0.256, respectively. Moreover, by comparing the prediction results of RF and SVM, it could be seen that RF performed better than SVM in Lower Liaohe Plain. The advantages of RF models have been mentioned in many studies. With different landscape features and limited sample sizes, RF seems to outperform other models [70]. Lamichhane et al. observed that the RF model outperformed Stepwise Multiple Linear Regression Kriging in the prediction of SOC in an alluvial plain area of Nepal [52]. Another study predicted and quantified the distribution of farmland SOC at a national scale, and found that the RF model performed best among all the tested models, followed by Cubist, and then Extreme Gradient Boosting models [3]. Ferreira et al. compared the performance of regression tree, RF, and SVM, and found that the RF model generated the most accurate maps of SOC [10]. These results highlighted the advantages of RF model to predict the relationships between environmental and soil variables, especially at large geographical scales. With the increase in the research scale, the relationships between soil and landscape are relatively complicated. The RF model can solve the nonlinear problem well, which can not only reflect the overall trend of SOC spatial distribution, but also express the local variation more accurately. However, considering that different models may adapt to different variable combinations and research scales, it is necessary to compare and analyze the prediction accuracy of different models.

4.3. Limitations of This Study and Future Perspectives

Although landscape analysis methods provide opportunities to characterize farmland landscapes at different scales, few studies have considered the impact of landscape patterns on SOC. Thus, the landscape pattern metrics selected in this study may not be comprehensive due to the lack of a relevant research base. In addition, although the study confirmed that integration of landscape pattern metrics can improve the model accuracy, the overall R² has yet to be improved. Some variables, such as cropping system [16], phenological parameters [11], population, and gross domestic product information [15] can be included in future spatial mapping studies as human-related variables. Due to the existence of the scale effect, studies at different spatial scales can provide more accurate support for clarifying the relationship between farmland landscapes and SOC.

5. Conclusions

In this study, natural and landscape variables were selected to predict the spatial distribution of SOC, using RF and SVM, on Lower Liaohe Plain of northeast China. Landscape pattern metrics were used as proxy indicators reflecting the intensity of human activities on farmland SOC. The results confirmed the validity of landscape pattern metrics in farmland SOC mapping. On the basis of natural variables, combined with landscape metric variables, R² of RF and SVM increased by 20.63% and 20.75%, respectively. RMSE and MAE also showed a downward trend. Furthermore, RF had better prediction performance than SVM, and the predicted SOC range based on the optimal model was 5.400–15.665 g/Kg. In addition, the importance ranking results showed that natural variables (temperature and precipitation) shaped the basic pattern of farmland SOC and were the most important variables in spatial prediction, while the contribution rate of the landscape pattern metric AI exceeded that of ELE and became the most important landscape variable. CONTIG-MN and CONTAG were also more important than other terrain variables. The results highlight the importance of considering anthropogenic variables in spatial mapping of low-lying agricultural areas. This study provides support for improving the accuracy of farmland SOC mapping at a large geographical scale, and has a positive impact on landscape management policies in the study area.