Application of a combinatorial approach for soil organic carbon mapping in hills

Highlights

• A combined method is developed to map SOC under complex topography.

• The quantified environmental factors are employed in SOC mapping.

• The complexities relationships of SOC and local factors are considered.

• This approach performs better on capturing spatial details.

• The combined method is more accurate for SOC mapping in hills.

Abstract

Accurate mapping of soil organic carbon (SOC) is critical to improve C management and develop sustainable management policies. However, it is constrained by local variations of the model parameters under complex topography, especially in hills. This study applied a methodological framework to optimize the spatial prediction of SOC in the hilly areas during 1981–2012 by quantifying the relative importance of environmental factors, which include both qualitative factors and quantitative variables. Results showed that SOC increased twofold with a moderate spatial dependence during the past 32 years. During this period, land use patterns, soil groups, topographic factors, and vegetation coverage had significant impacts on the SOC changes (p < 0.01). Specifically, the impact of land use patterns has exceeded the impact of soil groups and became the dominant factor affecting SOC changes. Meanwhile, impacts from the topographic factors and vegetation coverage have substantially declined. Based on those results, a combinatorial approach (LS_RBF_HASM) was developed to map SOC using radial basis function neural network (RBF) and high accuracy surface modelling (HASM), and to generate more detailed spatial mapping relationships between SOC and the affecting factors. Compared with ordinary kriging (OK), land use-soil group units (LS) and HASM combined (LS_HASM), multiple linear regression (MLR) and HASM combined with LS (LS_MLR_HASM); LS_RBF_HASM showed a better performance with a decline of 6.3%–37.7% prediction errors and more accurate spatial patterns due to the quantitative combination of auxiliary environmental variables and more information on the SOC variations within local factors captured by RBF and HASM. Additionally, MLR may partially undermine the relationship of the internal spatial structure due to the highly nonlinear relation between SOC and environmental variables. This methodological framework highlights the optimization of more environmental factors and the calculation of spatial variability within local factors and provides a more accurate approach for SOC mapping in hills.

Introduction

Accumulation of soil organic carbon (SOC) is essential for mitigating climate change, improving crop production, and sustaining environmental services (McBratney et al., 2003; Lal, 2004; Sanderman and Chappell, 2013). Accurately assessing SOC content is important to regulating soil carbon and adapting to global environmental changes (Florinsky et al., 2002; Zhang et al., 2014; Wiaux et al., 2014; Rovai et al., 2018). The SOC is highly spatial heterogeneous due to the spatial heterogeneities in environmental factors and human activities (Wang et al., 2009; Mishra et al., 2010; Luo et al., 2021). Therefore, accurate representing of the spatial distribution of the SOC in earth system models remains a critical scientific challenge (Zhang et al., 2012; Gelaw et al., 2014; Adhikari et al., 2020).

Many tools based on geographic information systems (GIS) have been developed and tested in analysing soil processes and derive soil properties with spatial continuity (McBratney et al., 2003; Martin et al., 2014; Were et al., 2015). Ordinary Kriging (OK) is a classic method that predicts soil properties in unsampled areas from the neighbour soil observations based on the spatial autocorrelations among measurements of soil sampling points with no environmental factors used as auxiliary variables (McBratney et al., 2003), and thus shows a limited simulation accuracy in some areas with high spatial heterogeneity (Shi et al., 2011, Li et al., 2013; Rial et al., 2017; Keskin et al., 2019). Multiple Linear Regression (MLR) is the most frequently used method for predicting SOC (Grunwald, 2009; Liu et al., 2013; Lamichhane et al., 2019) using both the environmental correlation between soil variables and environmental parameters as well as the spatial autocorrelation in the residuals of the soil variables (McBratney et al., 2003; Minasny et al., 2013; Li et al., 2013). Geographically Weighted Regression (GWR) is developed to analyze spatial heterogeneity. Both MLR and GWR have improved performance by incorporating environmental variables and reducing spatial autocorrelations, respectively. However, relationships between soil properties and environmental variables are not just linear and spatial heterogeneity (McBratney et al., 2003; Li et al., 2017), interior uncertainties such as non-linearity between soil properties and environmental variables are widespread (Drake et al., 2006; Were et al., 2015). Machine learning techniques such as Neural Network (NN) and random forest (RF) can overcome the shortcomings of parametric and non-parametric statistical methods (Sharma et al., 2003; Grimm et al., 2008) including spatial autocorrelation, non-linearity, and overfitting (Behrens et al., 2005; Drake et al., 2006; Vågen et al., 2013; Yang et al., 2016). Although the NN applied to map SOC showed good performance (Li et al., 2013; Were et al., 2015), SOC residuals within local variables still exist. High Accuracy Surface Modelling (HASM) is a new model developed in terms of the fundamental theorem of surfaces (Yue et al. 2007, 2013, 2013) and has been confirmed with advantages in handling spatial residuals (Shi et al., 2011; Li et al. 2013, 2017). These studies have promoted the performance of SOC mapping, but a recent study showed that GWR and RF models with similar performances could not predict SOC accurately under complex topography (Mirchooli et al., 2020). It is thus unlikely that a single model can be developed to be applicable in soil mapping accurately under complex environmental conditions (Mishra et al., 2010; Li et al., 2017; Keskin et al., 2019; Lamichhane et al., 2019).

Soil mapping allows to regionalizing SOC by relating it to environmental variables for improving the prediction accuracy. However, many soil-forming environmental variables are not adequately represented in current models (Lamichhane et al., 2019; Adhikari et al., 2020), and about 80% of studies used less than three types of factors (McBratney et al., 2003; Schillaci et al., 2017; Lamichhane et al., 2019). Moreover, the relationships between environmental variables and soils vary across space (Sumfleth and Duttmann, 2008; Mashimbye et al., 2014; Liu et al., 2016; Lozano-García et al., 2016). Local factors such as topography and vegetation index can lead to additional spatial variations in small areas (Song et al., 2017; Adhikari et al., 2020; Bangroo et al., 2020; Heuvelink et al., 2021; Gibson et al., 2021). The relative importance of different factors is also often region-specific and scale-dependent (Anta et al., 2020; Balkovič et al., 2020; Luo et al., 2020; Abera et al., 2021). Such difference of local factors needs to be considered in SOC mapping. Although a few researchers have developed combined models and local factors for the spatially distributed modelling of SOC (Li et al., 2017; Lamichhane et al., 2019; Heuvelink et al., 2021), how to incorporate these issues within a more generic framework for SOC prediction with high spatial resolution remains insufficient.

Hilly areas are widely distributed around the world, and provide an array of ecosystem services to humans and animals including logistics supply and ecological buffering. Previous studies have revealed the complex relationships between environmental variables and SOC in hill areas (Chaplot et al., 2010; Li et al., 2013; Gibson et al., 2021; Moura-Bueno et al., 2021). This study aims to (1) identify the relative importance of different soil-forming environmental factors (i.e. land use, soil group, parent material, topographic and vegetation index) on the SOC in a hilly area in southwest China (Li et al., 2016), and (2) develop a combined approach based on HASM and NN to map hill's SOC change more accurately by incorporating environmental variables.

The following sections introduce the materials and methods (Section 2) including study area, data sources and processing, statistical analysis, methods for SOC, and performance assessment. Then Section 3 describes and demonstrates the SOC data statistics, related environmental factors, models for SOC mapping, and models validation. Section 4 discusses drives of SOC spatiotemporal variability and the performance of the methodological framework. Finally, Section 5 provides conclusions.