Improving spatial estimation of soil organic matter in a subtropical hilly area using covariate derived from vis-NIR spectroscopy

Highlights

• Principal component score of soil vis-NIR data served as covariate to estimate OM.

• Ordinary kriging (OK) and co-kriging (CK) were performed.

• CK used target variable from half of the total samples.

• CK provided much improved estimation over OK.

Spatial variability of soil properties is complex in the hilly areas of subtropical China that are heavily influenced by human activities. Large datasets are needed for achieving accurate spatial estimates of soil properties in these areas, and it can be difficult due to the budget and time restrictions. Soil diffuse reflectance contains the integrated information of soil properties, thus can provide useful auxiliary variables to improve the spatial estimation of soil properties. This study aimed to improve the spatial estimation of soil organic matter (OM) in a typical subtropical hilly area of China, with principal component (PC) extracted from soil visible and near-infrared (vis-NIR) reflectance spectroscopy used as co-variable. Spatial estimation was performed using ordinary kriging (OK) and co-kriging (CK). A set of 125 samples collected from soil surface layer was used. To simulate the under-sampling situation of target variable, a subset was constructed by randomly sampling 50% of the total samples. The results showed that compared to OK, CK provided the decreased root mean square error (RMSE) by about 45% in cross validations and improved estimation map with much lower estimation variances. This study indicates that for the soil properties and functional attributes that are well correlated with soil reflectance and are sparsely sampled, combined use of soil reflectance spectroscopy technique and multivariate geostatistical method will provide a powerful solution for accurate spatial estimation of soil OM.

Introduction

Soil organic matter (OM) is the key substance of soil, and also an important indicator of soil functions (Bot & Benites, 2005). The OM content is influenced by natural factors including climate, topography, vegetation type, soil physical and chemical properties, and by human activities such as land use and management practices. These factors affect the spatial distribution pattern of OM to various degrees at different spatial scales.

Interest in accurate mapping soil OM is increased in regions requiring agricultural, ecological, and environmental assessment. For the purpose of improving mapping accuracy and reducing cost of soil sampling and laboratory analysis, a number of multivariate spatial statistical techniques, such as regression-kriging, co-kriging, kriging with external drift, and geographically weighted regression, are developed using easily accessible auxiliary variables including terrain indices, remote sensing data, and proximal sensing soil variables (Mueller and Pierce, 2003, Hengl et al., 2004, Terra et al., 2004, Tarr et al., 2005, Simbahan et al., 2006, Phachomphon et al., 2010, Li, 2010, Pei et al., 2010, Mishra et al., 2010, Bilgili et al., 2011, Zhang et al., 2011, Piccini et al., 2014). Studies show that the improvement of mapping accuracy is strongly influenced by the correlation strength of OM with auxiliary variables, and that which auxiliary variable provides the most accurate estimates must be evaluated across different environments (Hengl et al., 2004, Li, 2010).

The hilly red soil region of subtropical China has complex topography and varied land use, and has been under intense pressure for years due to the large rural population (Xi, 1990). The spatial variability of soil OM in this area is complex due to the natural process and human activities (Zhang, Shi, Yu, Wang, & Xu, 2010), and the spatial estimation using univariate geostatistics or multivariate geostatistics with environmental covariates may not be desirable without dense soil sampling and high analytical costs. In this case, we consider that an auxiliary soil variable measured more cheaply and easily than the target soil variable, such as visible and near-infrared (vis-NIR) diffuse reflectance, may be favourable for accurate mapping with lower cost.

The low-cost Vis-NIR diffuse reflectance needs little or no sample preparation for soil analysis. A literature review given by Stenberg, Viscarra Rossel, Mouazen, and Wetterlind (2010) shows that the technique is useful to measure soil water and mineral composition and to derive robust calibrations for OM and clay content. Soil vis-NIR spectral library has been developed at regional, national, and global level to help progress soil spectroscopy from an almost purely research tool to a more widely adopted and useful technique for soil monitoring, mapping and management (Shepherd and Walsh, 2002, Viscarra Rossel, 2009, Shi et al., 2014, Stevens et al., 2013). Soil spectral data based on laboratory or field measured reflectance may serve as covariates for improving the spatial estimation of soil properties. There are several studies that combined the use of soil vis-NIR spectroscopy and geostatistical technique. Bilgili et al. (2011) used the partial least square regression (PLSR) estimated values of soil properties from vis-NIR reflectance as co-variables to improve spatial prediction accuracy. Chen, Chang, and Liu (2013) predicted the spatial pattern of soil cadmium using spectral data at a selected wavelength as covariate. Conforti et al. (2015) developed a linear mixed effect model to estimate soil OM by combining the significant latent variables extracted from the PLSR analyses of vis–NIR data and the spatial association of the residuals from PLSR.

In those studies, concerning the selection of auxiliary variable, there are in principle two approaches: the use of the estimates by soil reflectance spectroscopy, or the use of the reflectance data at specific wavelength with high correlation to target variable. The former is likely to add the error of spectroscopy quantitative analysis model to the spatial estimation of target variable; the latter ignores the fact that soil vis–NIR spectra are largely non-specific, quite weak and broad due to overlapping absorptions of soil constituents and their often small concentrations in soil (Viscarra Rossel & Behrens, 2010). Therefore, here we attempted to use the principal components (PCs) of soil vis-NIR spectra as auxiliary variable of multivariate geostatistics, since principal component analysis (PCA) is an efficient method to remove multi-collinearity and reduce data redundancy in soil vis-NIR spectra. Vis-NIR spectroscopy is a logical technique to describe OM in soil, because the great majority of bonds that are absorbed in the near-infrared are organic, generally between C and O, N, or H (Malley, Martin, & Ben-Dor, 2004). It is difficult to assign specific absorption bands for OM, because it consists of mixtures of complex molecules and therefore shows overlapping of absorption bands (Mulders, 1987). In this study, the PC that showed good correlation with OM served as a covariate for the spatial estimation of soil OM.

Co-kriging is the extension of kriging in which correlated variables are used to improve the estimation or to compensate for missing data on some variables (Myers, 1984). Theoretically, when the target variable is under-sampled compared to a correlated auxiliary variable, by using the co-kriging method the spatial information of the auxiliary variable is transferred to the target through the cross-correlated information to minimise the estimation variances. Co-kriging is a preferred method in many practical situations where an auxiliary variable is collected more cheaply, quickly or easily than the target variable (Vauclin et al., 1983, Yates and Warrick, 1987, Atkinson et al., 1992, Tarr et al., 2005, Bilgili et al., 2011).

In this study, we hypothesised that a combination of the PCs of soil vis-NIR spectroscopy and multivariable geostatistics can lead to more accurate estimates of spatial variation of soil OM in a typical red soil hilly area at subtropical China. The ordinary co-kriging (CK) method was applied, and the PC of soil vis-NIR reflectance data which had the strongest correlation with OM was used as covariate. To simulate the situation of under-sampling the target variable when CK, we randomly sub-sampled 50% of the total soil samples. The ordinary kriging (OK) method using the target variable alone was also applied for comparison purpose.