The effectiveness of digital soil mapping with temporal variables in modeling soil organic carbon changes
Introduction
Soil is the largest terrestrial carbon pool and has the potential to offset the greenhouse effect. Many recent studies have focused on soil organic carbon (SOC) changes in order to provide accurate estimations of SOC and to develop mechanistic understandings of the impacts of environmental changes on SOC (Luo et al., 2017, Riggers et al., 2019). Research in support of these goals has led to the development of digital soil mapping (DSM) approaches for predicting SOC changes in time and space under climate and land cover change scenarios. These studies have projected the current predictions of SOC stocks into the future and past at various scales (e.g., Gray and Bishop, 2016, Yigini and Panagos, 2016, Sanderman et al., 2017, Adhikari et al., 2019, Huang et al., 2019). And have contributed to the expansion of the DSM approach to predicting SOC changes over time. DSM approaches have been developed by relating SOC measurements to environmental variables using statistical functions (Minasny et al., 2013). These approaches assume that the environmental drivers of the spatial gradients of SOC remain constant over time. Based on this assumption, the temporal variation in SOC can be modeled by applying a DSM model derived from a particular time to fit the variables retrieved from a different period. This approach allows us apply links between SOC changes and DSM more broadly.
Despite the valuable knowledge gained from these pioneering efforts, the development of predictions for novel periods presents a substantial challenge to the DSM approach because it is difficult to introduce models to temporal domains of SOC change that differ from the spatial domains in which the models were fitted. According to the DSM framework (McBratney et al., 2003), SOC can be theoretically predicted from environmental factors and soil conditions. When the authors of previous studies (e.g., Gray and Bishop, 2016, Yigini and Panagos, 2016, Adhikari et al., 2019) applied the DSM framework to predict SOC change, certain environmental variables were assumed to be relatively static (e.g., topography and soils) during the study periods, whereas certain variables were considered dynamic (e.g., climate, vegetation, and land cover). The updating of dynamic variables in DSM models highlights the potential to make predictions based on environmental change, but it is still difficult to capture the impacts of environmental controls on SOC change with these models.
Many studies have shown that the spatial variability of SOC can be explained by environmental variables (e.g., Luo et al., 2017, Gray et al., 2019, Viscarra Rossel et al., 2019). For example, soil carbon levels were shown to decrease with temperature and increase with precipitation in China’s terrestrial ecosystems (Tang et al., 2018); these authors also found that high species richness and high biomass inputs enhanced soil carbon stocks and that soil carbon is sensitive to human activities. At temporal scales, changes in these environmental factors also have important influences on soil carbon. Rising temperatures lead to the loss of soil carbon (Crowther et al., 2016). Human activities, such as farming and afforestation, have significantly influenced the accumulation of soil carbon (Piao et al., 2009, Sanderman et al., 2017). These studies provide evidence that the environmental variables connected to the spatial gradients of soil carbon may be related to the drivers of temporal changes in soil carbon stocks.
Here, the basic assumptions of DSM approaches used to predict SOC changes were considered in order to obtain environmental variables, including static variables (topography and soil factors) and dynamic variables (land cover, vegetation, and climate). Additionally, new variables, namely, temporal variables, that represent the temporal characteristics of the dynamic variables were calculated. We hypothesized that a DSM model that included temporal variables would provide better results than the same model without these variables. Environmental variables are critical factors that determine the spatial variability of SOC, and these variations exist within a system of environmental factors that change and interact in both space and time. Consequently, the variability of SOC is sensitive not only to environmental conditions but also to changes in environmental conditions.
To test the hypothesis of this study, observations of SOC stocks in China at a depth of 0–20 cm were collected from two time periods (the 1980s and 2010s). Prior to analysis, temporal variables were calculated by assessing changes in climate, vegetation, and land cover. Furthermore, a covariate balance analysis was conducted to assess and subsequently diminish the potential sampling bias of these variables between the two periods. The spatial variation in SOC stocks was then related to temporal variables. Models were developed using topographic and soil variables as static variables; climate, vegetation and land cover as dynamic variables; and temporal patterns of climate, vegetation and land cover as temporal variables. We forecasted historical models to recent data and hindcasted recent models to historical data over thirty years by changing the dynamic variables. In both cases, the model performance was evaluated using temporally independent data from a separate period. The performance of these models was further compared to the performance of commonly used DSM models that contain only static and dynamic variables.
Section snippets
Soil organic carbon data
Soil data collected from 1648 plots across China were retrieved from the World Soil Information Service database (WoSIS, Batjes et al., 2020). This legacy soil dataset was produced mainly during China’s second national soil survey project in the early 1980s. This dataset includes in-depth information on the SOC contents (g kg−1) of different pedological horizons. The SOC content was measured using the Walkley-Black method (Shi and Song, 2016). There were only 148 sites with bulk density (BD)
Spatial prediction of SOC stocks
To overcome multicollinearity among predictors, RFE was used to screen for the independent variables to be used in the models. The feature selection results are shown in Table 1. Of the fifteen potential temporal variables, eight and fourteen temporal variables were selected variables for use in Mod-SDT-1980 and Mod-SDT-2010, respectively.
The mean CCC values of the internal validations of the SOC stocks predictions were 0.54 and 0.56 for Mod-SD-1980 and Mod-SDT-1980, respectively (Fig. 5).
The effectiveness of temporal projection
The results showed that the external validation metrics for the model performance were generally worse than the internal validation metrics. It suggests that model evaluation based solely on internal validation may overestimate the performance of DSM models in predicting SOC stocks. The deterioration of model performance when predicting temporally independent data (as indicated by the external evaluations of the forecasts and hindcasts) may have been due to discrepancies in the influence of
Conclusions
The application of DSM to predict changes in SOC stocks is simple and could play a critical role in predicting likely changes in soil carbon under changing environmental conditions. The work presented in this study demonstrates that the assessment of model performance based on temporally independent data is likely to be poorer than the assessment of model performance using independent validation data from the corresponding time. Relatedly, future work must use appropriate methods to assess the
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This research was supported by the National Key Research and Development Program of China (No. 2017YFB0504204), the National Natural Science Foundation of China (No. 42171054), the One Hundred Talents Program of the Chinese Academy of Science (2015, No. Y674141001), the Dragon 5 Program (No. 59197), and the Priority Academic Program Development of Jiangsu Higher Education Institutions.
References (40)
- et al.
Assessing soil organic carbon stock of Wisconsin, USA and its fate under future land use and climate change
Sci. Total Environ.
(2019) - et al.
Modeling soil attribute depth functions with equal-area quadratic smoothing splines
Geoderma
(1999) - et al.
Driving factors of soil organic carbon fractions over New South Wales, Australia
Geoderma
(2019) - et al.
High-resolution and three-dimensional mapping of soil texture of China
Geoderma
(2020) - et al.
On digital soil mapping
Geoderma
(2003) - et al.
Digital mapping of soil carbon
Adv. Agron.
(2013) - et al.
Multi-model ensemble improved the prediction of trends in soil organic carbon stocks in German croplands
Geoderma
(2019) - et al.
Impacts of urbanization on soil organic carbon stocks in the northeast coastal agricultural areas of china
Sci. Total Environ.
(2020) - et al.
Assessment of soil organic carbon stocks under future climate and land cover changes in Europe
Sci. Total Environ.
(2016) - et al.
Modeling soil organic carbon change in croplands of china, 1980–2009
Global Planet. Change
(2012)
TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015
Sci. Data
Standardised soil profile data to support global mapping and modelling (WoSIS snapshot 2019)
Earth Syst. Sci. Data
Random forests
Mach. Learn.
Quantifying global soil carbon losses in response to warming
Nature
Change in soil organic carbon stocks under 12 climate change projections over New South Wales, Australia
Soil Sci. Soc. Am. J.
Gene selection for cancer classification using support vector machines
Mach. Learn.
Cited by (24)
A fine digital soil mapping by integrating remote sensing-based process model and deep learning method in Northeast China
2024, Soil and Tillage ResearchDynamic modeling of topsoil organic carbon and its scenarios forecast in global Mollisols regions
2023, Journal of Cleaner Production