Elsevier

Geoderma

Volume 405, 1 January 2022, 115407
Geoderma

The effectiveness of digital soil mapping with temporal variables in modeling soil organic carbon changes

https://doi.org/10.1016/j.geoderma.2021.115407Get rights and content

Highlights

  • We identified temporal patterns of environmental factors used to map SOC change.

  • Spatiotemporal variations in SOC along the gradients of temporal covariates were similar across space and time.

  • Model projections with temporal covariates resulted in more accurate estimates.

Abstract

The effectiveness of using changes in environmental conditions to explain the spatiotemporal variability in soil organic carbon (SOC) with digital soil mapping (DSM) requires investigation. In this study, temporal variables representing temporal patterns of climate, vegetation, and land cover factors were explored. Models to predict SOC stocks were developed using a random forest algorithm and data from China during two periods (the 1980s and 2010s). We forecasted and hindcasted the developed models and assessed their temporal projections against temporally independent data. Models were developed for both periods using different sets of variables (with/without temporal variables), and their temporal projections were compared. The important temporal variables were identified by applying the recursive feature elimination algorithm. The results showed that the performances of temporal projections for the 1980s and 2010s were improved by approximately 17% and 47%, respectively, when temporal variables were included in the models. Spatially, the maps of changes in SOC stocks derived from the models that included temporal variables presented stronger associations with temporal changes in climate, vegetation, and land cover than those derived from the models that did not include temporal variables. This work highlights that variation in SOC stocks can be linked to temporal patterns of environmental factors. The findings also provide evidence that the application of temporal patterns of environmental factors to DSM models can be useful for the large-scale prediction of changes in SOC.

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)

  • J.T. Abatzoglou et al.

    TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015

    Sci. Data

    (2018)
  • Adair, E.C., Parton, W.J., Del Grosso, S.J., Silver, W.L., Harmon, M.E., Hall, S. A., et al., 2008. Simple three‐pool...
  • N.H. Batjes et al.

    Standardised soil profile data to support global mapping and modelling (WoSIS snapshot 2019)

    Earth Syst. Sci. Data

    (2020)
  • L. Breiman

    Random forests

    Mach. Learn.

    (2001)
  • Chen, P.F., 2019. Monthly NPP 1 km raster dataset of China’s terrestrial ecosystems (1985-2015). Global Change Data...
  • T.W. Crowther et al.

    Quantifying global soil carbon losses in response to warming

    Nature

    (2016)
  • J.M. Gray et al.

    Change in soil organic carbon stocks under 12 climate change projections over New South Wales, Australia

    Soil Sci. Soc. Am. J.

    (2016)
  • Greifer, N., 2019. WeightIt: Weighting for Variable Balance in Observational Studies. Available at:...
  • Greifer, N., 2020. Variable Balance Tables and Plots. Available at:...
  • I. Guyon et al.

    Gene selection for cancer classification using support vector machines

    Mach. Learn.

    (2002)
  • Cited by (24)

    View all citing articles on Scopus
    View full text