Elsevier

CATENA

Volume 182, November 2019, 104141
CATENA

Exploring the driving forces and digital mapping of soil organic carbon using remote sensing and soil texture

https://doi.org/10.1016/j.catena.2019.104141Get rights and content

Highlights

  • Remote sensing covariates and field data were integrated for SOC mapping.

  • A Gradient Boosting Machine (GBM) was calibrated for SOC mapping.

  • The results showed that EVI and sand were the most influential factors.

  • 45.2% of SOC variation was estimated by remote sensing covariates.

Abstract

The main objectives of this paper were 1) to estimate soil organic carbon (SOC) using remote sensing covariates, soil properties, and topographic factors , and 2) to evaluate the interaction and the relative influence of the selected factors on the spatial variation of SOC. Thirteen factors were considered for digital mapping of SOC in the west Urmia Lake in Iran. To quantify multicollinearity among the predictor variables, Variance Inflation Factor (VIF) was calculated. Among them, nine independent factors were remained including silt, sand, slope, enhanced vegetation index (EVI), brightness, wetness, land cover, and latitude and longitude. A machine learning algorithm called Gradient Boosting Machine (GBM) was calibrated for understanding the spatial dynamic and prediction of SOC. Model performance showed that GBM explained 43.5% (R2) of the SOC variation, and root mean square error (RMSE) was 0.23%. Results showed that EVI and sand were the most influential factors of the SOC variation while slope and land cover were the least important ones. Furthermore, significant interaction among EVI-wetness-SOC and EVI-sand-SOC was detected. On the other hand, 45.2% of SOC variation was estimated by remote sensing covariates. These results suggested that GBM was a promising approach for an in-depth understanding of the SOC variation over space.

Introduction

Soil organic carbon (SOC) is one of the major indicators of soil quality and productivity (Herrick and Wander, 1997). The presence of SOC improves the physicochemical properties of soils because it facilitates the maintenance of nutrients, increases the water infiltration rate and controls soil resilience against environmental degradation through improved soil aggregation, porosity and structure (Jegajeevagan et al., 2013; Parras-Alcantara et al., 2016). SOC also plays an important role in regulating the global carbon cycle and controlling the emission of greenhouse gases (Bradford et al., 2016; Filippi et al., 2016; Lal, 2004). It is more important in arid and semi-arid regions where lack of water is a critical problem and soil particles can easily be moved by wind owing to the absence of SOC (Bruun et al., 2015; Saia et al., 2014).

Urmia Plain, situated in the west of Urmia Lake, is one of the biggest plains in Iran and an important agricultural hub in the region. In the last few decades, this region has experienced several episodes of environmental degradation as exemplified by drying up of Urmia Lake. Therefore, maintaining agricultural production yields at optimum level with the least water usage and minimum damage to the environment has become a crucial issue. To achieve this, the lands must be kept in the best condition. This is not possible unless SOC, which can compensate for the unnecessary use of chemical fertilizers leading to soil salinization and environmental pollution, is taken seriously. Thus, owing to the limited amount of suitable land for the food production, mapping and monitoring SOC content is vital from both agricultural and environmental perspectives (Dono et al., 2016; Novara et al., 2017).

Geostatistical methods have been one of the common ways for spatial estimation of SOC. Kriging, known as the best linear unbiased predictor, has been upgraded in different forms and successfully calibrated for SOC prediction worldwide (Dai et al., 2014; Liu et al., 2017; Mirzaee et al., 2016; Mishra et al., 2017; Zeng et al., 2016). Such methods have shown superlative performance where considerable SOC measurements are available (Hoffmann et al., 2014; Piccini et al., 2014). However, field sampling and laboratory measurements are expensive and time consuming (Miklos et al., 2010; Mulder et al., 2011). In addition, SOC shows high degrees of variability in response to soil factors that ignoring them reduces the accuracy of the SOC predictions (Allen et al., 2010; Jandl et al., 2013; Viaud et al., 2010). According to the literature, topographic data, land cover, soil structure and texture, available water capacity, parent material, cation exchange capacity and soil type have been used as environmental auxiliaries in SOC mapping (Aksoy et al., 2016; Rial et al., 2017; Schillaci et al., 2017; Sindayihebura et al., 2017; Yigini and Panagos, 2016). Covariates such as topographic factors, the spectral behavior of soil properties, vegetation condition and soil moisture can mostly be extracted from remote sensing (RS) data. RS technology is one of the best alternatives for gaining quick and cost-effective information regarding soil properties while covering wide areas including inaccessible locations. The correlation between soil properties (for instance organic matter and moisture content) and RS covariates has been reported by the previous studies notably those on soil salinity (Bouaziz et al., 2011; Rahmati and Hamzehpour, 2016; Wang and Xu, 2008), soil erosion and runoff (Li et al., 2009; King et al., 2005), soil surface temperature (Mao et al., 2005), and soil nutrients (Xu et al., 2017, Xu et al., 2018). However, the relationship between the outlined covariates and soil properties is often complex.

Parallel to the advancements in RS technology, progress in machine learning (ML) models such as Gradient Boosting Machine (GBM, also known as Boosted Regression Tress (BRT)), Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Networks (ANNs) has greatly improved the estimation of soil parameters through a procedure known as digital soil mapping (DSM; McBratney et al., 2003; Mulder et al., 2011; Sindayihebura et al., 2017; Yang et al., 2016; G. Zhang et al., 2017; Keskin et al., 2019). Guo et al. 2015) integrated RF and residual kriging to estimate SOC at regional scale where their proposed framework resulted in a better performance than stepwise linear regression that was used as a benchmark. Schillaci et al. (2017) used the BRT together with soil properties, topographic and RS data to map SOC at a semi-arid Mediterranean region. Wang et al. (2018) used a set of ML techniques to map SOC in the semi-arid rangelands of eastern Australia. Their findings showed higher performance of RF and BRT models than SVM. Keskin et al. (2019) reviewed eight ML and statistical techniques for digital mapping of soil carbon and found that RF, SVM and BRT yielded the highest accuracy. Nevertheless, previous applications of ML techniques have focused on evaluating the models based on their accuracy while the relative influence and interaction effect of the covariates have been less investigated.

In attempt to model the SOC variation, contribution of this paper is to profoundly understand how the SOC changes in relation to the variation of its underlying factors. Hence, the aim of this research was twofold besides considering extra variables mined from RS imagery. Firstly, to evaluate the relative influence and interaction of predictors such as soil properties, RS covariates and topographic factors on the spatial variation of SOC. Secondly, to predict SOC for the unsampled locations using GBM and creating their associated uncertainty map.

Section snippets

Study area

Study area covers 900 km2 of the Urmia Plain, west of Lake Urmia in northwestern Iran (Fig. 1) which extends between 45° 13′ to 45° 55′ E and 37° 20′ to 37° 53′ N. The dominant land cover types include forest, grassland and cultivated land. The mean annual precipitation rate is approximately 367 mm while the mean annual temperature for the coldest and warmest months are −5.2 °C and 32 °C, respectively. Potential evaporation in the area is approximately 900–1170 mm. Elevation above sea level in

Feature selection

As a result of collinearity and multicollinearity examination, NDVI, greenness, clay, and elevation were removed and the remaining factors were taken into consideration. Fig. 2 shows the correlation plot between the remaining factors including silt, sand, slope, EVI, brightness, wetness, and latitude and longitude. Since the land cover is a categorical predictor, it is not shown on Fig. 2. Results from the VIF exploration measurement indicated that VIF was below 6 for all the selected factors.

Enhanced vegetation index

GBM can be used to generate the partial response curve, a graph showing the relationship between the variability of SOC and each of the explanatory factors. As shown in Fig. 4, EVI with 35.5%, was the most important factor influencing SOC variation. EVI is not only chlorophyll-sensitive, indicating the greenness of the plant canopy, but also pertinent to the structural variations of the canopy, for example plant physiognomy and canopy architecture (Huete et al., 2002). The response curve of SOC

Conclusion

Along with advances in RS technology, DSM (in particular SOC mapping) has gained an unprecedented importance. Time and cost savings and widespread coverage of satellite imageries are among the advantages that highlight the role of RS in the field of soil sciences. On the other hand, calibrating well-established ML models (i.e., GBM and RF), increase the precision of SOC mapping and boost our understanding regarding the underlying drivers of SOC variation. In this paper, a combination of soil

Acknowledgments

This research was granted by Tarbiat Modares University, Grant Number IG-39713.

References (85)

  • A. Huete et al.

    Overview of the radiometric and biophysical performance of the MODIS vegetation indices

    Remote Sens. Environ.

    (2002)
  • H. Keskin et al.

    Digital mapping of soil carbon fractions with machine learning

    Geoderma

    (2019)
  • C. King et al.

    Remote-sensing data as an alternative input for the STREAM runoff model

    Catena

    (2005)
  • H. Li et al.

    Predicting runoff in ungauged catchments by using Xinanjiang model with MODIS leaf area index

    J. Hydrol.

    (2009)
  • D. Liu et al.

    Spatial distribution of soil organic carbon and analysis of related factors in croplands of the black soil region, Northeast China

    Agric. Ecosyst. Environ.

    (2006)
  • A.B. McBratney et al.

    On digital soil mapping

    Geoderma

    (2003)
  • J. Meersmans et al.

    A multiple regression approach to assess the spatial distribution of soil organic carbon (SOC) at the regional scale (Flanders, Belgium)

    Geoderma

    (2008)
  • S. Mirzaee et al.

    Spatial variability of soil organic matter using remote sensing data

    Catena

    (2016)
  • U. Mishra et al.

    Spatial representation of organic carbon and active-layer thickness of high latitude soils in CMIP5 earth system models

    Geoderma

    (2017)
  • V.L. Mulder et al.

    The use of remote sensing in soil and terrain mapping—a review

    Geoderma

    (2011)
  • A. Novara et al.

    Agricultural land abandonment in Mediterranean environment provides ecosystem services via soil carbon sequestration

    Sci. Total Environ.

    (2017)
  • L. Parras-Alcantara et al.

    Long-term effects of soil management on ecosystem services and soil loss estimation in olive grove top soils

    Sci. Total Environ.

    (2016)
  • B.T. Pham et al.

    Landslide susceptibility modeling using reduced error pruning trees and different ensemble techniques: hybrid machine learning approaches

    Catena

    (2019)
  • Ch. Piccini et al.

    Estimation of soil organic matter by geostatistical methods: use of auxiliary information in agricultural and environmental assessment

    Ecol. Indic.

    (2014)
  • M. Rial et al.

    Understanding the spatial distribution of factors controlling topsoil organic carbon content in European soils

    Sci. Total Environ.

    (2017)
  • C. Schillaci et al.

    Spatio-temporal topsoil organic carbon mapping of a semi-arid Mediterranean region: the role of land use, soil texture, topographic indices and the influence of the remote sensing data to modeling

    Sci. Total Environ.

    (2017)
  • H. Shafizadeh-Moghadam et al.

    Novel forecasting approaches using combination of machine learning and statistical models for flood susceptibility mapping

    Journal of environmental management

    (2018)
  • A. Sindayihebura et al.

    Comparing digital soil mapping techniques for organic carbon and clay content: case study in Burundi's central plateaus

    Catena

    (2017)
  • H. Shafizadeh-Moghadam

    Improving spatial accuracy of urban growth simulation models using ensemble forecasting approaches. Computers

    Environment and Urban Systems

    (2019)
  • F. Stevens et al.

    Detecting and quantifying field-related spatial variation of soil organic carbon using mixed-effect models and airborne imagery

    Geoderma

    (2015)
  • R. Taghizadeh-Mehrjardi et al.

    Digital mapping of soil organic carbon at multiple depths using different data mining techniques in Baneh region, Iran

    Geoderma

    (2016)
  • F. Wang et al.

    Development and application of a remote sensing-based salinity prediction model for a large estuarine lake in the US Gulf of Mexico coast

    J. Hydr.

    (2008)
  • B. Wang et al.

    High resolution mapping of soil organic carbon stocks using remote sensing variables in the semi-arid rangelands of eastern Australia

    Sci. Total Environ.

    (2018)
  • Y. Xu et al.

    Evaluating the effect of remote sensing image spatial resolution on soil exchangeable potassium prediction models in smallholder farm settings

    J. Environ. Manag.

    (2017)
  • Y. Xu et al.

    Estimating soil total nitrogen in smallholder farm setting using remote sensing spectral indices and regression kriging

    Catena

    (2018)
  • R.M. Yang et al.

    Comparison of boosted regression tree and random forest models for mapping topsoil organic carbon concentration in an alpine ecosystem

    Ecol. Indic.

    (2016)
  • Y. Yigini et al.

    Assessment of soil organic carbon stocks under future climate and land cover changes in Europe

    Sci. Total Environ.

    (2016)
  • C. Zeng et al.

    Mapping soil organic matter concentration at different scales using a mixed geographically weighted regression method

    Geoderma

    (2016)
  • G. Zhang et al.

    Recent progress and future prospect of digital soil mapping: a review

    J. Integr. Agric.

    (2017)
  • H. Zhang et al.

    Prediction of soil organic carbon in an intensively managed reclamation zone of eastern China: a comparison of multiple linear regressions and the random forest model

    Sci. Total Environ.

    (2017)
  • E. Aksoy et al.

    Combining soil databases for topsoil organic carbon mapping in Europe

    PLoS One

    (2016)
  • D.E. Allen et al.

    A review of sampling designs for the measurement of soil organic carbon in Australian grazing lands

    Rangel. J.

    (2010)
  • Cited by (65)

    View all citing articles on Scopus
    View full text