Limits...
Prediction of Soil Organic Carbon at the European Scale by Visible and Near InfraRed Reflectance Spectroscopy.

Stevens A, Nocita M, Tóth G, Montanarella L, van Wesemael B - PLoS ONE (2013)

Bottom Line: The best calibrations achieved a root mean square error ranging from 4 to 15 g C kg(-1) for mineral soils and a root mean square error of 50 g C kg(-1) for organic soil materials.Model errors are shown to be related to the levels of soil organic carbon and variations in other soil properties such as sand and clay content.This study is a first step towards providing uniform continental-scale spectroscopic estimations of soil organic carbon, meeting an increasing demand for information on the state of the soil that can be used in biogeochemical models and the monitoring of soil degradation.

View Article: PubMed Central - PubMed

Affiliation: Georges Lemaître Centre for Earth and Climate Research, Earth and Life Institute, UCLouvain, Louvain-la-Neuve, Belgium.

ABSTRACT
Soil organic carbon is a key soil property related to soil fertility, aggregate stability and the exchange of CO2 with the atmosphere. Existing soil maps and inventories can rarely be used to monitor the state and evolution in soil organic carbon content due to their poor spatial resolution, lack of consistency and high updating costs. Visible and Near Infrared diffuse reflectance spectroscopy is an alternative method to provide cheap and high-density soil data. However, there are still some uncertainties on its capacity to produce reliable predictions for areas characterized by large soil diversity. Using a large-scale EU soil survey of about 20,000 samples and covering 23 countries, we assessed the performance of reflectance spectroscopy for the prediction of soil organic carbon content. The best calibrations achieved a root mean square error ranging from 4 to 15 g C kg(-1) for mineral soils and a root mean square error of 50 g C kg(-1) for organic soil materials. Model errors are shown to be related to the levels of soil organic carbon and variations in other soil properties such as sand and clay content. Although errors are ∼5 times larger than the reproducibility error of the laboratory method, reflectance spectroscopy provides unbiased predictions of the soil organic carbon content. Such estimates could be used for assessing the mean soil organic carbon content of large geographical entities or countries. This study is a first step towards providing uniform continental-scale spectroscopic estimations of soil organic carbon, meeting an increasing demand for information on the state of the soil that can be used in biogeochemical models and the monitoring of soil degradation.

No MeSH data available.


Related in: MedlinePlus

Relative Root Mean Square Error of Prediction (RMSEP) per land cover, for arbitrary classes of SOC and sand content.The sand classes are 0–25%, 25–50%, 50–75%, 75–100% and the SOC classes are 0–25 g C kg−1, 25–50 g C kg−1, 50–200 g C kg−1. The relative RMSEP is the RMSEP divided by the mean of observed SOC values of models developed with (red bars) and without auxiliary predictors (blue bars). Each panel regroups mineral samples of a given SOC interval and land cover type. The number of training samples (n) for each class of SOC content is given in each panel.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC3686688&req=5

pone-0066409-g006: Relative Root Mean Square Error of Prediction (RMSEP) per land cover, for arbitrary classes of SOC and sand content.The sand classes are 0–25%, 25–50%, 50–75%, 75–100% and the SOC classes are 0–25 g C kg−1, 25–50 g C kg−1, 50–200 g C kg−1. The relative RMSEP is the RMSEP divided by the mean of observed SOC values of models developed with (red bars) and without auxiliary predictors (blue bars). Each panel regroups mineral samples of a given SOC interval and land cover type. The number of training samples (n) for each class of SOC content is given in each panel.

Mentions: The differences in spectral response observed in Figure 4–5 had logically a strong impact on model errors. To illustrate this, we computed the relative RMSEP for mineral soil models for intervals of SOC and sand content. The relative RMSEP is the RMSEP divided by the mean of the observed SOC content in a given class. For models using the spectra only for prediction, the relative RMSEP of the models was stable across the SOC content classes but it increased with the sand content (Figure 6). This confirms the results of other studies [48], [54] that found larger SOC prediction errors for soils with the highest sand contents. The effect of sand content on SOC prediction accuracy was more pronounced at low SOC content due to the relatively low absorption rates of organic matter and the masking from other soil components ([55]; Figure 4–5). It is therefore expected that spectral libraries of soils characterized by a low SOC content will perform poorly when samples have large variations in particle size distribution. It can be also observed that the use of sand content as auxiliary predictor drastically improved model predictions for sandy soils (Figure 6), explaining the increase in model accuracies compared to models based on the spectral matrix only (Table 4).


Prediction of Soil Organic Carbon at the European Scale by Visible and Near InfraRed Reflectance Spectroscopy.

Stevens A, Nocita M, Tóth G, Montanarella L, van Wesemael B - PLoS ONE (2013)

Relative Root Mean Square Error of Prediction (RMSEP) per land cover, for arbitrary classes of SOC and sand content.The sand classes are 0–25%, 25–50%, 50–75%, 75–100% and the SOC classes are 0–25 g C kg−1, 25–50 g C kg−1, 50–200 g C kg−1. The relative RMSEP is the RMSEP divided by the mean of observed SOC values of models developed with (red bars) and without auxiliary predictors (blue bars). Each panel regroups mineral samples of a given SOC interval and land cover type. The number of training samples (n) for each class of SOC content is given in each panel.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC3686688&req=5

pone-0066409-g006: Relative Root Mean Square Error of Prediction (RMSEP) per land cover, for arbitrary classes of SOC and sand content.The sand classes are 0–25%, 25–50%, 50–75%, 75–100% and the SOC classes are 0–25 g C kg−1, 25–50 g C kg−1, 50–200 g C kg−1. The relative RMSEP is the RMSEP divided by the mean of observed SOC values of models developed with (red bars) and without auxiliary predictors (blue bars). Each panel regroups mineral samples of a given SOC interval and land cover type. The number of training samples (n) for each class of SOC content is given in each panel.
Mentions: The differences in spectral response observed in Figure 4–5 had logically a strong impact on model errors. To illustrate this, we computed the relative RMSEP for mineral soil models for intervals of SOC and sand content. The relative RMSEP is the RMSEP divided by the mean of the observed SOC content in a given class. For models using the spectra only for prediction, the relative RMSEP of the models was stable across the SOC content classes but it increased with the sand content (Figure 6). This confirms the results of other studies [48], [54] that found larger SOC prediction errors for soils with the highest sand contents. The effect of sand content on SOC prediction accuracy was more pronounced at low SOC content due to the relatively low absorption rates of organic matter and the masking from other soil components ([55]; Figure 4–5). It is therefore expected that spectral libraries of soils characterized by a low SOC content will perform poorly when samples have large variations in particle size distribution. It can be also observed that the use of sand content as auxiliary predictor drastically improved model predictions for sandy soils (Figure 6), explaining the increase in model accuracies compared to models based on the spectral matrix only (Table 4).

Bottom Line: The best calibrations achieved a root mean square error ranging from 4 to 15 g C kg(-1) for mineral soils and a root mean square error of 50 g C kg(-1) for organic soil materials.Model errors are shown to be related to the levels of soil organic carbon and variations in other soil properties such as sand and clay content.This study is a first step towards providing uniform continental-scale spectroscopic estimations of soil organic carbon, meeting an increasing demand for information on the state of the soil that can be used in biogeochemical models and the monitoring of soil degradation.

View Article: PubMed Central - PubMed

Affiliation: Georges Lemaître Centre for Earth and Climate Research, Earth and Life Institute, UCLouvain, Louvain-la-Neuve, Belgium.

ABSTRACT
Soil organic carbon is a key soil property related to soil fertility, aggregate stability and the exchange of CO2 with the atmosphere. Existing soil maps and inventories can rarely be used to monitor the state and evolution in soil organic carbon content due to their poor spatial resolution, lack of consistency and high updating costs. Visible and Near Infrared diffuse reflectance spectroscopy is an alternative method to provide cheap and high-density soil data. However, there are still some uncertainties on its capacity to produce reliable predictions for areas characterized by large soil diversity. Using a large-scale EU soil survey of about 20,000 samples and covering 23 countries, we assessed the performance of reflectance spectroscopy for the prediction of soil organic carbon content. The best calibrations achieved a root mean square error ranging from 4 to 15 g C kg(-1) for mineral soils and a root mean square error of 50 g C kg(-1) for organic soil materials. Model errors are shown to be related to the levels of soil organic carbon and variations in other soil properties such as sand and clay content. Although errors are ∼5 times larger than the reproducibility error of the laboratory method, reflectance spectroscopy provides unbiased predictions of the soil organic carbon content. Such estimates could be used for assessing the mean soil organic carbon content of large geographical entities or countries. This study is a first step towards providing uniform continental-scale spectroscopic estimations of soil organic carbon, meeting an increasing demand for information on the state of the soil that can be used in biogeochemical models and the monitoring of soil degradation.

No MeSH data available.


Related in: MedlinePlus