Predicting and Mapping Soil Carbon Using Visible Near Infrared Spectroscopy at Different Scales

Fan Deng

Research output: Book/anthology/dissertation/reportPh.D. thesis


Soil is a huge pool of carbon, far exceeding that of vegetation or the atmosphere. Soil organic carbon (SOC) is a key parameter for soil quality and thus the soil’s capacity to produce crops. Therefore, the management of SOC stocks not only has an impact on global climate through the dynamic exchange of carbon between the soil and the atmosphere, but is also essential for ensuring global food security for a growing world population. Hence there is a growing demand for accurate, high resolution maps of SOC stocks, which in turn requires efficient and inexpensive methods for SOC data acquisition. This has stimulated the development of proximal soil sensing (PSS) technology and digital soil mapping (DSM). Visible-near infrared spectroscopy (Vis-NIR) is now widely seen as potential PSS methodology for SOC analysis, since it is remarkably versatile, robust, non-destructive, relatively easy to use, particularly useful for field applications and facilitates high sampling densities.
The general aim of this thesis, as part of the research project “Temporal and spatial dynamics of soil organic carbon in cultivated landscapes” funded by the Danish Council for Independent Research, Technology and Production Sciences (FTP), was to determine the potential of Vis-NIR for predicting SOC at field scale and national scale. Specifically, the first two objectives were to investigate the effects of different spectral data pretreatment and of soil moisture content on the calibration of Vis-NIR spectra to SOC content. The third objective was to test whether SOC calibration models built for different subdivisions of the Danish soil spectral library according to pedological or geological stratification would improve estimation of SOC content from Vis-NIR scans. The fourth objective was to explore the use of Vis-NIR for monitoring temporal changes in SOC in a wide range of soils in Denmark. The fifth objective was to investigate the use of Vis-NIR for estimating SOC distribution in soil profiles as the basis for mapping SOC in three dimensions at the field scale. The soils used in this work included samples from a national soil archive in Denmark and from field campaigns conducted during this study in Denmark and partly also in Germany. These soils covered most of the geological variability in Denmark. The spectrometer models LabSpec 5100, Veris and FieldSpec Pro were used to acquire Vis-NIR soil spectra. The reference SOC contents were measured by dry combustion with one of three instruments: LECO CN-1000 furnace, LECO CN-2000 or Thermo Flash 2000 Organic Element Analyzer. Principal component analysis (PCA), partial least squares regression (PLSR) and regression rules were the multivariate data analysis methods used in this thesis.
The first study investigated the effect of soil spectra pretreatment on our ability to build prediction models for SOC. To this end, 3000 air-dried soil samples previously collected in the 7-km national grid sampling scheme of Denmark were scanned with a LabSpec 5100 (range 350-2500 nm) using a Muglight as the light source. Following this, five types of spectra pretreatment techniques, namely multiplicative scatter correction (MSC), standard normal variate (SNV), detrending, moving average smoothing and Savitzky-Golay derivation, were used to remove noise from the spectra. Principal component analysis and PLSR were then applied for pattern recognition and building calibration models for each of the pretreatment techniques. The results of the validation process suggested that MSC preprocessing led to the best performing calibration model with the best predictive power among the five models. In general, these data pretreatment techniques can remove some of the noise in the spectra, but they also make the model more complicated to interpret as some useful information is removed too.
The second experiment dealt with the effect of soil moisture on SOC prediction by vis-NIR spectroscopy for a wide range of soil moisture contents. We selected 44 soil samples from Denmark and Germany with a gradient of clay and SOC contents. Samples were air-dried and sieved <2 mm. Eight different water potentials were established in repacked soil columns. These were: pF 1, pF 1.5, pF 2.5, and pF 3 by draining rewetted samples to the defined water potential; the remaining potentials either by saturating samples with water, air-drying, oven-drying or freeze-drying soils. Spectra were acquired with a LabSpec 5100 and a contact probe as light source. Calibration models for SOC prediction were built by PLSR for each of these water potentials. Water content had a substantial impact on the soil spectra, thus influencing SOC calibration. The best calibrations for SOC were obtained at moisture contents corresponding to pF 2.5 and freeze-drying. The good performance of the model at pF 2.5 indicated that the best in situ measurements for soil spectra may be obtained in spring and autumn, when soils are slightly drier than field capacity.
We assumed that the prediction capabilities of the Danish soil spectra library could be improved by dividing it into rather homogeneous subpopulations and building separate calibration models for each of these. Hence, the library was subdivided according to the categories mineral/organic soils, FAO soil classification, geo-region and parent material. In total, 2565 soil samples from 750 profiles of the national grid sampling scheme were used for this purpose. Multivariate data regression was carried out by applying regression rules in the analytical software Cubist, an extension to the R software environment. Each validation data set comprised 25% of the data within each subset. Both the mineral/organic and parent material subdivisions yielded better overall predictions than the general model for the whole sample population, and there was large variation in the predictive power of calibration models within each subdivision. Overall, the subdivision into mineral and organic soils resulted in the best model performance.
The fourth study tested the application of vis-NIR for monitoring SOC change. We used the Danish soil spectral library as well as agricultural soils collected by the national grid sampling scheme in both 1986 and 2009 at the depth intervals 0-25 cm and 25-50 cm. Air-dried and <2 mm sieved soils were scanned with a LabSpec 5100 using a Muglight. Regression rules were applied in Cubist for building SOC calibration models. In the topsoil there was no significance difference between the measured and model-predicted SOC in both years. However, there were significant differences in the subsoil. Empirical Bayesian kriging (EBK) was used to map SOC across Denmark, both for measured and model-predicted SOC. The spatial patterns of these were very similar, and the differences between measured and model-predicted SOC values were small. Based on these results, we concluded that Vis-NIR is sufficiently accurate for detecting changes in SOC in agricultural soils over a period of two decades.
The last experiment investigated the potential of using Vis-NIR for mapping the 3-dimensional distribution of SOC at the field scale. For this, 120 undisturbed soil columns (1 m depth) were collected from an arable field in central Jutland. The sampling design was based on conditioned Latin Hypercube sampling (cLHS), utilising auxiliary soil data including electricity conductivity, slope, elevation and curvature profile to capture the full variation in SOC in the field. Field-moist soil columns were scanned in the laboratory at 5 cm intervals. The Kennard-Stone algorithm was then applied to select the most representative Vis-NIR spectra based on the variance range of the score plot for identifying subsamples from the soil columns for reference SOC analysis by dry combustion.
The external parameter orthogonalisation (EPO) algorithm was able to correct for soil moisture effects in these soil cores, but did not improve the calibration of SOC. Interestingly, the prediction ability for SOC increased when the Danish spectral library was spiked with local samples from Vindum. This indicates that the full variation in Danish soils is not yet fully represented in the library. The 3-dimensional map generated from Vis-NIR predicted SOC for soils at 0-25, 25-30 and 95-100 cm depth and revealed the spatial variability of SOC in the field.

This thesis demonstrates the great potential of Vis-NIR spectroscopy as an efficient, inexpensive and accurate method for estimating SOC content and SOC changes over time. It includes the first systematic study of how a wide range of soil moisture contents affect the calibration of Vis-NIR spectra to SOC content. The innovative method developed involving subdivision of the soil spectral library allowed the library to be used more efficiently for both laboratory and field soil sensing. The first attempt to use Vis-NIR for monitoring SOC changes over a 20-year period is also presented. The efficient field sampling design and spectra selection method developed was successfully used for 3-dimensional SOC mapping based on Vis-NIR. However, further work is needed, especially with VerisProbe, to explore how soil structure affects soil spectra, and how soil moisture affects other soil properties. A better design for longer and shorter interval soil monitoring by Vis-NIR should also be explored, in order to give a better understanding of spatial and temporal changes in SOC. This would greatly facilitate 3-dimensional mapping of SOC and improve our understanding of SOC stocks in complex landscapes.
Original languageEnglish
PublisherAarhus Universitet, Institut for Agroøkologi
Number of pages163
ISBN (Print)978-87-92869-71-5
Publication statusPublished - 27 Mar 2013


Dive into the research topics of 'Predicting and Mapping Soil Carbon Using Visible Near Infrared Spectroscopy at Different Scales'. Together they form a unique fingerprint.

Cite this