Visible-near infrared reflectance spectroscopy for assessment of soil properties in a semi-arid area of Turkey

https://doi.org/10.1016/j.jaridenv.2009.08.011Get rights and content

Abstract

Reflectance spectroscopy can be used to nondestructively characterize materials for a wide range of applications. In this study, visible-near infrared reflectance spectroscopy (VNIR) was evaluated for prediction of diverse soil properties related to four different soil series of the Entisol soil group within a single field in northern Turkey. Soil samples were collected from 512 locations in a 25 × 25 m sampling grid over a 32 ha (800 × 400 m) area. Air-dried soil samples were scanned at 1 nm resolution from 350 to 2500 nm, and calibrations between soil physical and chemical properties and reflectance spectra were developed using cross-validation under partial least squares regression (PLSR) and multivariate adaptive regression splines (MARS). Raw reflectance and first derivative reflectance data were used separately and combined for all samples in the data set. Data were additionally divided into two random subsets of 70 and 30% of the full data, which were each used for calibration and validation. Overall, MARS provided better predictions when under cross-validation. However, PLSR and MARS results were comparable in terms of prediction accuracy when using separate data sets for calibration and validation. No improvement was obtained by combining first derivative and raw data. Strongest correlations were obtained with exchangeable Ca and Mg, cation exchange capacity, and organic matter, clay, sand, and CaCO3 contents. When soil data were classified into groups, VNIR spectroscopy estimated class memberships well, especially for soil texture. In conclusion, VNIR spectroscopy was variably successful in estimating soil properties at the field scale, and showed potential for substituting laboratory analyses or providing inexpensive co-variable data.

Introduction

Soil characteristics often exhibit high spatial variability, even across single agricultural fields. Mapping soil fertility indicators and quantifying soil parameters that control the fate of chemicals are therefore important for site-specific soil management and protection of the environment. Typically, large numbers of samples must be collected and analyzed in order to capture this spatial variability and adequately estimate soil properties. Conventional methods may be expensive and require large amounts of labor and chemicals for performing these tasks (Viscarra Rossel and McBratney, 1998a, Viscarra Rossel and McBratney, 1998b). Visible and near infrared reflectance (VNIR) spectroscopy shows promise as a low-cost method that can be used to substitute or complement traditional soil characterization methods. Once calibrated, it can be used to predict multiple soil characteristics simultaneously and explain within-field spatial variability (Bowers and Hanks, 1965, Ben-Dor and Banin, 1995, Chang et al., 2001, Islam et al., 2003, Shepherd and Walsh, 2002).

The scale of application for this method can affect its utility. Shepherd and Walsh (2002) and Brown et al. (2006) used VNIR spectroscopy based on samples from many soil types over large geographical areas. This type of sampling generally provides a wide range of soil indicator values, which promotes good regressive results. On the other hand, wide distribution of soils in a sample set challenges the methodology by requiring greater universality in the statistical prediction relations, especially when including different parent materials (Reeves and Van Kessel, 1999, Shepherd and Walsh, 2002). In precision agriculture, the interest is often limited to characterizing single or multiple fields within a relatively small geographical area. This poses different challenges in that the data ranges for soil properties are generally small, but the prediction equation may only need to be locally applicable.

With VNIR, soil reflectance characteristics are determined over the entire visible (350–700 nm) and near infrared (700–2500 nm) region with the use of a monochromator. Raw data, first-, and second-derivatives each provide valuable information that can be analyzed separately or combined using multivariate statistical methods or data mining techniques. Soil constituents have unique absorption features in these wavelength regions due to overtones related to stretching and bending vibrations in molecular bonds such as C–C, C–H, N–H and O–H (Dalal and Henry, 1986).

Chang et al. (2001) predicted more than thirty soil properties simultaneously with variable levels of success using a principal component analysis method with cross-validation. They reported successful predictions (R2 > 0.80) for total organic carbon and nitrogen (g kg−1), gravimetric soil water content, soil water content at −1.5 Mpa, exchangeable calcium, cation exchange capacity (CEC) and silt and sand content. Brown et al. (2006) used over 4100 surface and subsurface soils from across the United States, Africa and Asia to evaluate the accuracy of VNIR empirical models for global soil characterization and reported strong predictability for kaolinite, montmorillonite, clay content, as well as CEC, soil organic carbon, inorganic carbon, and extractable Fe.

Others also used VNIR spectroscopy to successfully predict organic carbon and nitrogen (Reeves et al., 2002), Fe2O3, Al2O3, CaCO3 (Ben-Dor and Banin, 1995), potentially mineralizable nitrogen (Morón and Cozzolino, 2002, Reeves and Van Kessel, 1999), heavy metals, micronutrients (Cozzolino and Morón, 2003, Kooistra et al., 2001, Udelhoven et al., 2003), C:N ratio and soil biological properties (Chodak et al., 2001, Ludwig et al., 2002). Additionally, the prediction of soil constituents that do not absorb within the VNIR range may be possible through their correlations with spectrally active constituents (Ben-Dor and Banin, 1995).

Different mathematical pre-processing techniques have been applied to either raw reflectance spectra or soil absorbance spectra (log 1/reflectance) to remove noise within spectra originating from effects of illumination or non-homogeneous distributions of particle sizes. The most commonly used are first- and second-derivatives with or without smoothing, which have provided optimum predictions in some studies (Chang et al., 2001, Reeves et al., 2002, Shepherd and Walsh, 2002). However, Kooistra et al. (2001) reported best predictions for clay, organic matter, Cd and Zn without pre-processing of spectra. Soil reflectance and soil variables have been calibrated using statistical methods like Multiple Linear Regression, Polynomial Regression, Principal Component Regression (Chang et al., 2001, Daniel et al., 2004, Islam et al., 2003, Udelhoven et al., 2003) and data mining techniques (Brown et al., 2006, Shepherd and Walsh, 2002). Partial least squares regression (PLSR) has been the most commonly used multivariate statistical method in calibrating soil reflectance to individual soil parameters and the estimation of those parameters, mainly due to its superiority over traditional methods in dealing with high dimensional multi-collinearity. It is superior to Principal Component Regression in that it uses the information in both predictor and response variables.

Multiple Adaptive Regression Splines (MARS) is a non-parametric multivariate regression method. It is capable of modeling both linear and nonlinear relationships between response and predictor variables by fitting local regression curves to spectral subregions and including higher order interactions among predictors. It has been successfully applied in various fields (Deichmann et al., 2002, Luoto and Hjort, 2005, Shepherd and Walsh, 2002, Yang et al., 2003) and generally provides better results in modeling compared to other linear and non-parametric regression techniques like Generalized Linear Models (GLM), Artificial Neutral Networks (ANN), and Classification and Regression Trees (CART).

The objectives of this study were to (i) determine whether VNIR spectroscopy can be used as a rapid, inexpensive alternative or supplement to traditional methods for measuring soil properties collected from a single field in northern Turkey, (ii) to evaluate whether the combined use of raw and first derivative spectra can help improve the estimations of parameters, (iii) to compare the predictive abilities of MARS with the more commonly used PLSR method, and (iv) to evaluate the utility of VNIR for single field characterization compared to other scales.

Section snippets

Study site and sampling

Soil samples were collected from a study area on the experimental farm of the Tokat Province Research Institute in northern Turkey (40° 32′ N lat, 36° 32′ E long), located at an average elevation of 580 m above sea level. The study area has a semi-arid climate with a mean annual precipitation, evaporation, temperature and relative humidity of 445 mm, 881 mm, 12 °C and 61%, respectively (Tokat Research Institute, 2000).

The study area (Fig. 1) covers 32 ha (800 m × 400 m), which was divided into 25 × 25 m

Soil properties

Summary statistics of soil variables and Pearson correlation coefficients between them are provided in Table 1, Table 2, respectively. The variables pH, K, EC and Na generally have narrow ranges in the data set as opposed to clay, sand, organic matter, CaCO3, CEC, and exchangeable Ca and Mg contents. Among exchangeable cations, Ca content was the highest (avg. 30.7 me 100 g−1) and made up the largest part of the CEC of soil samples. Soil EC was low for all soil variables.

Significant correlations

Conclusion

This study focused on the use of VNIR for predicting soil properties within a single agricultural field, and evaluated several different methods of spectral pretreatment and data analysis. Important environmental soil variables such as clay and SOM content were well predicted using hyperspectral VNIR spectroscopy. The results were generally in line with those of the other studies, even though they were conducted at different scales and in other geographic regions. The comparison of MARS and

Acknowledgements

This study was in part funded by the Cornell Computational Agriculture Initiative and by the Scientific Research Administration of Gaziosmanpasa University, Tokat, Turkey. The senior author is grateful for the support from the Turkish Higher Education Council (YOK).

References (43)

  • G.J. Bouyoucos

    The hydrometer method as a new and rapid method for determining the colloidal content of soils

    Soil Science

    (1926)
  • S.A. Bowers et al.

    Reflection of radiant energy from soils

    Soil Science

    (1965)
  • C.W. Chang et al.

    Near-Infrared reflectance spectroscopy – principal components regression analyses of soil properties

    Soil Science Society of America Journal

    (2001)
  • M. Chodak et al.

    Use of near infrared spectroscopy to determine biological and chemical characteristics of organic layers under spruce and beech stands

    Journal of Plant Nutrition and Soil Science

    (2001)
  • J. Cohen

    A coefficient of agreement for nominal scales

    Educational and Psychological Measurement

    (1960)
  • D. Cozzolino et al.

    The potential of near-infrared reflectance spectroscopy to analyse soil chemical and physical characteristics

    Journal of Agricultural Science

    (2003)
  • R.C. Dalal et al.

    Simultaneous determination of moisture, organic carbon, and total nitrogen by near infrared reflectance spectrophotometry

    Soil Science Society of America Journal

    (1986)
  • K.W. Daniel et al.

    Analysis of VNIR (400–1100 nm) spectral signatures for estimation of soil organic matter in tropical soils of Thailand

    International Journal of Remote Sensing

    (2004)
  • P. Dardanne et al.

    Multivariate calibration and chemometrics for near infrared spectroscopy: which method?

    Journal of Near Infrared Spectroscopy

    (2000)
  • B.W. Dunn et al.

    The potential of near-infrared reflectance spectroscopy for soil analysis – a case study from the Riverine Plain of south-eastern Australia

    Australian Journal of Experimental Agriculture

    (2002)
  • J.H. Friedman

    Multivariate adaptive regressions splines

    Annals of Statistics

    (1991)
  • Cited by (191)

    View all citing articles on Scopus
    View full text