Prediction of soil cation exchange capacity using visible and near infrared spectroscopy

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Available online at www.sciencedirect.com

ScienceDirect journal homepage: www.elsevier.com/locate/issn/15375110

Special Issue: Proximal Soil Sensing Research Paper

Prediction of soil cation exchange capacity using visible and near infrared spectroscopy Yahya Ulusoy a,*, Yu¨cel Tekin a,1, Zeynal Tu¨msavas‚ b, Abdul M. Mouazen c a

Vocational School of Technical Sciences, Uludag University, Bursa, Turkey Agricultural Faculty, Uludag University, Bursa, Turkey c Cranfield Soil and AgriFood Institute, Cranfield University, United Kingdom b

article info

This study was undertaken to investigate the application of visible and near infrared (vis

Article history:

eNIR) spectroscopy for determining soil cation exchange capacity (CEC) under laboratory

Published online xxx

and on-line field conditions. Measurements were conducted in two fields with clay texture in field 1 (F1) and clay-loam texture in field 2 (F2) both in Turkey. Partial least squares (PLS) regression analyses with full cross-validation were carried out to establish CEC models

Keywords:

using three datasets of F1, F2 and F1 þ F2. Analytically-measured, laboratory viseNIR and

Cation exchange capacity

on-line viseNIR predicted maps were produced and compared statistically by kappa co-

On-line soil sensor

efficient. Results of the CEC prediction using laboratory viseNIR data gave good prediction

Soil mapping

results, with averaged r2 values of 0.92 and 0.72, root mean squared errors of prediction

ViseNIR spectroscopy

(RMSEP) of 1.89 and 1.54 cmol kg1 and residual prediction deviations (RPD) of 3.69 and 1.89 for F1 and F2, respectively. Less successful predictions were obtained for the on-line measurement with r2 of 0.75 and 0.7, RMSEP of 4.79 and 1.76 cmol kg1 and RPD of 1.45 and 1.56 for F1 and F2, respectively. Comparisons using kappa statistics test indicated a significant agreement (k ¼ 0.69) between analytically-measured and laboratory viseNIR predicted CEC maps of F1, while poorer agreement was found for F2 (k ¼ 0.43). A moderate spatial similarity was also found between analytically-measured and on-line viseNIR predicted CEC maps in F1 (k ¼ 0.50) and F2 (k ¼ 0.49). This study suggests that soil CEC can be satisfactorily analysed using viseNIR spectroscopy under laboratory conditions and with somewhat less precision under on-line scanning conditions. © 2016 IAgrE. Published by Elsevier Ltd. All rights reserved.

* Corresponding author. Tel.: þ90 2242942307. E-mail addresses: [email protected] (Y. Ulusoy), [email protected] (Y. Tekin), [email protected] (Z. Tu¨msavas‚), [email protected] (A.M. Mouazen). 1 Tel.: þ90 2242942307. http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005 1537-5110/© 2016 IAgrE. Published by Elsevier Ltd. All rights reserved. Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

2

1.

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Introduction

Cation exchange capacity (CEC), which refers to the soil's ability to hold positively-charged ions, is an important indicator of soil fertility. It is an important property for sitespecific management of soil nutrients in precision agriculture. Generally, CEC values increase with the increase in the content of the colloidal fraction (amount and type of clay and organic matter) in the soil. The conventional analytical methods used for the determination of CEC such as the sodium saturation method (Chapman, 1965) are expensive, difficult and time consuming, because different cations must be extracted and determined. Therefore, researchers have attempted to find alternative methods that are simple, fast, and cost-effective. Proximal soil sensing methods are promising techniques for application with decision support tools that help with soil fertility management. In a review of the application of proximal soil sensing to assess soil properties, Kuang et al. (2012) identified visible and near infrared (viseNIR) spectroscopy as the most promising measurement technique to supply accurate and meaningful data on CEC for successful decision support on soil fertility management and viseNIR spectroscopy has become the most attractive proximal soil sensing method for acquiring data on key soil properties including CEC. Recent studies, e.g., Mouazen, Maleki, De Baerdemaeker, and Ramon (2007), Viscarra-Rossel and Chen (2011), Tekin, Kuang, and Mouazen (2013) and Kodaira and Shibusawa (2013), have shown that viseNIR spectroscopy provides accurate quantification of the main physical and chemical soil properties and that it is a useful tool for digital soil mapping and for precision agriculture applications. This is because viseNIR spectra of soils contain large sets of spectral information representing broad bands of overtones and combinations of fundamental vibrations occurring in the mid infrared (MIR) range of the electromagnetic spectrum (Stenberg, Viscarra Rossel, Mouazen, & Wetterlind, 2010). Quantitative information on CEC can be extracted with suitable multivariate regression methods, which have an advantage over simple bivariate relationships and are suitable for peak intensity measurements such as those existing in MIR range (SorianoDisla, Janik, Viscarra-Rossel, MacDonald, & McLaughlin, 2014). Partial least squares (PLS) regression is the most common technique adopted to model the relationships between infrared spectral intensity characteristics of soil components and soil properties through derived PLS loadings, scores and regression coefficients (Janik, Forrester, & Rawson, 2009). So far, the use of viseNIR spectroscopy for the determination of soil CEC has achieved varying degrees of success, depending on the conditions, under which the evaluations were carried out (e.g., in the laboratory or on-line in the field) (Bilgili, Van ~ asveras-Sa  nchez, Es, Akbas, Durak, & Hively, 2010; Can  Barron, del Campillo, & Viscarra-Rossel, 2012; Leone, Viscarra-Rossel, Amenta, & Buondonno, 2012; Marin z-Garcı´a, & Mouazen, 2013; Gonzalez, Kuang, Quraishi, Muno Savvides, Corstanje, Baxter, Rawlins, & Lark., 2010). Previous studies showed that the accuracy for CEC assessment under laboratory scanning conditions varied widely in the prediction set (r2 ¼ 0.13e0.90; RMSEP ¼ 1.22e10.43 cmol kg1;

RPD ¼ 0.60e2.7) (e.g., Awiti, Walsh, Shepherd, & Kinyamario, 2008; Ben-Dor & Banin, 1995; Chang, Laird, Mausbach, & Hurburgh, 2001; Mouazen, De Baerdemaeker, & Ramon, 2006; Waruru, Shepherd, Ndegwa, Kamoni, & Sila, 2014). There are several reasons for the different results reported, which include the type of spectrophotometer, variability of CEC in the dataset, modelling technique applied and soil type. MarinGonzalez et al. (2013) reported moderately good prediction of CEC (r2 ¼ 0.62; RMSEP ¼ 0.97 cmol kg1, and RPD ¼ 1.61) for online measurement in a field in Bedfordshire in the UK. Kweon, Lund, and Maxton (2013) reported successful on-line measurements of six out of nine fields with r2 of 0.86 or higher and RPD of 2.78 or greater. To the best of our knowledge, no report of on-line measurement of CEC in arid and semi-arid environments has been presented in the literature. The aim of this study was to evaluate the accuracy of an on-line viseNIR spectroscopy sensor to assess and map the CEC of the soil in two fields located in semi-arid regions of Turkey. The paper also compared modelling results and maps developed for CEC with individual- and mixed-field datasets.

2.

Materials and methods

2.1.

Experimental sites

This study was carried out in two fields in Turkey in 2014, one located in the village of Karacabey in Bursa Province (F1; 10.06 ha; Latitude: 40 090 10.800 N and Longitude: 28 230 01.000 E), and the second field a circular one located in the village of Ayrancı in Karaman Province (F2; 50 ha; Latitude: 37 320 39.100 N and Longitude: 33 400 43.600 E). Both fields were irrigated. According to soil maps of the General Directorate of Rural Service of Turkey, the soil types of Karacabey field and Ayrancı field are classified as vertisols and fluvisols (Alluvial), respectively. Karacabey field is of a clay soil texture, whereas Ayrancı field is of a clay loam texture (Table 1). The crops cultivated in 2014 in Karacabey and Ayrancı fields were barley and wheat, respectively.

2.2.

On-line soil sensor and measurement

The on-line soil sensor consisted of a subsoiler with an optical probe attached to a chisel (Mouazen, 2006). The subsoiler, acting as a soil-cutting tool, together with the optical probe were installed on a frame (Mouazen, De Baerdemaeker, & Ramon, 2005), which had been manufactured at Uludag University. This on-line sensor was mounted on the three-point linkage of a tractor for field measurements (Fig. 1). An AgroSpec mobile, fibre-type viseNIR spectrophotometer (Tec5 Technology for Spectroscopy, Germany) having a spectral range of 350e2200 nm was used in this study. A differential global positioning system (DGPS) (EZ-Guide 250, Trimble, USA) with a sub-metre accuracy was employed to record the position of the measured on-line spectra. An AgroSpec software, a platform for the mobile spectrometer system, was used to acquire soil spectral and DGPS data. More detailed information about the on-line sensor can be found in Quraishi and Mouazen (2013).

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

3

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Table 1 e Field and soil characteristics for Karacabey field (F1) and Ayrancı Field (F2). Field F1 F2

Area (ha)

Crop

Texture type

Sand (%)

Silt (%)

Clay (%)

Average field MC (%)

Average field SOC (%)

10 50

Wheat Barley

Clay Clay-loam

26.6 34.4

30.4 33.8

43.0 31.8

22.2 8.16

1.41 1.30

MC, moisture content. SOC, soil organic carbon.

On-line raw spectra of the soil were collected along parallel transects (20 m apart) at a constant speed of approximately 3 km h1. A measurement depth of 0.15 m was set and controlled during the measurement by two pairs of metal wheels installed on each side of the frame. The probe was positioned parallel to the surface with the help of a water gauge. A total of 92 and 238 disturbed soil samples were collected from F1 and F2, respectively, from the bottom of trenches opened by the subsoiler. Each soil sample was collected over 1 m distance. Sampling points were chosen randomly at approximately every 20 m along the trench lines. The coordinates of the sampling points were recorded with the same DGPS. Figure 2 shows the sampling points and measured transects in F1 and F2.

2.3.

Laboratory reference measurements

Samples from each field were subdivided into two sets. The first set was used for laboratory reference measurements of CEC, moisture content (MC), soil organic carbon content (SOC) and particle size distribution (PSD), whereas the second set was used for optical measurements in the laboratory. The CEC of the soil samples was determined using the Eppendorf Elex 6361 flame photometer (Eppendorf, Hamburg, Germany), according to the sodium saturation method (Chapman, 1965). Average field SOC was measured by the WalkleyeBlack method (Nelson & Sommers, 1982), whereas MC was measured by oven-drying the samples at 105  C for 24 h (Black, 1965). The PSD was evaluated by sieving and sedimentation (British Standard, 1998). The PSD analysis results for both fields were used to determine their textural class according to the United States Department of Agriculture (USDA) soil textural classification system (Table 1).

2.4.

Optical measurement in the laboratory

Soil samples were scanned in the laboratory using the same viseNIR spectrophotometer used during the on-line field measurements. Before scanning, each soil sample was preprocessed to get rid of stubble, plant roots and residues, and gravel, and thoroughly mixed. Next, each soil sample was subsampled three times and each placed in plastic dishes (12 mm deep and 12 mm in diameter). The loose soil in the dishes was first compacted and then carefully levelled to form a smooth scanning surface (Mouazen, Karoui, Deckers, De Baerdemaeker, & Ramon, 2007; Mouazen et al., 2005). A white Spectralon® disc of almost 100% reflectance was used to optimise the instrument before scanning and every 30 min afterwards. Each dish was scanned 10 times, and the readings obtained were averaged in one spectrum. The final spectrum, used for further analysis, was an average of the spectra obtained for the three dishes.

2.5.

Soil spectral pre-processing

Analyses of soil spectra were performed on fresh soil samples either scanned in the field with the on-line soil sensor or in the laboratory under static conditions. Unscrambler® software Version 9.8 (Camo AS, Oslo, Norway) was used for spectral pre-processing and model development. Pre-processing of the spectra was performed in order to remove noisy portions of the spectrum and to eliminate some sources of variation not related to the variables of interest. Noise was present in the 305e421 and 1745e2200 nm ranges due to low reflectance of the soil and lower sensitivity of the instrument at these wavelengths. Spectra within wavelengths of 421e1745 nm were selected for the calibration, after removing the noisy

Fig. 1 e The on-line visible and near infrared (viseNIR) spectroscopy soil sensor attached to the three point linkage of a tractor, mimicking the design of Mouazen (2006). Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

4

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Fig. 2 e Soil sampling positions for (a) Karacabey field (F1) and (d) Ayrancı Field (F2); on-line soil measurement transects for (b) F1 and (e) F2; position of soil samples collected for validation for (c) F1 and (f ) F2.

parts at the two ends of the spectrum. After spectral truncation, the spectra showed clearly discernible reflectance dips and peaks in both visible and NIR ranges (Fig. 3). Different data pre-processing options were adopted to develop calibration models. These included reducing the wavelengths by averaging three adjacent wavelengths into one wavelength for 421e1000 nm range and six for

1001e1745 nm range. This reduction in the number of wavelengths was adopted as it resulted in the best model performance. Wavelength-averaging was followed successively by maximum normalisation, the SavitzkyeGolay 1st derivative (Savitzky & Golay, 1964), and finally, smoothing, in order to reduce the negative influence of noise and remove additive baseline shift. The 1st derivative was adopted using a second-order polynomial with a polynomial order of two fitted to the spectra. The SavitzkyeGolay smoothing method was applied for two smoothing points using a second-order polynomial.

2.6.

Fig. 3 e One laboratory and five on-line soil spectra collected from one sampling location in each field of both refers to on-line spectra and to Karacabey field (F1; refers to onlaboratory spectra) and Ayrancı Field (F2; to laboratory spectra) as examples. line spectra and

Modelling

The PLS regression analysis with full cross-validation was used for modelling. Before each PLS analysis, the entire spectra were split randomly into two sets of 80% (calibration set) and 20% (prediction set). The calibration set of F1 consisted of 74 randomly selected soil samples, whereas the remaining 18 samples were used as the prediction set. In an attempt to enhance the prediction capability of the CEC model in F2 having small variability, two calibration models were developed. For the first model, 198 randomly selected soil samples collected from F2 were used as the calibration set, while the remaining 40 samples were used as the prediction set. The second model was established by adding 56 samples from F1 within the same CEC range to the F2 samples. This model was designated as the F1 þ F2 model. The aim was to compare prediction accuracies between the CEC model developed with individual field samples (F2 model) with the

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

corresponding model developed with a mixture of samples from two fields of different soil types (F1 þ F2 model). The performance and accuracy of the CEC calibration model were evaluated in cross-validation using samples of the calibration set and in prediction using samples of the prediction set. Model performance was evaluated by means of the coefficient of determination (r2), root mean square error of prediction (RMSEP, cmol kg1) and ratio of residual prediction deviation (RPD), which is standard deviation (SD) divided by RMSEP (cmol kg1). The evaluation of model performance, in terms of prediction ability, was based on the following criteria: excellent if RPD > 2.0, good if 1.4  RPD < 2.0, and unreliable if RPD < 1.4 (Chang et al., 2001).

2.7. maps

Development of cation exchange capacity (CEC)

Three categories of CEC maps were generated for both fields: Category I e Analytically-measured CEC maps: these maps were developed using the entire dataset collected for laboratory reference measurement (i.e., 92 in F1 and 238 in F2). Category II e Comparison maps: these maps included analytically-measured, laboratory vis-NIR predicted and on-line viseNIR predicted maps developed with the prediction datasets of F1 (18 samples) and F2 (40 samples), and Category III e On-line full-data point maps: these maps were developed with the entire on-line viseNIR predicted points for F1 (6486) and F2 (16,830). The inverse distance weighting (IDW) interpolation method was used to develop the analytically-measured, laboratory viseNIR-predicted, and on-line viseNIR-predicted maps of categories I and II. The full-data point maps (category III) were developed using kriging interpolation method after semivariogram development. The reason for choosing IDW for categories I and II and kriging for category III was the limited number of data points in the former two categories as compared to category III. The IDW interpolation tools are referred to as deterministic interpolation methods because they are directly based on the surrounding measured values or on specified mathematical formulas that determine the smoothness of the resulting surface. The kriging method is from the second family of interpolation methods, which consists of geostatistical techniques, based on statistical models that include autocorrelation. All maps were developed using ArcGIS 10 (ESRI, USA) software. The advanced parameters option in ArcGIS allowed control of the semivariogram used for kriging. In order to observe the visual relationship between different maps, ArcGIS Geostatistical Analyst General Quantileequantile (QeQ) tools were used. However, in order to quantify spatial similarities between pairs of maps, kappa (k) is the common analysis (Mouazen et al., 2009). Kappa expresses proportionally how much better the results are compared to a purely random classification. The larger the k value is, the more accurate is the classification; hence, similarity between maps. Using Statistical Package for the Social Sciences (SPSS) (IBM, USA), analyses were carried out to

5

calculate kappa (k) values (Cohen, 1960) in order to compare the statistical relationship of the pairs of maps.

3.

Results and discussion

3.1.

Quality of on-line soil spectra

Figure 3 shows laboratory and on-line measured soil spectra for the two fields. Of the five on-line viseNIR soil spectra shown in Fig. 3, one spectrum is for the same location as the laboratory soil sample collected for wet chemistry analyses while the remaining four spectra are collected at 2 s interval before and after the reference point. It can be observed that there is considerable similarity between one spectrum of the five on-line spectra and the laboratory spectrum. With small changes in sample position of about 0.5 m, differences in the spectra are expected, potentially reflecting actual differences in the soil. This necessitates high sampling resolution to explore existing spatial variability of soil provided by on-line soil sensors. However, as compared to the laboratorymeasured spectrum, the on-line soil spectra show upward or downward shift around 970 nm, which is in line with the results of Mouazen et al. (2009), using a diode array spectrometer from Zeiss (Zeiss Corona 45 viseNIR Fibre, Germany). They attributed this shift to the variation in sensor-to-soil distance or angle due to vibration induced by tractor driving over an irregular soil surface, and may be overcome by addition or subtraction of reflectance values for the second part of soil spectra that starts after the shift. Clear differences between the spectra collected from the two fields can be observed, and are likely to be mainly due to differences in soil type and mineralogy, soil colour, MC, and physico-chemical composition of the soil in the two fields. The diffuse reflectance of field F1 was much smaller and characterised by smaller absorption dips at 1450 (second overtone of water) than that of field F2 (Fig. 3). This can be easily explained by the larger MC of F1 (average of 22.2%), than that of F2 (average of 8.16%). With increasing MC, diffuse reflectance decreases due to increase in absorption associated with increase in soil darkness (Mouazen, Karoui, De Baerdemaeker, & Ramon, 2006; Nocita, Stevens, Noon, & Wesemael, 2013; Tekin, Tumsavas‚, & Mouazen, 2012).

3.2.

Variability of cation exchange capacity

Detailed descriptive statistics for the analytically-measured CEC are provided in Table 2. The minimum, maximum, mean, and SD of the F1 CEC values vary more widely (14.48, 40.39, 27.92 and 6.02 cmol kg1, respectively) than the corresponding values for the F2 (9.42, 28.55, 17.58 and 2.56 cmol kg1, respectively). The smaller the field variability of a soil property, the less likely it will be to obtain good calibration models (Kuang & Mouazen, 2011). Kuang and Mouazen (2011) reported that apart from higher r2 and RPD, higher RMSEP values are to be expected for a larger concentration range and SD of a soil property as compared with smaller concentration ranges. In fact, small variability in a soil property may lead to unsuccessful regression. This necessitates focussing on viseNIR mapping of soil properties only for fields

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

6

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Table 2 e Sample statistics of laboratory and on-line measured cation exchange capacity (CEC) of the calibration and prediction sets of both Karacabey field (F1) and Ayrancı Field (F2). Field All samples

Cross-validation set

Laboratory prediction set

On-line prediction set

F1 F2 F1 F1 F2 F1 F1 F2 F1 F1 F2 F1

Sample number

Min (cmol kg1)

Max (cmol kg1)

Mean (cmol kg1)

SD (cmol kg1)

92 238 294 74 198 254 18 40 40 18 40 40

14.48 9.42 9.42 13 13.12 13.12 14.48 13.9 14.55 14.48 13.9 14.55

40.39 28.55 34.39 35.98 23.98 29.26 40.39 22.28 22.47 40.39 22.28 22.47

27.92 17.58 18.77 27.70 17.57 18.83 28.87 17.53 17.65 28.87 17.53 17.65

6.02 2.56 4.03 5.23 4.64 3.83 6.97 2.91 2.90 6.94 2.90 2.89

þ F2

þ F2

þ F2

þ F2

SD, standard deviation.

with considerable variability to allow for successful regression models to be established, from which successful decision support for variable rate applications can be made.

3.3.

Model performance in calibration and prediction

The model performance in cross-validation for laboratory and on-line predictions for both fields are shown in Table 3. Figure 4 shows scatter plots of analytically-measured vs. viseNIR-predicted CEC using the prediction set (18 samples in F1 and 40 samples in F2) for the laboratory scanned and on-line scanned soil spectra. According to the classification of RPD values proposed by Chang et al. (2001), the performance of the CEC model in cross-validation for both fields can be classified as excellent (RMSEP ¼ 1.21 cmol kg1 and RPD ¼ 4.38 for F1; RMSEP ¼ 1.41 cmol kg1 and RPD ¼ 3.29 for F2; RMSEP ¼ 1.45 cmol kg1 and RPD ¼ 2.64 for F1 þ F2). Islam, Singh, and McBratney (2003) reported lower prediction results for CEC (r2 ¼ 0.64, RPD ¼ 1.6) when using an independent validation set of a total of 161 samples. Using principal component regression analysis performed on 802 samples, Chang et al. (2001) reported sufficient prediction results (r2 ¼ 0.81, RMSEP ¼ 3.82 cmol kg1, and RPD ¼ 2.28). Zornoza et al. (2008) successfully predicted CEC with comparable

results (r2 ¼ 0.92, RMSEP ¼ 0.06 cmol kg1 and RPD ¼ 3.46) to those obtained in the current work. Using 582 soil samples, Awiti et al. (2008) presented similar calibration results, but with higher error (r2 ¼ 0.90, RMSEP ¼ 3.24 cmol kg1). Waruru et al. (2014) reported good prediction results in the prediction ¼ 0.7, set but again with larger error (r2 1 RMSEP ¼ 9.6 cmol kg and RPD ¼ 1.7). Kweon et al. (2013) reported successful on-line predictions of CEC in six fields in the US with r2  0.86 and RPD  2.78. Using 146 samples collected from one field in the UK, Marin-Gonzalez et al. (2013) obtained good prediction accuracy for laboratory and on-line measurements (RPD ¼ 1.70 and 1.61 and r2 ¼ 0.72 and 0.62, respectively). Result of the current study for the prediction of CEC have shown that the performance of the viseNIR models obtained under on-line measurement conditions was not as good as that obtained under laboratory measurement conditions (Table 3). The performance of the on-line CEC models with RPD ¼ 1.45 for F1, 1.56 for F2, and 1.36 for F1 þ F2 can be classified as good (for F1 and F2 data) and unreliable (for F1 þ F2 data) predictions. This deterioration in prediction performance can be explained by the presence of several ambient factors and uncontrolled conditions influencing online measurement. These include among others tractor

Table 3 e Summary of cation exchange capacity (CEC) model performance for cross-validation, laboratory and on-line predictions in both Karacabey field (F1) and Ayrancı Field (F2).

Cross-validation

Laboratory prediction

On-line prediction

Field

r2

RMSEP (cmol kg1)

RPD

Intercept

Slope

LV's

F1 F2 F1 þ F2 F1 F2 F1 þ F2 F1 F2 F1 þ F2

0.82 0.58 0.83 0.92 0.72 0.74 0.75 0.70 0.73

1.21 1.41 1.45 1.89 1.54 1.58 4.79 1.86 2.13

4.38 3.29 2.64 3.69 1.89 1.84 1.45 1.56 1.36

0.58 0.85 2.99 1.41 6.23 6.95 2.56 0.55 1.8

0.98 0.95 0.83 0.95 0.63 0.60 0.79 0.98 1.14

5 7 9 4 6 7 7 6 7

RMSEP, root mean square error of prediction. RPD, residual prediction deviation. LV's, latent variables.

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

7

Fig. 4 e Scatter plots of (a) analytically-measured vs. visible and near infrared (viseNIR) predicted cation exchange capacity (CEC) of the prediction set for laboratory (r2 ¼ 0.92) and ( b) on-line scanned soil spectra (r2 ¼ 0.75) of Karacabey field (F1) (18 samples); (c) laboratory scanned (r2 ¼ 0.72) and (d) on-line scanned soil spectra (r2 ¼ 0.73) of Ayrancı Field (F2) (40 samples) of the F2 calibration model; (e) laboratory scanned (40 samples; r2 ¼ 0.74) and (f ) on-line scanned soil spectra (r2 ¼ 0.73) of the F1 þ F2 calibration model.

vibration, sensor-to-soil distance and/or angle variation, presence of plant roots, debris and stones, and ambient light reaching the spectrophotometer detector (Mouazen, Maleki, et al., 2007; Stenberg et al., 2010). In addition, the combined dataset of F1 þ F2 did not lead to improved prediction

performance of the CEC model in F2 either under laboratory or on-line field measurement conditions (Table 3). This may suggest relying on one field dataset rather than mixing two datasets, especially when the concentration ranges of CEC from the two sample sets are identical. This will indeed

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

8

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

increase the complexity of the sample set, and results in less successful calibration models.

3.4.

Regression coefficients plot

It was reported that CEC is a soil property that does not have a direct spectral response in the NIR range (Stenberg et al., 2010). Therefore, CEC has been measured with various degrees of success through co-variation with other soil properties possessing direct spectral responses such as organic matter, clay content and clay type, and moisture content (Stenberg et al., 2010). In the visible range, successful measurement of these ‘secondary’ soil properties, which are associated with colour changes, is due to electron transition (Viscarra-Rossel, Walvoort, McBratney, Janik, & Skjemstad, 2006). For example, an increase in soil organic carbon leads to a darker soil colour, which has been attributed to increased absorption in the blue waveband around 450 nm. Changes in colour have also been associated with mineralogy (i.e., iron oxides), which can be observed around the red absorption band at 680 nm (Mouazen, Karoui, et al., 2007; Mouazen et al., 2005). The regression coefficients after PLS analysis for the cross-validation models of F1, F2, and F1 þ F2 are shown in Fig. 5. There was considerable similarity between F1 and F2, although the depth of absorption for F1 was considerably greater than that of F2. On the other hand, considerably different regression coeffecients curve can be observed for F1 þ F2 when compared individually with those of F1 and F2. In the visible range, the highest regression coefficient occurred around the blue waveband near 445 nm for F1 only, which can be explained by the darker colour of vertisols in F1, as compared to F2. It has been reported that the absorption band at 450 nm could also be caused by paired and single Fe3þ electron transitions to a higher energy state (Sherman & Waite, 1985). Three other absorption bands in

600 400

Intensity, au

200 0 -200

the visible range around 520, 588, and 624 nm can be observed in the regression coefficients plots of both fields, which may also be linked to colour changes due to mineralogy. These absorption wavebands may well explain the link with iron oxide minerals, hence, contributing to successful CEC prediction. In the NIR range, the intensity of regression coefficients for some important bands was higher than those observed in the visible range. The highest regression coefficient was observed around 1400 nm, which is linked to the OeH stretching mode of water in the second overtone of the NIR region. This observation is supported by the fact that CEC is directly linked with clay type and content, which directly affects the ability of soil to absorb water and also adsorb nutrients. However, the regression coefficient at this band is larger for F1 (clay soil with higher MC of 22.2%, as shown in Table 1) than for F2 (clayloam soil of low MC of 8.16%), which may further explain why CEC is predicted with higher r2 and RPD in F1 than in F2 (Table 3). Three other distinguishing absorption bands around 1519, 1591, and 1639 nm were observed, particularly for F1, which may also contribute to the larger r2 and RPD for CEC prediction in this field than in F2. Viscarra Rossel and Behrens (2010) have attributed the absorption around 1500 nm to amine (NeH) and that around 1650 nm and 1100 nm to aromatic CeH (Fig. 5), all associated with organic matter content, which directly affects the CEC content of the soil (Moore, Dolling, Porter, & Leonard, 1998). Helling, Chesters, and Corey (1964) indicated that the mean relative contribution of organic matter to total soil CEC in soils varied from 19% at pH 2.5e45% at pH 8.0 for mean organic matter and clay contents of the soils of 3.28% and 13.3%, respectively. Examining the Pearson correlation (r) between analytically-measured CEC on the one hand and SOC, clay content and moisture content on the other (Table 4) reveals high correlations in F1 only. The highest r value of 0.948 is calculated for correlation between CEC and clay, confirming that CEC is measured with viseNIR spectroscopy through covariation, particularly with clay and to less extent with MC (r ¼ 0.687) and SOC (r ¼ 0.575). However, for F2 and F1 þ F2 datasets this is not true, as negligible correlations are calculated between CEC and clay, MC and SOC. This suggests that CEC is not necessarily measured only through co-variation with other soil properties that have direct spectral responses in the NIR range (i.e., clay, MC and SOC). Further studies may be needed to determine whether CEC has any direct spectral responses.

-400 -600 -800 200

400

600

800

1000

1200

1400

1600

1800

Wavelength, nm Fig. 5 e Plot of regression coefficients against wavelength after partial least-squares (PLS) regression analysis by visible and near infrared (viseNIR) spectroscopy for crossvalidation models of cation exchange capacity (CEC) for Karacabey field (F1) (thick solid line), Ayrancı field (F2) (thin solid line), and F1 þ F2 (dotted line).

Table 4 e Pearson correlation (r) between analyticallymeasured cation exchange capacity (CEC) on one hand and soil organic carbon (SOC), moisture content (MC) and clay content on the other hand, presented for Field 1 (F1), Field 2 (F2) and F1 þ F2 datasets. Soil properties

CEC*clay CEC*SOC CEC*MC

Pearson correlation (r) F1

F2

F1 þ F2

0.948 0.575 0.687

0.051 0.076 0.023

0.076 0.022 0.048

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

3.5.

Mapping

3.5.1. Comparison of laboratory reference and viseNIR CEC predicted maps The comparison of the analytically-measured and viseNIR CEC predicted maps using the prediction set of 18 samples of F1 shows similar spatial patterns with distinguishable high and low CEC zones (Fig. 6). However, a closer agreement can be observed between the analytically-measured and laboratory viseNIR predicted maps, as compared to the on-line viseNIR predicted maps. The similarity of the analyticallymeasured (92 samples) and on-line (6486 points) CEC maps can be readily seen (Fig. 7). Again, both maps show similar spatial distribution of CEC, and similar to the three maps of Fig. 6, with the central-southern part of the field of a higher CEC range, as compared to the two northern corners. In F2 (Fig. 8), soil maps of analytically-measured and laboratory and on-line viseNIR-predicted CEC demonstrate some

9

spatial similarity. Nonetheless, the range of CEC variation in F2 is much smaller than that in F1. All F2 soil maps showed zones with high CEC values in the southeast part of the field, twice those of the northwest part of the field. The on-line viseNIR full-data point prediction maps for F2 (Fig. 9b and c) show a rather different spatial distribution of CEC, as compared to corresponding maps produced with a small number of data points (see Fig. 8d and e). However, similarity between the full-data point maps (Fig. 9b and c) and the corresponding map of the analytical measurement using 238 samples (Fig. 9a) is reasonably distinguishable. The full-data point maps indicate that the southern part of the field has the highest CEC level while the central part of the field has the smallest level of CEC, when compared to zones at the edges of the field (Fig. 9). The high sampling resolution obtained with the on-line soil sensor provided more detailed information about CEC spatial distribution, although both F1 and F1 þ F2 models led to a somewhat similar spatial distribution of CEC.

Fig. 6 e Comparison between (a) analytically-measured, ( b) laboratory visible and near infrared (viseNIR) predicted and (c) on-line viseNIR predicted maps of cation exchange capacity (CEC), based on 18 samples of the prediction set for Karacabey field (F1). Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

10

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Fig. 7 e Comparison between (a) analytically-measured (92 samples) and ( b) full-data point on-line visible and near infrared (viseNIR) predicted (6486) maps of cation exchange capacity (CEC) for Karacabey field (F1).

These results suggest that this technology is useful to provide more detailed soil information for an improved site-specific management of soil nutrients. A general QeQ plot was used to assess the visual similarity of the distributions of the F1 datasets (Fig. 10). QeQ plots compare the quintiles of data distribution with the quintiles of standardised theoretical distribution from a specified family of distributions (ESRI®, 2014). The QeQ plots for analyticallymeasured CEC vs. laboratory viseNIR predicted CEC (Fig. 10a) using 18 samples of the prediction set showed normal distribution of the datasets, indicating a high similarity between the two maps. Similarly, a QeQ plot comparing between analytically-measured CEC (all 92 samples) and online predicted full-data point CEC maps (6486 points) show a reasonable similarity (Fig. 10b) due to the normal distribution exhibited. The QeQ plot of analytically measured CEC (40 samples) vs. on-line viseNIR predicted maps (40 samples) for the F2 (Fig. 11a) and F1 þ F2 (Fig. 11b) models showed a normal distribution indicating high similarity. A similar distribution was observed between analytically measured (238) and on-line (16,830) viseNIR predicted maps for both F2 (Fig. 11c) and F1 þ F2 (Fig. 11d) models. However, deviation from the straight line (45 line) distribution was observed for few points in Fig. 11c and d, indicating a poorer spatial similarity, which might be attributed to the larger number of points used in Fig. 11a and b.

3.5.2.

Statistical comparison of CEC maps

Landis and Koch (1977) categorised kappa values of 0.0 as no agreement, 0.00e0.20 as slight agreement, 0.21e0.40 as fair agreement, 0.41e0.60 as moderate agreement, 0.61e0.80 as substantial agreement, and 0.81e1.00 as excellent agreement.

The output of the kappa statistics test comparison between analytically-measured CEC and laboratory viseNIR predicted CEC maps for F1 indicates substantial agreement (k ¼ 0.69), whereas moderate agreement is recorded for F2 (k ¼ 0.43) and F1 þ F2 (k ¼ 0.46) (Table 5). In the prediction set, the k values for analytically-measured CEC and on-line viseNIR predicted CEC show moderate agreement for the three modelling cases, with almost equal k values ranging between 0.49 and 0.51 (Table 5). The worst k value is obtained for the analytically-measured vs. on-line viseNIR predicted CEC full-data point maps for F2 with k ¼ 0.39, a value that can be classified as slight agreement. As discussed earlier, predictions of CEC are more successful in F1 than in F2; this is reflected in the quality of soil maps developed as the k value for the analytically-measured vs. on-line viseNIR predicted CEC full-data point map for F1 field is much larger (k ¼ 0.55) than that for F2.

4.

Conclusions

This paper evaluates the accuracy of visible and near infrared (viseNIR) spectroscopy for the measurement of soil cation exchange capacity (CEC) in two fields with heavy soil types (i.e., clay (F1) and clay loam (F2) in semi-arid environments in Turkey). Results were evaluated under laboratory and on-line field measurement conditions, which support the following conclusions: 1 ViseNIR spectroscopy can be successfully used for the measurement of CEC in heavy soils under semi-arid environmental conditions. However, to generalise this conclusion, measurements across a wider range of soil types are needed because of the high variability in soil.

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

11

Fig. 8 e Comparison maps for Ayrancı Field (F2) of (a) analytically-measured cation exchange capacity (CEC); ( b) laboratory visible and near infrared (viseNIR) predicted CEC based on F2 calibration model and (c) F1 þ F2 model; (d) on-line viseNIR predicted CEC based on F2 model and (e) F1 þ F2 model. All maps are shown for the prediction set of 40 samples.

2 A greater accuracy in prediction of CEC using viseNIR spectroscopy was observed under laboratory conditions than under on-line conditions. Results showed averaged r2 values of 0.92 and 0.72, RMSEP of 1.89 and 1.54 cmol kg1

and RPD of 3.69 and 1.89 for F1 and F2, respectively, which indicates an excellent and good laboratory prediction of CEC. The on-line measured spectral data in the fields yielded good prediction with r2 of 0.75 and 0.7, RMSEP of

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

12

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Fig. 9 e Comparison of (a) analytically-measured maps based on 238 measurement points with ( b) on-line predicted fulldata point maps of cation exchange capacity (CEC) (16,830 points) for Ayrancı Field (F2) based on F2 calibration model and (c) F1 þ F2 calibration model.

Fig. 10 e QeQ plots of (a) analytically-measured vs. on-line visible and near infrared (viseNIR) predicted cation exchange capacity (CEC), using 18 samples, and ( b) analytically-measured vs. on-line viseNIR predicted full-data point CEC for Karacabey field (F1). Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

13

Fig. 11 e QeQ plots of (a) analytically-measured (40 samples) vs. on-line visible and near infrared (viseNIR) predicted cation exchange capacity (CEC) for F2 and ( b) F1 þ F2 models and (c) analytically-measured (238 samples) vs. full-data point viseNIR predicted CEC for F2 and (d) F1 þ F2 models.

Table 5 e Results of Kappa statistic comparing symmetric measures of maps of cation exchange capacity (CEC) with different mapping options for both Karacabey field (F1) and Ayrancı Field (F2). Pairs Chemically measured e laboratory viseNIR predicted CEC Chemically measured e on-line viseNIR predicted CEC in prediction set Chemically measured e full-points viseNIR predicted CEC of full-points

Field

Approx. T1

k

F1 F2 F1 þ F2 F1 F2 F1 þ F2 F1 F2 F1 þ F2

376.048 182.195 197.946 283.710 209.406 218.581 308.644 177.680 177.705

0.69 0.43 0.46 0.50 0.49 0.51 0.545 0.393 0.423

k: Kappa. T1: using the asymptotic standard error assuming the null hypothesis.

4.79 and 1.76 cmol kg1 and RPD of 1.45 and 1.56 for F1 and F2, respectively. 3 Pearson correlation analysis between CEC and SOC, MC and clay suggested that CEC may not necessarily be measured only through co-variation with other soil properties that are known to have direct spectral responses in the NIR range (i.e., clay, MC and SOC). Whether CEC has direct spectral responses requires further studies. 4 ViseNIR-predicted maps of CEC were similar to the corresponding analytically-measured CEC maps. Comparisons using kappa statistics tests indicated a significant agreement (k ¼ 0.69) between analytically-measured and laboratory-viseNIR predicted CEC maps of F1 while a poorer agreement was found for F2 (k ¼ 0.43). A moderate spatial similarity was also found between analytically-measured and on-line viseNIR predicted CEC maps in F1 (k ¼ 0.50) and F2 (k ¼ 0.49). However, the high-resolution full-data point maps were more detailed and showed different spatial distributions, as compared to the maps developed with limited numbers of points.

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

14

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

Currently, research is underway to establish links between CEC and crop canopy characteristics and yield, by fusion of on-line viseNIR spectroscopy data on soil CEC with data on crop canopy and yield collected with satellite imagery, and yield sensor, respectively.

Acknowledgements This research was sponsored by ICT-AGRI (62-FARMFUSE) (The European Commission's ERA-NET scheme under the 7th ¨ BITAK (The Scientific and TechFramework Programme), TU nological Research Council of Turkey e Contract no. 112O471), and the UK Department of Environment, Food and Rural Affairs (Contract no. IF0208). The research received funds for the FarmFUSE project entitled “Fusion of multi-source and multisensor information on soil and crops for an optimised crop production system.”

references

Awiti, A. O., Walsh, M. G., Shepherd, K. D., & Kinyamario, J. (2008). Soil condition classification using infrared spectroscopy: a proposition for assessment of soil condition along a tropical forest-cropland chronosequence. Geoderma, 143, 73e84. Ben-Dor, E., & Banin, A. (1995). Near-infrared analysis as a rapid method to simultaneously evaluate several soil properties. Soil Science Society of America Journal, 59, 364e372. Bilgili, A. V., Van Es, H. M., Akbas, F., Durak, A., & Hively, W. D. (2010). Visible-near infrared reflectance spectroscopy for assessment of soil properties in a semiarid area of Turkey. Journal of Arid Environments, 74, 229e238. Black, C. A. (1965). Methods of soil analysis: Part I physical and mineralogical properties. Madison, Wisconsin, USA: American Society of Agronomy. British Standard. (1998). Soil quality e BS 7755: Physical methods e Section 5.4 determination of particle size distribution in mineral soil material e Method by sieving and sedimentation. UK: British Standards Institution. ~ asveras-Sa  nchez, J. C., Barro  n, V., del Campillo, M. C., & Can Viscarra-Rossel, R. A. (2012). Reflectance spectroscopy: a tool for predicting soil properties related to the incidence of Fe chlorosis. Spanish Journal of Agricultural Research, 10, 1133e1142. Chang, C., Laird, D. A., Mausbach, M. J., & Hurburgh, C. R., Jr. (2001). Near-infrared reflectance spectroscopy-principal components regression analyses of soil properties. Soil Science Society of America Journal, 65, 480e490. Chapman, H. D. (1965). Cation-exchange capacity. In C. A. Black (Ed.), Methods of soil analysis. Part 2 (pp. 891e901). New York: Academic Press. Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1), 37e46. ESRI. (2014). ArcGIS® version 10 e Help: Geostatistical analyst. USA: ESRI Inc.. Available online http://resources.arcgis.com/en/ help/ Accessed: 03.03.14. Helling, C., Chesters, G., & Corey, R. B. (1964). Contribution of organic matter and clay to soil cation-exchange capacity as affected by the pH of the saturating solution. Soil Science Society of America Journal, 28, 4. Islam, K., Singh, B., & McBratney, A. (2003). Simultaneous estimation of several soil properties by ultra-violet, visible,

and near-infrared reflectance spectroscopy. Australian Journal of Soil Research, 41, 1101e1114. Janik, L. J., Forrester, S. T., & Rawson, A. (2009). The prediction of soil chemical and physical properties from mid-infrared spectroscopy and combined partial least squares regression and neural networks (PLS-NN) analysis. Chemometrics and Intelligent Laboratory Systems, 97, 179e188. Kodaira, M., & Shibusawa, S. (2013). Using a mobile real-time soil visible-near infrared sensor for high resolution soil property mapping. Geoderma, 199, 64e79. Special Issue (SI). Kuang, B., Mahmood, H. S., Quraishi, Z., Hoogmoed, W. B., Mouazen, A. M., & van Henten, E. J. (2012). Sensing soil properties in the laboratory, in situ, and on-line: a review. In D. Sparks (Ed.), Advances in agronomy (vol. 114, pp. 155e224). UK: Academic Press. AGRON. Kuang, B., & Mouazen, A. M. (2011). Calibration of visible and near infrared spectroscopy for soil analysis at the field scale on three European farms. European Journal of Soil Science, 62, 629e636. Kweon, G., Lund, E., & Maxton, C. (2013). Soil organic matter and cation-exchange capacity sensing with on-the-go electrical conductivity and optical sensors. Geoderma, 199, 80e89. Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33, 159e174. Leone, A. P., Viscarra-Rossel, R. A., Amenta, P., & Buondonno, A. (2012). Prediction of soil properties with PLSR and viseNIR spectroscopy: application to Mediterranean soils from Southern Italy. Current Analytical Chemistry, 8, 283e299.  zMarin-Gonzalez, O., Kuang, B., Quraishi, M. Z., Muno  & Mouazen, A. M. (2013). On-line measurement Garcı´a, M.A., of soil properties without direct spectral response in near infrared spectral range. Soil & Tillage Research, 132, 21e29. Moore, G., Dolling, P., Porter, B., & Leonard, L. (1998). Soil acidity. In G. Moore (Ed.), Soilguide. A handbook for understanding and managing agricultural soils. Agriculture Western Australia Bulletin No. 4343. Mouazen, A. M. (2006). Soil survey device. International publication published under the patent cooperation treaty (PCT). World Intellectual Property Organization, International Bureau. International Publication Number: WO2006/015463, PCT/ BE2005/000129, IPC: G01N21/00, G01N21/00. Mouazen, A. M., De Baerdemaeker, J., & Ramon, H. (2005). Towards development of on-line soil moisture content sensor using a fibre-type NIR spectrophotometer. Soil & Tillage Research, 80, 171e183. Mouazen, A. M., De Baerdemaeker, J., & Ramon, H. (2006). Effect of wavelength range on the measurement accuracy of some selected soil properties using visual-near infrared spectroscopy. Journal of Near Infrared Spectroscopy, 14(3), 189e199. Mouazen, A. M., Karoui, R., De Baerdemaeker, J., & Ramon, H. (2006). Characterization of soil water content using measured visible and near infrared spectra. Soil Science Society of America Journal, 70, 1295e1302. Mouazen, A. M., Karoui, R., Deckers, S., De Baerdemaeker, J., & Ramon, H. (2007). Potential of visible and near infrared spectroscopy to derive colour groups utilising the Munsell soil colour charts. Biosystems Engineering, 97(2), 131e143. Mouazen, A. M., Maleki, M. R., Cockx, L., Van Meirvenne, M., Van Holm, L. H. J., Merckx, R., et al. (2009). Optimum three-point linkage set up for improving the quality of soil spectra and the accuracy of soil phosphorous measured using an on-line visible and near infrared sensor. Soil & Tillage Research, 103(1), 144e152. Mouazen, A. M., Maleki, M. R., De Baerdemaeker, J., & Ramon, H. (2007). On-line measurement of some selected soil properties using a VISeNIR sensor. Soil & Tillage Research, 93, 13e27. Nelson, D. W., & Sommers, L. E. (1982). Total carbon, organic carbon, and organic matter. In A. L. Page (Ed.) (2nd

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005

b i o s y s t e m s e n g i n e e r i n g x x x ( 2 0 1 6 ) 1 e1 5

ed.,Agronomy monograph, 9Methods of soil analysis, part 2 (pp. 539e579). Madison, WI: ASA. Nocita, M., Stevens, A., Noon, C., & Wesemael, B. V. (2013). Prediction of soil organic carbon for different levels of soil moisture using ViseNIR spectroscopy. Geoderma, 199, 37e42. Quraishi, M. Z., & Mouazen, A. M. (2013). Calibration of an on-line sensor for measurement of topsoil bulk density in all soil textures. Soil & Tillage Research, 126, 219e228. Savitzky, A., & Golay, M. J. E. (1964). Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry, 36, 1627e1639. Savvides, A., Corstanje, R., Baxter, S. J., Rawlins, B. G., & Lark, R. M. (2010). The relationship between diffuse spectral reflectance of the soil and its cation exchange capacity is scale-dependent. Geoderma, 154, 353e358. Sherman, D. M., & Waite, T. D. (1985). Electronic spectra of Fe3þ oxides and oxyhydroxides in the near infrared to ultraviolet. American Mineralogist, 70, 1262e1269. Soriano-Disla, J. M., Janik, L. J., Viscarra-Rossel, R. A., MacDonald, L. M., & McLaughlin, M. J. (2014). The performance of visible, near-, and mid-infrared reflectance spectroscopy for prediction of soil physical, chemical, and biological properties. Applied Spectroscopy Reviews, 49, 139e186. Stenberg, B., Viscarra Rossel, R., Mouazen, A. M., & Wetterlind, J. (2010). Visible and near infrared spectroscopy in soil science. Advances in Agronomy, 107, 163e215. Tekin, Y., Kuang, B., & Mouazen, A. M. (2013). Potential of on-line visible and near infrared spectroscopy for measurement of pH

15

for deriving variable rate lime recommendations. Sensors, 13, 10177e10190. Tekin, Y., Tumsavas‚, Z., & Mouazen, A. M. (2012). Effect of moisture content on prediction of organic carbon and pH using visible and near-infrared spectroscopy. Soil Science Society of America Journal, 76(1), 188e198. Viscarra Rossel, R. A., & Behrens, T. (2010). Using data mining to model and interpret soil diffuse reflectance spectra. Geoderma, 158, 46e54. Viscarra-Rossel, R. A., & Chen, C. (2011). Digitally mapping the information content of visibleenear infrared spectra of surficial Australian soils. Remote Sensing of Environment, 115, 1443e1455. Viscarra-Rossel, R. A., Walvoort, D. J. J., McBratney, A. B., Janik, L. J., & Skjemstad, J. O. (2006). Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma, 131, 59e75. Waruru, B. K., Shepherd, K. D., Ndegwa, G. M., Kamoni, P. T., & Sila, A. M. (2014). Rapid estimation of soil engineering properties using diffuse reflectance near infrared spectroscopy. Biosystems Engineering, 121, 177e185. Zornoza, R., Guerrero, C., Mataix-Solera, J., Scow, K. M., Arcenegui, V., & Mataix-Beneyto, J. (2008). Near infrared spectroscopy for determination of various physical, chemical and biochemical properties in Mediterranean soils. Soil Biology and Biochemistry, 40, 1923e1930.

Please cite this article in press as: Ulusoy, Y., et al., Prediction of soil cation exchange capacity using visible and near infrared spectroscopy, Biosystems Engineering (2016), http://dx.doi.org/10.1016/j.biosystemseng.2016.03.005