Predicting soil organic matter from cellular phone images under varying soil moisture

Geoderma xxx (xxxx) xxxx

Contents lists available at ScienceDirect

Geoderma journal homepage: www.elsevier.com/locate/geoderma

Predicting soil organic matter from cellular phone images under varying soil moisture Yuanyuan Fua,b, Perry Tanejac, Shaomin Lind, Wenjun Jib,e, Viacheslav Adamchukb, ⁎ Prasad Daggupatic, Asim Biswasd, a

Key Laboratory of Quantitative Remote Sensing in Agriculture of Ministry of Agriculture, Beijing Research Center for Information Technology in Agriculture, Beijing 100097, China b Department of Bioresource Engineering, McGill University, 21111 Lakeshore Road, Ste Anne de Bellevue, Quebec H9X 3V9, Canada c School of Engineering, University of Guelph, 50 Stone Road East, Guelph, Ontario N1G 2W1, Canada d School of Environmental Sciences, University of Guelph, 50 Stone Road East, Guelph, Ontario, N1G 2W1, Canada e College of Land Science and Technology, China Agriculture University, Beijing 100083, China

A R T I C LE I N FO

A B S T R A C T

Handling Editor: Cristine L.S. Morgan

Soil organic matter (SOM) is considered as the backbone of soil health and soil quality. Thus, its’ estimation is critical to support the development of management decision including precision agriculture. To overcome challenges of laborious, rather expensive and time-consuming laboratory measurements, recent advances in image acquisition systems provided a new dimension of image-based SOM prediction. However, challenges remain in using soil images taken directly in the field due to variable soil surface conditions including vegetation cover, illumination, and soil moisture. Soil moisture can significantly influence soil color and thus confounds the relationship between SOM and soil color. This study quantifies the effects of soil moisture on the relationship between SOM and color parameters derived from cell phone images and establishes suitable SOM prediction models under varying conditions of soil moisture contents (SMCs). To simulate the continuous variation of soil moisture in the field, air-dried ground soil samples were saturated and allowed to dry naturally. Images were captured with a cellular phone over time representing various SMCs (set of images). Final set of images were captured on oven-dried samples. Images were preprocessed using illumination normalization to avoid illumination inconsistencies and segmentation technique to remove non-soil parts of the images including black cracks, leaf residues and specular reflection before modelling. Five color space models including RGB, HIS, CIELa*b*, CIELc*h* and CIELu*v* were used to quantify soil color parameters. Univariate linear regression models were developed between SOM and color parameters and an optimal set of color parameters that are capable of resisting variation in SMC was determined. It was observed that SMC exerted a considerable influence on SOM prediction accuracy when its value reached > 10%. The threshold of 10% SMC was considered as the critical SMC. Consequently, stepwise multiple linear regression (SMLR) models were developed for soil samples with SMC below and above the critical SMC. For the soil samples at below the critical SMC, the color parameter R based model produced satisfactory prediction accuracy for SOM with R2cv, RMSEcv, and RPDcv values of 0.936, 4.44% and 3.926, respectively. For the soil samples at above the critical SMC, the SOM predictive model including SMC as a predictor variable showed better accuracy (R2cv = 0.819, RMSEcv = 7.747%, RPDcv = 2.328) than that without including SMC (R2cv = 0.741, RMSEcv = 9.382%, RPDcv = 1.922). This study showed potential of cellular phone to be used as a proximal soil sensor fast, accurate and non-destructive estimation of SOM both in the laboratory and field conditions.

Keywords: Soil organic matter Soil moisture content Cell phone images Color space models Stepwise linear regression

1. Introduction Soil organic matter (SOM) is the organic matter component of soil, incorporating plant and animal residues. It is considered as the

backbone of soil’ health and regulates various physical, chemical and biological processes and properties (Sudarsan et al., 2016). It also outlines the capacity of soil to supply nitrogen, the crop growth determinant. However, like other properties, SOM is highly spatially

⁎

Corresponding author. E-mail addresses: [email protected] (P. Taneja), [email protected] (S. Lin), [email protected] (W. Ji), [email protected] (V. Adamchuk), [email protected] (P. Daggupati), [email protected] (A. Biswas). https://doi.org/10.1016/j.geoderma.2019.114020 Received 7 April 2019; Received in revised form 24 September 2019; Accepted 19 October 2019 0016-7061/ © 2019 Elsevier B.V. All rights reserved.

Please cite this article as: Yuanyuan Fu, et al., Geoderma, https://doi.org/10.1016/j.geoderma.2019.114020

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

variable within a field (Conant et al., 2011). The information on spatial variability of SOM can help decide site-specific management of agricultural resources including the application of nitrogen fertilizer and achieve the tradeoff between crop production increase and environment pollution reduction (McBratney and Pringle, 1999), the critical component of precision agriculture. Traditional procedures for estimating SOM are laborious, costly and require time-intensive spatially dense soil sampling and laboratory analysis (O'Halloran et al., 2004). These often restrict a detailed characterization of SOM and its spatial variability in the field. Furthermore, larger field sizes make detailed characterization unaffordable for many stakeholders. In the last decade, with the development of technology, various soil sensors have been used to characterize SOM. This is because SOM absorbs light across all visible and near-infrared wavelengths (Biehl et al. 1985; Ben-Dor 2002), and consequently, soil spectroscopy has shown potential to characterize SOM both in-situ and in laboratory conditions (Ishida and Ando, 1999; Hummel et al., 2001; Barnes et al., 2003; Gregory et al., 2006; Shi et al., 2015). A strong correlation between SOM and mid-infrared (MIR) spectra has also been reported (Rossel et al., 2006a; Dhawale et al., 2015). However, the related complex data collection and processing techniques and expensive equipment restrict a wider usage in practical agricultural production scenarios. Besides, SOM prediction accuracy is limited using vis-NIRMIR spectroscopy under soil conditions, like variable soil moisture and surface roughness (Nocita et al., 2013; Rienzi et al., 2014; Rodionov et al., 2014). With technological progression and the advancement of image acquisition systems, image-based soil characterization techniques have garnered significant attention of global soil science community. Unlike soil diffuse reflectance spectroscopy, image acquisition devices like digital cameras or even cameras’ in cellular phones are easily accessible. Available literature on image-based SOM or soil organic carbon (SOC) predictions (Chen et al., 2000; Gregory et al., 2006; Rossel et al., 2006b; Rossel et al., 2008), soil color was used as a proxy to link SOM or SOC with images. Soils with darker color are generally associated with higher OM contents and often considered as fertile and suitable for plant growth (Schulze et al., 1993). Traditionally, soil color is quantified by the Munsell colorimetric system (Munsell Color Company, 1994) that requires subjective visual matching between the soil samples and limited standard color chips. Thus, the Munsell colorimetric system is suitable when precise color measurements and automatic color matching are required. Good quality images from advanced image acquisition systems helped overcome the limitations of the Munsell colorimetric system and improved prediction of SOM/SOC. For example, Gregory et al. (2006) predicted SOM content in southwestern Ontario fields using a high-resolution digital camera system containing NIR, red (R), green (G) and blue (B) wavebands under both controlled and field conditions. They reported that SOM content was related to the image intensity measured in all wavebands without involving multiple soil types. Using a digital camera in the laboratory and under the ‘ideal’ lighting condition, Rossel et al. (2006b) observed that the CIEL*u*v* and CIEL*c*h* models were more suitable for predicting SOC. In a follow-up study, Rossel et al. (2008) reported a comparable prediction of SOC using simple color space models from a digital camera. Stronger relationships between soil color and SOM were by many others (Aitkenhead et al., 2012; Aitkenhead et al., 2015; Gregory et al., 2006; Rossel et al., 2006b). This relationship was taken to a step forward and a cell-phone application named SOCIT (only pertinent to mineral soils in Scotland) was developed by Aitkenhead et al. (2013). More recently, our team developed an algorithm to quantify SOM and soil texture from image parameters using geostatistical and regression-based methods (Sudarsan et al., 2016). These image-based SOM or SOC prediction studies (Chen et al., 2000; Gregory et al., 2006; Rossel et al., 2006b; Rossel et al., 2008, Sudarsan et al., 2016) directly used soil color parameters to develop prediction models without considering the contribution of other

confounding factors like soil moisture, surface residue, surface roughness and light (Ben-Dor et al., 2008). These factors are known to influence spectral response in the visible range of the electromagnetic spectrum (400–700 nm). Among these factors, soil moisture is the most important one that restricts practical in-situ measurement of SOM. Usually, dry soils are lighter in color than wet soils (Al-Abbas et al., 1972; Stevens et al., 2008). As soil moisture content (SMC) increases, soil micro and macropores are gradually filled with water and alter the physical structure of the soil. Consequently, the relative refractivity at the soil particle surface also changes causing the change in soil color (Nocita et al., 2013). The SMC, thus, complicates the relationship between SOM and soil color and becomes a key determinant factor for the practical use of image-based SOM prediction. SMC was implicitly involved in the predictive models developed in the above-mentioned studies as the SMCs of soil samples were variable in these studies. However, it’s influence on SOM prediction was not considered and hence, the given study was planned in which SMC was explicitly considered for SOM prediction. In addition, a few studies indirectly supported the concept of critical moisture content, a threshold SMC above or below which the pattern of changing soil color are different (Chen et al., 2000; Nocita et al., 2013). There may be various reasons that contribute to the development of this threshold SMC. For example, soil physical structures are generally different before and after water fully fills the micro and macropores of soil. Thus, the soil spectral responses will have low sensitivity to soil type with increasing SMC, especially at higher than critical SMCs (Chen et al., 2000). It suggested that SMC might follow different patterns to influence soil color above and below the critical SMC. Lee et al. (2003) reported a similar non-linear decrease in soil reflectance with increasing SMC. Lobell and Asner (2002) noticed that soil reflectance showed an obvious difference in the visible region only until volumetric SMC reached 20%. Similarly, Nocita et al. (2013) reported a critical gravimetric SMC of 15%. While there are evidences on the influence of critical SMC on soil reflectance in the visible range, how this may impact the prediction of SOM using digital images was not considered before. Thus, the objectives of this study were to (1) evaluate the ability of cell phone images to predict SOM using color parameters; (2) quantify the effect of soil moisture on the accuracy of SOM prediction models based on color parameters; and (3) determine the critical moisture content influencing SOM prediction accuracy on color parameters and establish suitable SOM prediction models accordingly. 2. Materials and methods 2.1. Experimental setup and image acquisition A set of 25 soil samples from a previous field experiment (Ji et al., 2016) was considered in this study. These samples were part of a set of 56 and 64 topsoil samples (0–15 cm) that were collected using a hand shovel from Field 26 and Field 86 respectively, of Macdonald Campus Farm, McGill University, Quebec, Canada (45°25′N, 73°56′W). Both fields demonstrated a high spatial (within the field) variability in soil types and in SOM (Ji et al., 2016). While larger set of samples would represent the variability in soil conditions more accurately, the set of 25 soil samples were selected in this pilot study to exhibit the wide variation in SOM (3.3–62.7%). The SOM was measured using the loss on ignition (LOI) method (Schulte and Hopkins, 1996). Soil samples were evenly (~8 mm thickness) placed in petri dishes for image acquisition. Images (2322 × 4128 pixels) were captured with a cellular phone 10-megapixel camera set to a holder 32 cm above the sample and the images were saved as Joint Photographic Experts Group (JPEG) standard compression. The images of air-dried soil samples were first collected and then water was sprayed gradually and carefully (without disturbing the surface) to saturate the soil. A set of images were collected at the saturated water content and then three more sets 2

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

normalization.

during the natural drying process of samples corresponding to different SMCs. Finally, the samples were oven-dried to get 0% SMC. The weight of the soil samples (including the petri dish) was recorded during each stage of image acquisition. Gravimetric SMC was calculated for each set of samples based on the difference in the weights with the oven dried samples. Due to unknown bulk density of soil samples, only gravimetric SMC was considered in this study. A total of six sets of images corresponding to six different levels of SMC were collected during the experiment. To facilitate illustration hereafter, the six sets of images were respectively named as images of “Group 1”, “Group 2”, “Group 3”, “Group 4”, “Group 5” and “Group 6” with increasing SMC. Therefore, the images of oven-dried soil samples formed the “Group 1”, images of air-dried soil samples of formed the “Group 2”, three sets of images collected during the natural drying process formed the “Group 3”, “Group 4” and “Group 5”. The saturated soil samples formed the “Group 6”. There were 146 images in all, with 25 images for each Group except 24 images in “Group 2” and “Group 6”, and 23 images for “Group 1”. These 4 images were corrupted during the data transfer and could not be recovered. In this study, the SMCs of samples in each “Group” were not held at the same level. It was different from the SMC settings of other studies in which soil samples had the same SMC at the same wetting level and thus abrupt bi- or tri-modal soil moisture distributions were generated (e.g. Nocita et al., 2013; Rienzi et al., 2014; Rodionov et al., 2014). However, SMC in a field is likely to follow a normal or quasi-normal distribution. Moreover, soil samples with varying levels of SOM will have different water-holding capacities and drying characteristics. Setting up a fixed SMC would thus, bias the image acquisition and would not represent actual field conditions.

2.2.3. Image segmentation Each soil image contained some non-soil objects. For example, on close observation, small black cracks and leaf residues with low pixel values were observed on the soil images of the first five Groups. Similarly, images of the “Group 6” exhibited high pixel values from the specular reflection of water on the soil surface (Fig. 2). Though they occupied a small portion of the whole ROI, these non-soil objects were scattered all over the images. An image histogram-based segmentation method was used in this study to separate soil from non-soil objects based on the frequency of pixel values. Pixel values whose counts were smaller than a specified threshold (set to 1000 after several trials) were considered as the non-soil materials representing cracks and leaf residues and were removed from the ROI. For the images in the “Group 6”, the threshold was set to 40 after several trials. The pixel values larger than this threshold were regarded as those of the specular reflection areas and were removed from the ROI. 2.2.4. Color space conversions While the soil images were stored in RGB color space model with primary color parameter R, G, and B, the conversion to other color space models provided multiple and often secondary color parameters with various levels of information. In this study, the soil ROIs with and without illumination normalization and segmentation were converted from RGB color space model to other color space models including CIEL*a*b*, CIEL*c*h*, CIEL*u*v* and HIS (decorrelated RGB) (Barron and Torrent, 1986; Ford and Roberts, 1998). Table 1 provides a brief introduction to each color parameter of the five color space models. For details on the algorithms used for converting the RGB images to the above-mentioned color space models, we refer to the work of Rossel et al. (2006a,b) and Ford and Roberts (1998). The mean pixel value of each color parameter was used in the subsequent analysis. Furthermore, the color parameters of the soil ROIs without illumination normalization and segmentation were used as benchmark to evaluate the effectiveness of illumination normalization and segmentation methods.

2.2. Image processing and analysis Before the images were analyzed, preprocessing was carried out with four components in the order; 1) region of interest (ROIs) selection, 2) illumination normalization, 3) image segmentation and 4) color space conversions.

2.3. Evaluation of the effectiveness of illumination normalization and image segmentation techniques

2.2.1. Region of interest (ROIs) selection The ROI selection involved cropping a) 800 × 800 pixels from the center of the petri dish for further analysis and b) 200 × 200 pixels from the white board on which the petri dishes were placed during image acquisition to use as a reference in illumination normalization. Fig. 1 shows the images along with their corresponding ROIs of two soil samples from each Group representing different SMCs.

Pearson correlation coefficients were used to examine the usefulness of color parameters and evaluate the effectiveness of illumination normalization and image segmentation techniques in the prediction of SOM. The color parameters derived from images without and with illumination normalization and image segmentation were used in the correlation analyses to see the usefulness and regression models to evaluate the predictability of SOM. The change in correlation coefficients between the SOM and the color parameters derived from original images, images after illumination normalization and images after segmentation was recorded to examine the improvements. Both univariate and multivariate linear regression analyses were adopted in this study to examine the predictive performance of color parameters from the original images, images after illumination normalization, images after segmentation and images after both illumination normalization and segmentation.

2.2.2. Illumination normalization The lighting conditions in the images were slightly variable in the same Group, while more variability was observed among different Groups (Fig. 1) as the photos were taken on different days. Thus, it was necessary to minimize the edge effects and account for variation due to light source. Illumination normalization was carried out by dividing each pixel of a soil ROI by the mean intensity value of its corresponding reference, band by band according to the Eq. (1).

DN (i , j )INλ = DN (i , j )ORIλ /MeanDNRefλ

(1)

whereDN (i, j )INλ andDN (i, j )ORIλ are the illumination normalized and original pixel value at the ith row and jth column of a soil ROI, respectively; MeanDNRefλ is the mean pixel value of the corresponding reference for specific waveband λ (only R, G, B as others were derived from them). While use of a bright white colored board/paper as reference may strongly reflect the light and saturate the color bands, the off-white colored board used in this study did not reflect light strongly. Moreover, the purpose was to have a constant reference during every image capture and the background board was suitable. However, future studies should choose a color palate or band as a reference material which may provide more robust reference material for illumination

2.4. Univariate linear regression to analyze the effect of varying SMC on SOM prediction using color parameters Simple univariate linear regression was used to examine the effect of SMC on SOM prediction using color parameters. Separate regression models were developed for each data Group (“Group 1” through “Group 6”) and the optimal color parameters were determined by comparing the model performances. The parameters which continued to offer satisfactory prediction accuracies in terms of model performance regardless of the variation in SMC were regarded as optimal color 3

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

Fig. 1. Images showing ROIs (large white square box) and reference regions (small white square box) in petri dishes for two soil samples with SOM content of 3.3% (top row) and 8% (bottom row) under six different soil wetting conditions; (a) “Group 1”, (b) “Group 2”, (c) “Group 3”, (d) “Group 4”, (e) “Group 5”, and (f) “Group 6”.

Fig. 2. Images showing the segmentation results of soil samples under six different soil wetting conditions. (a) “Group 1”, (b) “Group 2”, (c) “Group 3”, (d) “Group 4”, (e) “Group 5”, and (f) “Group 6”.

2.5. Determination of critical SMC and development of SOM prediction models

parameters. Due to limited size, the leave-one-out cross-validation (LOOCV) was used for each regression model to validate the established relationships due to its good capability in providing nearly unbiased predictions of the error (Efron and Gong, 1983; Schlerf et al., 2005). Once selected, the performance the optimal color parameters were tested using a calibration-validation dataset. For example, the univariate linear model developed for oven-dried samples (“Group 1” as calibration dataset) using optimal color parameters were tested on other soil samples from all groups (i.e. “Group 2” through “Group 6” as validation dataset). This was carried out to further quantify the need to account for SMC while developing SOM prediction models using color parameters as well as to study how this influences model prediction accuracy.

The relationship among SOM, SMC, and color parameters was examined using scatter plots. The change in the relationship was recorded based on the distribution of color parameters as a function of SOM and SMC. The critical SMC was determined based on the change in distribution of color parameters contributed to the significant variations in correlation. Thus, critical SMC represents the SMC at which the soil moisture begins to significantly influence color parameters. After the critical SMC was determined (10% SMC), the whole data set comprising of 146 images of soil samples (each representing different SMC) was divided into two groups; 1) SMC > 10% and 2) SMC < 10%. For each group, 70% images were selected using the Kennard-Stone algorithm (Kennard and Stone, 1969). In selecting the 4

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

Table 1 Brief description of color parameter in the five color space models. Color space models

Color parameter

Definition

RGB

R G B H I S L* a* b* L* c* h* L* u* v*

one primary color with primary stimulus occurring at 700 nm, combination of lightness and chromaticity (hue and chroma) one primary color with primary stimulus occurring at 546 nm, combination of lightness and chromaticity (hue and chroma) one primary color with primary stimulus occurring at 436 nm, combination of lightness and chromaticity (hue and chroma) hue, H = (2G − R − B)/4 lightness, I = (R + G + B)/3 chroma, S = (R − B)/2 lightness, range from black to white chroma, redness (positive a*) or greenness (negative a*) chroma, yellowness (positive b*) or blueness (negative b*) same definition with L* in CIEL*a*b* color model chroma, c* = sqrt(a*^2 + b*^2) hue, h* = arctan(b*/a*) same definition with L* in CIEL*a*b* color model chroma, redness (positive u*) or greenness (negative u*) chroma, yellowness (positive v*) or blueness (negative v*)

HIS

CIEL*a*b*

CIEL*c*h*

CIEL*u*v*

RPDv were calculated by using Eq. (3) too. In the study, we followed the RPD classification of Chang and Laird (2002): those with RPD > 2 was considered to be a good quantitative model.

images, the image color parameters were used as the criteria of selection and the Euclidian Distance was used as the distance matrix. The calibration (70%) and validation (30%) images were selected for each Group in order to quantify the effect of SOM with similar SMC. Unlike considering whole dataset with more images, calibration-validation for each Group would help to quantify the effect of SOM with narrow SMC range. Stepwise Multiple Linear Regression (SMLR) was used to develop relationship between SOM and color parameters (R, G, B, H, I, S, L*, a*, b*, c*, h*, u* and v*) for the samples below and above the critical SMC. Additionally, SMLR model was also developed between the SOM and the color parameters including SMC (R, G, B, H, I, S, L*, a*, b*, c*, h*, u*, v* and SMC) for the samples above the critical SMC. At each step of the forward SMLR, one color parameter with the most statistically significant contribution and lowest p value was added to the model and the change in the model p values and F-statistics was recorded. In the study, a threshold p-value to include and exclude a color parameter was set to 0.05. There were strong correlations between the color parameters (Table 2) and thus required to consider multicollinearity. Multicollinearity is the ratio of variance in a model with multiple terms, divided by the variance of a model with only one term. In this study, the variance inflation factor (VIF) was used to assess the magnitude of multicollinearity among the variables. Studies suggest that a VIF value of greater than 10 is an indication of severe multicollinearity and results in poor prediction (Myers, 2000; Darvishzadeh et al., 2008). In addition to the use of calibration and validation datasets, all the samples of each Group were also used to validate the corresponding calibrated SMLR model using LOOCV. Therefore, there were model statistics for three sets of data; calibration using 70% images, validation using 30% images and LOOCV using all images together except one image.

3. Results and discussion 3.1. Descriptive statistics The descriptive statistics of SMCs of various Groups and SOM showed a high degree of variation with the coefficient of variation, CV (%) between 38.74% and 192.37% (Table 3). SOM content varied between 3.3%) to 62.7% with an average of 19.70% and a standard deviation of 18.35% representing the wide variability of the study area. With an acceptable approximation, the SMCs for all Groups except Group 2 and Group 3 were normally distributed (kurtosis approximately between −3 and 3) (Table 3). The SMCs of Group 3 had a very large CV of about 192.37%. On the other hand, SMC of Group 6 comparatively varied less, with a CV of around 38.74%. For each variable, the range of values observed in the data is also presented in Table 2 to increase the understanding of Root Mean Square Error (RMSE). 3.2. Evaluation of the effectiveness of illumination normalization and image segmentation techniques Soil organic matter was negatively correlated with most of the color parameters for oven-dried soil samples (Fig. 3(a)). It was in line with the fact that SOM has a significantly negative correlation with soil reflectance (Al-Abbas et al., 1972). The darker color of soil was thus correlated to the amount of organic matter which consequently exhibited lower reflectance. Furthermore, the absolute values of the correlation coefficient between SOM and most of the color parameters were close to 1. This indicated that the color parameters shared a strong linear relationship with SOM, and this could be used for SOM prediction through modelling. Besides, the absolute values of correlation coefficients improved after illumination normalization for most of the color parameters. However, the improvements were not obvious for some color parameters. It could be because soil images were collected in the laboratory setting and the variation in lighting conditions was not drastic during image acquisition. The absolute values of correlation coefficients for color parameters a* and h* changed more prominently, than that of other color parameters. It indicated that color parameters a* and h* were more sensitive to the change in lighting conditions over other color parameters. While a* is the chroma with positive values representing redness and negative values representing greenness; h* is the hue and is calculated from a* and b* (h* = arctan(b*/a*)), where b* is also chroma with positive values representing yellowness and

2.6. Model validation and comparison The model performance was evaluated by the coefficient of determination (R2), root mean square error (RMSE, Eq. (2)) and ratio of prediction to deviation (RPD, Eq. (3)). The subscripts c, v, and cv represented the corresponding statistic for calibration, validation, and LOOCV. n

RMSEcv =

∑ (Yest − Ymea)2/n i=1

(2)

where Yest is the predicted SOM; Ymea is the measured SOM, and n is the number of samples. RMSEc and RMSEv were calculated by using Eq. (2) too.

RPDcv = SD /RMSEcv

(3)

where SD is the standard deviation of measured SOM values. RPDc and 5

Geoderma xxx (xxxx) xxxx

1.00 (1.00) 0.95 (0.96) 1.00 (1.00) 0.54 (-0.73) 0.73 (-0.74)

1.00 (1.00)

Table 3 Descriptive statistics for SOM (%) and SMC (%) of each Group.

SOM % SMC% (Group SMC% (Group SMC% (Group SMC% (Group SMC% (Group SMC% (Group

1.00 0.52 0.98 0.93 1.00 0.99 0.94 0.90 0.99 0.85 0.99 0.16 0.87 0.84 0.78 0.89 0.98 R G B H I S L* a* b* c* h* u* v*

(1.00) (0.98) (0.80) (0.68) (0.97) (0.86) (0.99) (0.62) (0.87) (0.83) (-0.74) (0.92) (0.97)

1.00 0.98 0.90 1.00 0.78 1.00 0.03 0.81 0.77 0.83 0.83 0.95

(1.00) (0.90) (0.70) (1.00) (0.74) (1.00) (0.45) (0.76) (0.71) (-0.69) (0.82) (0.91)

1.00 (1.00) 0.83 (0.49) 0.98 (0.91) 0.62 (0.38) 0.97 (0.88) −0.13 (0.12) 0.66 (0.40) 0.62 (0.35) 0.83 (-0.48) 0.70 (0.51) 0.85 (0.64)

1.00 (1.00) 0.89 (0.66) 0.78 (0.63) 0.90 (0.69) −0.06 (0.12) 0.83 (0.69) 0.79 (0.61) 0.86 (-0.55) 0.76 (0.56) 0.91 (0.76)

1.00 0.77 1.00 0.04 0.80 0.76 0.82 0.83 0.94

(1.00) (0.73) (1.00) (0.45) (0.74) (0.69) (-0.68) (0.81) (0.90)

1.00 0.79 0.58 1.00 1.00 0.51 0.99 0.94

(1.00) (0.78) (0.85) (1.00) (0.98) (-0.75) (0.98) (0.94)

1.00 0.06 0.82 0.78 0.82 0.84 0.95

(1.00) (0.50) (0.79) (0.74) (-0.71) (0.85) (0.93)

1.00 (1.00) 0.51 (0.80) 0.57 (0.84) −0.32 (-0.56) 0.56 (0.86) 0.29 (0.67)

1.00 1.00 0.57 0.98 0.95

(1.00) (0.98) (-0.75) (0.97) (0.95)

c* b* a* L* S I H B G R Color Parameters

Table 2 Correlation coefficients between the color parameters for SMC < = 10% and SMC > 10% (in bracket).

1) 2) 3) 4) 5) 6)

Mean

SD

Range

Kurtosis

CV (%)

19.7 0.00 2.54 2.96 29.57 49.88 60.76

18.36 0.00 3.39 5.70 17.21 20.76 23.54

3.30–62.70 0.00–0.00 0.48–13.33 0.24–22.58 8.21–77.87 26.95–104.95 33.98–119.6

0.14 0.00 4.04 8.26 1.60 0.60 0.06

93.17 0.00 133.68 192.37 58.20 41.62 38.74

negative values representing blueness (Table 1). Slight change in the lighting conditions can have stronger influence on hue (color) and thus the chroma, the saturation or dominance of colors. For example, the obvious change in a* was observed for soil samples that more dry (Group 1 through 3) contributing to the change in color of the soil. However, obvious change in h* was observed for soil samples that had more moisture content (Group 4 through 6) (Fig. 3). The correlation analyses between SOM and color parameters derived from images without and with segmentation were studied for each Group. There were no obvious improvements in the absolute correlation coefficients between SOM and color parameters after segmentation, except for “Group 6” (Fig. 4). For the first five Groups, it can be explained by the fact that the cracks with low pixel values and leaf residues with high pixel values only took up small regions in the soil images as the soil samples were collected after removing the residues form the surface, then air-dried, ground and sieved. Besides, the differences in pixel values resulting from them might have neutralized each other during averaging of color parameter. Unlike other groups, the “Group 6” with the specular reflection of water resulted in relatively large regions with very high pixel values in the soil images. The large differences in pixel values between soil and non-soil regions may not have neutralized during averaging of color parameter. Although the image segmentation did not improve the correlation between SOM and color parameters for the first five Groups, the technique would bring a great value in improving the correlation with color parameters if the images are collected in-situ at the presence of residues. Laboratory soil preparation in this study almost removed all the leaf residues and did not show much cracks. Therefore, further analyses were carried out with image segmentation with the view of extending the protocol to field collected soil images. In general, with the increase of SMC, the absolute values of correlation coefficients between SOM and color parameters decreased (Fig. 5(a)), while they increased between SMC and color parameters (Fig. 5(b)). The color parameters R, G, I, L*, u* and v* had a stronger correlation with SOM over other color parameters. These results were consistent with other studies that also reported strong correlations between these color parameters and SOC or SOM (Gregory et al., 2006; Rossel et al., 2006b, 2008). Comparing the correlation coefficients for SMC and SOM with color parameters, there were generally stronger correlations for SOM with color parameters. It indicated that SOM still accounted for a larger proportion than SMC in influencing soil color measurements. A comparative analysis of the predictive performance of selected color parameters (R, G, B) derived from the original images, images after illumination normalization, images after segmentation and images after both illumination normalization and segmentation using univariate linear regression and multivariate linear regression (Table 4). Except for the images from Group 6 (representing samples from saturation moisture content), a considerable improvement was observed (Table 4) in the values of Rcv2 and decrease in the values of RMSE for univariate linear regression models was observed when parameters were derived from images after illumination normalization than that from the original images. Similarly, a slight improvement in the Rcv2 values combined with a minor decrease in RMSE values were also

(1.00) (-0.65) (0.96) (0.92)

u* h*

v*

Y. Fu, et al.

6

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

Fig. 3. Correlation coefficients between SOM and color parameters derived from images without (Without IN) and with illumination normalization (With IN) for (a) “Group 1”, (b) “Group 2”, (c) “Group 3”, (d) “Group 4”, (e) “Group 5”, and (f) “Group 6”.

In general, the comparison analysis showed that both correlation coefficients and prediction accuracies were improved when color parameters were derived from images that were both illumination normalized and segmented. Hence, it can be inferred that illumination normalization and segmentation techniques were effective and imperative before using color parameters for SOM prediction through modelling. The subsequent analyses were therefore conducted on soil images that were both illumination normalized and segmented. However, as the lighting conditions were not varied in this study, further studies are required to verify the effectiveness of the illumination normalization method under field conditions with uncontrolled and large variation in lighting conditions. 3.3. Univariate linear regression to analyze the effect of varying SMC on SOM prediction using color parameters

Fig. 4. Correlation coefficients between SOM and color parameters derived from images without and with segmentation for the “Group 6”.

A comparative analysis of the predictive performance of color parameters for SOM prediction using univariate linear regression is presented in Fig. 6. The RPD values of univariate color parameter-based models for Group 1 were > 2 for color parameters including R, G, B, H, I, L*, h*, u*, and v* individually (Fig. 6). Among these, the model based on color parameter L* performed the best (Fig. 6). However, the ability of all color parameters to predict SOM degraded with the increasing mean and SD of SMC, although, this degradation was not proportional to the variation in SMC. For the first three Groups, in which the mean and SD of SMC was no

observed when the image parameters were derived from segmented images than that from the original images. Similarly, a further increase in Rcv2 values and a small decrease in RMSE values was observed when parameters were derived from images that were both illumination normalized and segmented in comparison with original images, images that were only illumination normalized or only segmented. A similar trend was also observed for multivariate linear regression models with first two color parameters R, G; R, B; G, B taken together and then all three-color parameters combined. 7

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

Fig. 5. Correlation coefficients between color parameters and a) SOM and b) SMC for each Group.

Fig. 6. Performance of color parameters including a) RMSE and b) RPD based on univariate linear regression models developed for each Group.

Table 4 Correlation coefficients and RMSE (in bracket) between three selected color parameters (R, G and B) from one channel on the prediction of SOM using image parameters derived from the original, illumination normalized, segmented and segmented and illumination normalized images together. Group

Parameter

Original

IN

1

R G B R G B R G B R G B R G B R G B

0.95 0.97 0.90 0.93 0.93 0.83 0.94 0.95 0.85 0.79 0.84 0.74 0.90 0.90 0.60 0.45 0.14 0.01

0.98 0.98 0.95 0.95 0.95 0.88 0.96 0.97 0.92 0.84 0.88 0.82 0.90 0.91 0.69 0.40 0.17 0.00

2

3

4

5

6

(0.09) (0.08) (0.13) (0.11) (0.11) (0.17) (0.10) (0.10) (0.16) (0.19) (0.16) (0.21) (0.13) (0.13) (0.26) (0.30) (0.38) (0.00)

Seg (0.07) (0.06) (0.09) (0.09) (0.09) (0.14) (0.08) (0.07) (0.12) (0.16) (0.14) (0.17) (0.13) (0.12) (0.23) (0.31) (0.37) (0.00)

0.95 0.97 0.91 0.93 0.93 0.83 0.94 0.95 0.87 0.79 0.85 0.75 0.85 0.85 0.70 0.67 0.37 0.00

color parameter L* was related to lightness. These results were consistent with Rossel et al. (2006b) who pointed out that SOC was well correlated with the lightness parameters of different color space models. A reasonable SOC prediction accuracies can be obtained using color parameters R and L*. Considering that the color parameters R and L* based univariate linear models had a good predictive performance for each Group, these were, therefore, identified as optimal color parameters and were used to further quantify the effect of SMC on SOM prediction. To further quantify the effect of SMC on SOM prediction, the individual calibration models based on the color parameters R and L* developed using oven-dried soil samples (or “Group 1”) were tested on validation sets of wetted soil samples (“Group 2” through “Group 6”). Low prediction accuracies were obtained when both the models were applied to predict moist samples, especially for the latter three Groups (“Group 4” through “Group 6”) (Table 5). For example, the calibration model for color parameter L* resulted in RMSEv and RPDv values of 59.421% and 0.308, respectively, exhibiting a poor performance. Overall, the accuracy of models decreased with increasing range and variation of SMC in the validation datasets. It was also evident that the decrease was much larger as compared to the decrease when models were calibrated and validated on each Group individually and for all color parameters (Fig. 6 and Table 5). It indicated that the SOM prediction models calibrated for dry samples cannot be directly used to predict the SOM contents of moist samples, especially for the soil samples with relatively high SMCs. It also supports that not only SOM, SMC influences color parameters and should be considered when developing SOM prediction models.

Seg and IN (0.09) (0.07) (0.12) (0.11) (0.11) (0.17) (0.10) (0.09) (0.15) (0.19) (0.16) (0.20) (0.16) (0.16) (0.22) (0.23) (0.33) (0.00)

0.97 0.98 0.96 0.95 0.95 0.88 0.96 0.97 0.92 0.85 0.88 0.83 0.86 0.86 0.77 0.69 0.43 0.01

(0.07) (0.06) (0.09) (0.09) (0.09) (0.14) (0.08) (0.07) (0.11) (0.16) (0.14) (0.17) (0.16) (0.15) (0.20) (0.22) (0.31) (0.00)

more than 2.96% and 5.7% respectively, the color parameters R, G, B, I, L*, u* and v* based univariate linear models exhibited RPDcv > 2 signifying better predictive performance, relative to the models based on other color parameters. For the later three Groups, with relatively high variations in SMC, only color parameters R and L exhibited RPDcv values > 2. This indicated that the color parameters R and L* had better capability in resisting against variations in SMC compared to the other color parameters. The color parameter R contained combined information on both lightness and chromaticity (chroma and hue), and 8

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

well as resemblance to similar studies, an SMC of 10% was determined to be the critical SMC. Further analysis was carried out on two groups of samples at below and above the critical SMC. After the critical SMC was determined, whole set 146 images was divided into two sets; one including images with SMC less than or equal to the critical SMC and the other one including those with a SMC greater than the critical SMC. The 1st set contained 70 (SMC < = 10%) and the 2nd set contained 76 images (SMC > 10%). For the set with SMC < = 10%, the predictive model was log SOM = 2.2447–3.1038R. In the SLMR, the only color parameter selected in the predictive model was R regardless of including SMC as a predictor variable. For the set with SMC > 10%, the predictive model was log SOM = 0.8220 − 0.0928c* + 0.0016 h* + 0.0557u* + 0.0058SMC. The color parameters c* and h* were included in the model when SMC wasn’t included as predictor variable, while the SMC and color parameters c*, h* and u* were included in the model when SMC was used as a predictor variable. It was observed that even with VIF > =10, none of the selected variables were collinear. Overall, the model developed for the set with SMC < = 10% (based on color parameter R) performed better with higher prediction accuracy compared to that developed for samples with SMC > 10% (Table 6). These indicated that SMC was a limiting factor to SOM prediction based on color parameters. SMC was not included in the model during the variable selection of SMLR for the Group with SMC < = 10%. It suggested that SMC had a negligible effect on SOM prediction at lower values of SMC. For the set with SMC > 10%, an increase in R2cv value by 0.078 and a corresponding reduction in the RMSEcv value by 1.636% was observed for the model including SMC as a predictor over the model excluding SMC as a predictor (Table 4). This indicated that SOM prediction accuracy can be improved by including SMC in the model for soil samples possessing greater amounts of moisture. This finding was in line with that observed by Rodionov et al. (2014). They developed a moisture-dependent SOC prediction model using PLSR for vis-NIR reflectance and obtained an improved prediction accuracy for field-moist undisturbed soil samples. In laboratory conditions, the SMC of air-dried soil samples is usually less than 10% and SOM models based on color parameters can obtain satisfactory prediction accuracies. However, under field conditions, the SMC is usually above 10% and the range of its variation is also found to be more than 10%. For instance, Hawley et al. (1983) reported an SMC range between 15.1% and 34.4% in eight different watershed areas in Oklahoma. Thus, with the availability of portable soil moisture sensors such as TDR based instruments, the proposed SMC-dependent SOM prediction model based on cell phone images confers great potential for SOM prediction under both laboratory and field conditions.

Table 5 Performance of the color parameters R and L based univariate linear models developed for oven-dried samples (“Group 1”) after validation on the latter five wetted Groups.

Calibration (Group 1)

Validation (Group 2)

Validation (Group 3)

Validation (Group 4)

Validation (Group 5)

Validation (Group 6)

R2c RMSEc% RPDc RPIQc R2v RMSEv% RPDv RPIQv R2v RMSEv% RPDv RPIQv R2v RMSEv% RPDv RPIQv R2v RMSEv% RPDv RPIQv R2v RMSEv% RPDv RPIQv

R

L*

0.984 2.369 8.056 11.196 0.941 6.338 2.955 4.008 0.954 4.094 4.483 5.764 0.872 30.590 0.600 0.771 0.809 43.657 0.420 0.541 0.791 54.921 0.333 0.381

0.978 3.068 6.221 8.646 0.936 6.656 2.813 3.816 0.949 7.152 2.566 3.300 0.901 30.681 0.598 0.769 0.816 44.211 0.415 0.534 0.768 59.421 0.308 0.353

3.4. Determination of critical SMC and development of SOM prediction models The scatter plot between SOM and color parameter R as an example for all the Groups is shown in Fig. 7. There was a negative correlation between R and SOM for the first three Groups. However, for the latter three Groups, there was an increase in SOM without much decrease in the R values. Similar trends were observed between SOM and other color parameters (e.g. G and B). The change in correlation indicated that the negative correlation between SOM and color parameters cannot be held with increasing SMC. Also, SMC influences color parameters-based SOM prediction differently at below and above a specific SMC. The decrease in SOM prediction accuracy was not obvious for the first three Groups. The SMC for 95% of soil samples of these Groups was < 10%. Similar patterns were also observed in other studies. For instance, Nocita et al. (2013) grouped soil samples with a gravimetric SMC > =15% and developed a SOC prediction model with good accuracy. Rienzi et al. (2014) demonstrated that predicting SOC over a range of 10% soil moisture variability did not substantially change prediction quality. Thus, based on the pattern observed in this study as

4. Conclusions Soil exhibits spectral reflectance characteristics in the visible part of the spectrum which can be used to predict various soil properties through modeling. The concept was used to save a lot of labor- and time-intensive and costly soil analysis. This study explored the feasibility of soil images captured using a cell phone to predict SOM under varying soil moisture conditions. The color parameters derived from the images demonstrated high correlation with both SOM and SMC. Soil moisture was a limiting factor in the use of cell phone images for SOM prediction under varying SMC. Good prediction accuracies were observed for drier soil samples or soil samples at below the critical SMC of 10% identified in this study. This could be highly suitable for laboratory estimation of SOM. However, field soil samples may vary largely in SMC and should be considered in the prediction model. Combining the current study result with a soil moisture sensor will provide quick estimation tool of SOM in field conditions. Thus, our findings serve as an informative guide on how to use cell phone images for SOM prediction. However, further studies are recommended for predicting more soil properties from various soil and climatic regions to dispatch this as a

Fig. 7. Scatter plots between SOM and color parameter R under varying soil moisture. 9

Geoderma xxx (xxxx) xxxx

Y. Fu, et al.

Table 6 Performance of SMLR SOM prediction models considering SMC. SMC < = 10%

Without involving SMC Involving SMC

SMC > 10%

R2cv

RMSEcv%

RPDcv

RPIQcv

R2cv

RMSEcv%

RPDcv

RPIQcv

0.936 0.936

4.440 4.440

3.926 3.926

4.527 4.527

0.741 0.819

9.382 7.747

1.922 2.328

2.350 2.846

matter content in southwestern Ontario fields using imagery from high-resolution digital cameras. Can. J. Soil Sci. 86, 573–584. Hawley, M.E., Jackson, T.J., McCuen, R.H., 1983. Surface soil moisture variation on small agricultural watersheds. J. Hydrol. 62 (1), 179–200. Hummel, J.W., Sudduth, K.A., Hollinger, S.E., 2001. Soil moisture and organic matter of surface and subsurface soils using an NIR soil sensor. Comput. Electron. Agric. 32, 149–165. Ishida, T., Ando, H., 1999. Use of disjunctive cokriging to estimate soil organic matter from Landsat Thematic Mapper image. Int. J. Remote Sens. 20, 1549–1565. Ji, W., Adamchuk, V.I., Biswas, A., Dhawale, N.M., Sudarsan, B., Zhang, Y., Rossel, R.A.V., Shi, Z., 2016. Assessment of soil properties in situ using a prototype portable MIR spectrometer in two agricultural fields. Biosyst. Eng. 152, 14–27. Kennard, R.W., Stone, L.A., 1969. Computer aided design of experiments. Technometrics 11 (1), 137–148. Lee, W.S., Sanchez, J.F., Mylavarapu, R.S., Choe, J.S., 2003. Estimating chemical properties of Florida soils using spectral reflectance. Trans. ASAE 46 (5), 1443. Lobell, D.B., Asner, G.P., 2002. Moisture effects on soil reflectance. Soil Sci. Soc. Am. J. 66, 722–727. McBratney, A.B., Pringle, M.J., 1999. Estimating average and proportional variograms of soil properties and their potential use in precision agriculture. Precis. Agric. 1, 125–152. Munsell Color Company, 1994. Munsell Soil Color Charts, revised ed. Macbeth Division of Kollmorgen, New Windsor, NY. Myers, R.H., 2000. Classical and Modern Regression with Applications (Duxbury Classic). Duxbury Press, Pacific Grove. Nocita, M., Stevens, A., Noon, C., van Wesemael, B., 2013. Prediction of soil organic carbon for different levels of soil moisture using Vis-NIR spectroscopy. Geoderma 199, 37–42. O'Halloran, I.P., Von Bertoldi, A.P., Peterson, S., 2004. Spatial variability of barley (Hordeum vulgare) and corn (Zea mays L.) yields, yield response to fertilizer N and soil N test levels. Can. J. Soil Sci. 84, 307–316. Rienzi, E.A., Mijatovic, B., Mueller, T.G., Matocha, C.J., Sikora, F.J., Castrignanò, A., 2014. Prediction of soil organic carbon under varying moisture levels using reflectance spectroscopy. Soil Sci. Soc. Am. J. 78, 958–967. Rodionov, A., Pätzold, S., Welp, G., Damerow, L., Amelung, W., 2014. Sensing of soil organic carbon using visible and near-infrared spectroscopy at variable moisture and surface roughness. Soil Sci. Soc. Am. J. 78 (3), 949–957. Rossel, R.V., Walvoort, D.J.J., McBratney, A.B., Janik, L.J., Skjemstad, J.O., 2006a. Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma 131, 59–75. Rossel, R.V., Minasny, B., Roudier, P., McBratney, A.B., 2006b. Colour space models for soil science. Geoderma 133, 320–337. Rossel, R.V., Fouad, Y., Walter, C., 2008. Using a digital camera to measure soil organic carbon and iron contents. Biosyst. Eng. 100, 149–159. Schlerf, M., Atzberger, C., Hill, J., 2005. Remote sensing of forest biophysical variables using HyMap imaging spectrometer data. Remote Sens. Environ. 95, 177–194. Schulte, E.E. and Hopkins, B.G., 1996. Prediction of organic matter by Weight Loss-OnIgnition Soil Organic Matter: Analysis and Interpretation. Magdoff FR, Tabatsai MA, Hanlon EA (eds). SSSA Inc., Madison, Wis Special Publication, 46 ,21–31. Schulze, D.G., Nagel, J.L., Van Scoyoc, G.E., Henderson, T.L., Baumgardner, M.F., Scott, D.E., 1993. Significance of organic matter in determining soil colors. In: Bigham, J.M., Ciolkosz, E.J. (Eds.), Soil Color. Soil Science Society of America, Madison, WI, pp. 71–90. Shi, Z., Ji, W., Viscarra Rossel, R.A., Chen, S., Zhou, Y., 2015. Prediction of soil organic matter using a spatially constrained local partial least squares regression and the Chinese vis–NIR spectral library. Eur. J. Soil Sci. 66, 679–687. Stevens, A., van Wesemael, B., Bartholomeus, H., Rosillon, D., Tychon, B., Ben-Dor, E., 2008. Laboratory, field and airborne spectroscopy for monitoring organic carbon content in agricultural soils. Geoderma 144, 395–404. Sudarsan, B., Ji, W., Biswas, A., Adamchuk, V., 2016. Microscope-based computer vision to characterize soil texture and soil organic matter. Biosyst. Eng. 152 (1), 41–50.

widespread technique for the estimation of soil properties. Additionally, current study lighting condition did not include high variability and thus the illumination normalization was not examined well. Further studies are warranted to include the field images with high variability in lighting conditions and soil moisture status. Acknowledgement This project was supported Natural Sciences and Engineering Research Council (NSERC) of Canada (RGPIN-2014-4100) granted to Biswas and funds provided through the Merit Scholarship Program for Foreign Students granted to Fu by the Quebec Fund for Research on Nature and Technology. The authors are also grateful to Hsin-Hui Huang, Francisco de la Macorra and Brett Bennett for their assistance with data collection. References Aitkenhead, M.J., Coull, M.C., Towers, W., Hudson, G., Black, H.I.J., 2012. Predicting soil chemical composition and other soil parameters from field observations using a neural network. Comput. Electron. Agric. 82, 108–116. Aitkenhead, M.J., Donnelly, D., Coull, M., Black, H., 2013. E-SMART: environmental sensing for monitoring and advising in real-time. IFIP Adv. Inf. Commun. Technol. 413, 129–142. Aitkenhead, M.J., Donnelly, D., Sutherland, L., Miller, D.G., Coull, M.C., Black, H.I.J., 2015. Predicting Scottish topsoil organic matter content from colour and environmental factors. Eur. J. Soil Sci. 66, 112–120. Al-Abbas, A.H., Swain, H.H., Baumgardner, M.F., 1972. Relating organic matter and clay content to multispectral radiance of soils. Soils Sci. 114, 477–485. Barnes, E.M., Sudduth, K.A., Hummel, J.W., Lesch, S.M., Corwin, D.L., Yang, C., Daughtry, C.S., Bausch, W.C., 2003. Remote-and ground-based sensor techniques to map soil properties. Photogramm. Eng. Remote Sens. 69, 619–630. Barron, V., Torrent, J., 1986. Use of the Kubelka—Munk theory to study the influence of iron oxides on soil colour. J. Soil Sci. 37, 499–510. Ben-Dor, E., 2002. Quantitative remote sensing of soil properties. Adv. Agron. 75, 173–243. Ben-Dor, E., Taylor, R.G., Hill, J., Demattê, J.A.M., Whiting, M.L., Chabrillat, S., Sommer, S., 2008. Imaging spectrometry for soil applications. Adv. Agron. 97, 321–392. Biehl, L.L., Stoner, E., 1985. Reflectance properties of soils. Adv. Agron 38, 1–44. Chang, C.W., Laird, D.A., 2002. Near-infrared reflectance spectroscopic analysis of soil C and N. Soil Sci. 167, 110–116. Chen, F., Kissel, D.E., West, L.T., Adkins, W., 2000. Field-scale mapping of surface soil organic carbon using remotely sensed imagery. Soil Sci. Soc. Am. J. 64, 746–753. Conant, R.T., Ogle, S.M., Paul, E.A., Paustian, K., 2011. Measuring and monitoring soil organic carbon stocks in agricultural lands for climate mitigation. Front. Ecol. Environ. 9 (3), 169–173. Darvishzadeh, R., Skidmore, A., Schlerf, M., Atzberger, C., Corsi, F., Cho, M., 2008. LAI and chlorophyll estimation for a heterogeneous grassland using hyperspectral measurements. ISPRS J. Photogramm. Remote Sens. 63, 409–426. Dhawale, N.M., Adamchuk, V.I., Prasher, S.O., Viscarra Rossel, R.A., Ismail, A.A., Kaur, J., 2015. Proximal soil sensing of soil texture and organic matter with a prototype portable mid-infrared spectrometer. Eur. J. Soil Sci. 66, 661–669. Efron, B., Gong, G., 1983. A leisurely look at the bootstrap, the jackknife, and crossvalidation. Am. Statist. 37, 36–48. Ford, A., Roberts, A., 1998. Colour Space Conversions. Westminster University, London, pp. 1–31. Gregory, S.D.L., Lauzon, J.D., O'Halloran, I.P., Heck, R.J., 2006. Predicting soil organic

10