Using a mobile real-time soil visible-near infrared sensor for high resolution soil property mapping

Using a mobile real-time soil visible-near infrared sensor for high resolution soil property mapping

Geoderma 199 (2013) 64–79 Contents lists available at SciVerse ScienceDirect Geoderma journal homepage: www.elsevier.com/locate/geoderma Using a mo...

4MB Sizes 3 Downloads 59 Views

Geoderma 199 (2013) 64–79

Contents lists available at SciVerse ScienceDirect

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

Using a mobile real-time soil visible-near infrared sensor for high resolution soil property mapping Masakazu Kodaira ⁎, Sakae Shibusawa Institute of Agriculture, Tokyo University of Agriculture and Technology, 3-5-8 Saiwai-cho, Fuchu-shi, Tokyo 183-8509, Japan

a r t i c l e

i n f o

Article history: Received 8 November 2011 Received in revised form 22 September 2012 Accepted 28 September 2012 Available online 20 November 2012 Keywords: Real-time soil sensor Vis–NIR spectroscopy PLS regression High-resolution soil map Site-specific soil management Precision farming

a b s t r a c t In this study, we developed twelve spectroscopic models based on visible and near-infrared (Vis–NIR: 305–1700 nm) soil reflectance spectra to predict and map at a high spatial resolution soil properties that are useful for site-specific soil management and precision agriculture. We collected using a real-time soil sensor (RTSS) with a differential global positioning system (DGPS). The investigated soil properties were moisture content (MC), soil organic matter (SOM), pH, electrical conductivity (EC), cation exchange capacity (CEC), total carbon (C-t), ammonium nitrogen (N-a), hot water extractable nitrogen (N-h), nitrate nitrogen (N-n), total nitrogen (N-t), available phosphorus (P-a), and phosphorus absorptive coefficient (PAC). The experimental site is a commercial upland field with alluvial soil located in Hokkaido, Japan. To develop the calibration models, 144 soil spectra were collected with the Vis–NIR spectrometer in the RTSS. Partial least squares regression (PLSR) coupled with full (leave-one-out) cross-validation were used to establish the relationships between the Vis–NIR soil reflectance spectra and the soil properties, whose values were obtained by soil chemical analysis. We show the coefficient of correlation, coefficient of determination (R2), root mean square error and residual prediction deviation (RPD). The accuracy of the spectroscopic models ranged from R 2 0.45 to 0.93 and RPD from 1.0 to 3.6. Our results were compared with previous studies, which include field-based and lab-based results. The accuracy of our predictions as measured by the RPD for MC, SOM, CEC, C-t, N-a, N-n, N-t, P-a, and PAC was similar or better than those obtained in previous studies. RPD results for pH, EC and N-h were only slightly poorer. © 2012 Elsevier B.V. All rights reserved.

1. Introduction In precision agriculture, rapid, non-destructive, cost-effective and convenient soil analysis techniques are needed for soil management, crop quality control using fertilizer, manure and compost, and variablerate input for soil variability in a field. Conventional soil management involves the following steps: 1) a soil sample is prepared using samples collected from a field or an individual crop compartment; 2) a representative value of several soil properties (e.g., N, P, K, pH) is obtained by soil chemical analysis for field crop management; and 3) the amount of fertilizer, manure and compost input is adjusted for each field or crop compartment as field crop management. However, to reduce fertilizer costs, increase crop production and achieve environmental impact reduction, changing the fertilizer application is not sufficient. Also, to create several high-resolution soil maps by conventional soil management procedure, a huge number of soil samples must be collected according to the many separate grids of a field or crop compartment.

⁎ Corresponding author. Tel./fax: +81 42 367 5762. E-mail address: [email protected] (M. Kodaira). 0016-7061/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.geoderma.2012.09.007

The high cost, time-consuming soil chemical analysis, and laborious collection of soil samples are significant problems. In an attempt to solve these problems, different types of soil sensing systems have been developed for application to site-specific management in precision agriculture. In a report by Adamchuk et al. (2004), most of these ‘on-the-go’ soil sensors are based on electrical or electromagnetic sensing. However, they are not able to measure multiple soil properties. In contrast, optical and radiometric sensors are promising candidates for estimating multiple soil properties from reflectance spectra data. These ‘on-the-go’ soil sensors might be less accurate than conventional laboratory methods, but are still advantageous due to their highresolution spatial information. Optical and radiometric sensors also require a higher initial equipment cost compared to the other sensors. Furthermore, the analysis results are based on chemometrics (Norris, 1989). Due to a lack of understanding of exactly how optical and radiometric sensors relate to composition and absorption wavelength, this type of sensing has not gained credibility. Recent studies that have analyzed the relationship between the absorption band of soil chemical analysis and chemometrics methods (e.g., multiple linear regression (MLR), principal component regression (PCR) and partial least squares regression (PLSR: Wold, 1975; Wold et al., 1983, 2001)) are gradually resolving this ambiguity (Shonk et al., 1991; Viscarra Rossel et al., 2006).

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

In addition, recent mathematical analyses and advanced chemometric techniques are worthy of mention, e.g., artificial neural networks (ANN) (Daniel et al., 2003; Viscarra Rossel and Behrens, 2010), back propagation neural network (BPNN) analysis (Mouazen et al., 2010), moving window partial least squares regression (MWPLSR) (Jiang et al., 2002), changeable size moving window partial least squares (CSMWPLS) (Du et al., 2004), searching combination moving window partial least squares (SCMWPLS) (Du et al., 2004) and genetic algorithm-based wavelength selection (GAWLS) (Kawamura et al., 2006). These techniques for clarifying the relationship between the spectral absorption wavelength and soil chemical composition (e.g., absorption of C\H, O\H and N\H bonds) make it possible to select the appropriate wavelength region. Therefore, to estimate many soil properties, optical and radiometric sensing is a proximal technique compared with conventional chemical analysis. The potential and accuracy of Vis–NIR diffuse reflectance spectroscopy for soil chemical properties, individual soil components and soil physical properties using ultraviolet (UV), Vis, NIR, mid-infrared (MIR) and combined spectra are also well documented. In a recent research report, simultaneous assessment on the prediction accuracy of various soil properties using UV, Vis, NIR and MIR regions (250– 25,000 nm range) was summarized by Viscarra Rossel et al. (2006). Their review includes not only the results of on-the-go (field-based) soil sensing but also laboratory analysis (lab-based). Most previous studies used dried, ground and sieved soil samples to measure soil reflectance spectra data in the laboratory, and developed calibration models using multivariate statistics. Techniques to collect soil reflectance spectral data in the laboratory using precision optical sensors and several multivariate statistics analysis methods were able to reduce both the cost and time (Chang et al., 2001; Chodak et al., 2003; D'Acqui et al., 2010; Daniel et al., 2003; Islam et al., 2004; Maleki et al., 2006; Matsunaga and Uwasawa, 1992; Mouazen et al., 2007; Shepherd and Walsh, 2002; Wetterlind et al., 2010). However, to draw highresolution soil maps, large amounts of soil samples must be carried to the laboratory. To achieve cost reduction, measurement must be field-based, which means measuring the soil reflectance spectra using fresh soil in agricultural fields. A few researchers have used fresh soil samples to measure soil reflectance spectra (Arakawa et al., 2010; Christy, 2008; Imade Anom et al., 2001; Kodaira et al., 2009; Mouazen et al., 2007). To further reduce the time and cost, soil reflectance spectra must be collected automatically. A recent study describes the development of an on-line (real-time) soil sensor that can collect large amounts of soil reflectance spectra using a Vis–NIR spectrophotometer mounted on a tractor, an agricultural applicator machine or an automobile (Christy, 2008; Mouazen et al., 2005; Shibusawa et al., 1999). Our soil sensing system is mounted on a tractor.

65

The real-time soil sensor (RTSS) used in this study is the successor model of the prototype (Shibusawa et al., 1999) designed and developed by our laboratory researchers and several partner companies. The prototype RTSS operated at a speed of 0.05 m s−1. Imade Anom et al. (2001) used the 01-model to draw soil maps of MC, SOM, N-n, pH and EC in a paddy field on the Experimental Farm at Kyoto University in Japan. The 02-model is a cost-reduction model and the measurement wavelength band was limited to nine bands. Umeda et al. (2011) used the product model (SAS1000) to investigate the availability of information between the total nitrogen soil map and wheat yield map at the Agricultural Experiment Station in Gunma Prefecture in Japan. The RTSS of the 01- and 02-model and the SAS1000 operated at a speed of 0.28 m s−1. In this study, we used the SAS1000 model RTSS. As a different feature, the traveling speed in the study is 0.56 m s −1, the fastest speed among the conventional models for the RTSS series. The prediction of several soil properties using conventional realtime or on-line soil sensors were made for MC, SOM, pH, EC, N-n, N-t, C-t and P-a. However, the measurements for soil properties are less than the number of soil properties that are measured using a spectrophotometer in the laboratory. The aims of this study are as follows: (1) Using spectra acquired with our RTSS and develop spectroscopic calibration models for moisture content (MC), soil organic matter (SOM), pH, electrical conductivity (EC), cation exchange capacity (CEC), total carbon (C-t), ammonium nitrogen (N-a), hot water extractable nitrogen (N-h), nitrate nitrogen (N-n), total nitrogen (N-t), available phosphorus (P-a), and phosphorus absorptive coefficient (PAC). (2) To compare our results to those of other studies using the coefficient of determination, root mean square error and residual prediction deviation. (3) To produce soil property maps using the results from the spectroscopic models and the RTSS. 2. Materials and methods 2.1. Real-time soil sensor The RTSS (SAS 1000, SHIBUYA MACHINERY Co., Ltd.) is mounted on a tractor using a three-point hitch (category II type). The RTSS equipment captures several datasets simultaneously as a real-time soil measurement system for precision agriculture. It is composed of the sensor unit's housing, a touch panel, the soil penetrator and the housing for the probes (Fig. 1). Electrical power for the RTSS was supplied from a gasoline inverter generator (EU-16i, Honda Motor Co., Ltd). The PTO (power

Fig. 1. RTSS-mounted tractor with cross-sectional view of the soil penetrator and probe housing.

66

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

take-off) of the tractor is connected to the hydraulic power of the RTSS and is used to drive the chisel unit into the soil. The sensor unit's housing includes the core devices of the RTSS system, such as a personal computer (Windows XP, Microsoft), 150-W Al-coated tungsten halogen lamp (USHIO Inc), micro CCD camera (TOSHIBA Co.), differential global positioning system (DGPS: DSM132, Trimble) receiver, and two spectrophotometers. The DGPS antenna is mounted on the roof. The spectrophotometer for Vis has a 256-pixel linear photodiode array (MMS 1, Carl Zeiss AG). The spectra ranged from 305 to 1100 nm with intervals of approximately 3 nm. A 128-pixel linear diode array of multiplexed InGaAs (MMS, Carl Zeiss AG) for NIR was used. The spectra ranged from 950 to 1700 nm with intervals of

approximately 6.0 nm pixel pitch. Also, the spectra data collected from the two spectrometers was converted to 5-nm-interval data by interpolation. The spectra ranged from 350 to 1700 nm with intervals of approximately 5 nm. During the experiment, the touch panel displayed uniform soil surface images from the micro CCD camera, and also recorded the images to a memory card. The displayed images were used to watch for emergencies and stability work, such as blockage by obstacles, and the images provided information for eliminating data from the data analysis, for example calibration and prediction outliers. The screen of the touch panel was on throughout the experiment, allowing us to see what was happening, as shown in Fig. 2. When the RTSS was

Fig. 2. Images of uniform soil surface captured by micro CCD camera. (a), (b) uniform soil surface suitable for measuring soil reflectance spectra, (c) a case of catching a straw of wheat, (d) a case of catching a stone, (e) a case of catching a larva, (f) soil stuck inside the sensor unit's housing.

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

crossing furrows, soil occasionally becomes stuck in the sensor's housing (Fig. 2f). In this case, we stopped the operation of the RTSS and removed the soil from inside the housing to start measurement again. The soil penetrator tip with a flat plane edge ensured uniform soil cuts, and the soil flattener following behind produced a uniform surface. In the housing for the probes, five micro optical devices were arranged. Two optical fiber probes were used to provide an illuminated area of about 50 mm in diameter on the underground soil surface. An additional optical fiber probe was used for collecting underground soil reflectance in the Vis–NIR range. The CCD camera lens was adjusted to monitor a 75-mm focus point on the uniform soil surface. A laser line marker was mounted to monitor changes in distance between the uniform soil surface and the micro optical devices. The RTSS was designed to collect underground soil reflectance spectra at depths of 0.05 to 0.35 m at a spacing of 0.05 m. In this experiment, we set the soil flattener depth to 0.15 m from tire of the RTSS. A great advantage in this experiment was setting the speed to 0.56 m s −1, which is double the conventional traveling speed of 0.28 m s −1. As a result, the RTSS measurement work reached a rate of 1 ha h −1. However, as a condition, preparation, measurement, travel time between 4 fields, cleaning and shipping were included (Kodaira et al., 2009).

2.2. Experimental site description The experimental site (total 10 fields, 31.48 ha) is a commercial upland field with an alluvial soil type located within the Tokachi Plain in Memuro-cho, Kasai District, Hokkaido, Japan (Fig. 3). The crop rotation system used at the site is five crops for 5 years: winter wheat–sugar beet–soy bean–potato–green manure. To develop the twelve soil properties calibration models, the experiment was conducted on field no. 3 (4.43 ha, 303.0 × 146.2 m; 42°50′55.32″ N and 143°00′13.68″ E) after harvesting winter wheat in October 2007, field no. 4 (4.51 ha, 303.0 × 148.8 m; 42°50′52.85″ N and 143°00′19.86″ E) after harvesting winter wheat in August with field no. 3 after harvesting sugar beet in October 2008, and field no. 3 after harvesting soy bean in October 2009. The soil texture of the two fields was described as follows: field no. 3 was 80.77% sand, 11.82% silt and 7.41% clay, and field no.4 was 82.05% sand, 10.82% silt and 7.13% clay. We also conducted experiments using the RTSS for the prediction of soil maps in other fields. However, this paper reports on the twelve soil maps for field no. 3 and no. 4. Also, we confirmed that there were mistakes in the N-n measurement using high-performance liquid chromatography (HPLC) by our laboratory in the sample dataset of field no. 3 in 2007. In this paper, the measured value by soil analysis for N-n and other soil properties are not included in the dataset for developing the twelve soil properties calibration models.

67

Fig. 4. Tramline used as a reference for RTSS travel and a data collecting point: The tramline matched the line (24 m spacing) of the crop protection sprayer (boomsprayer). Small black dots indicate the collection position for underground soil reflectance spectra data. Large black dots indicate the collection position for soil samples corresponding to the position of underground soil reflectance spectra data.

2.3. Collection of Vis–NIR spectra using the RTSS with soil samples For the purpose of site-specific soil management, the traveling line of the RTSS and collection spacing of soil samples were decided based on the grower's opinion, the grid center system and the specifications of the crop protection sprayer (boom sprayer) used. In this study case, it was decided that the traveling line of the RTSS would correspond to the tramline of the boom sprayer (spread width: 24 m). The grid cell for soil sampling was established as 25 × 24 m (0.06 ha) in accordance with the grower's opinion and the grid center system in the Hokkaido area (Hara, 1999). In this experiment, field nos. 3 and 4 were each divided into 72 grid cells. Therefore, the RTSS traveled at 24 m spacing on same tramline with the boom sprayer. Vis–NIR underground soil reflectance spectra data were acquired automatically every 1 data 4 s −1 (approximate 2.24 m interval) in the traveling direction. As shown in Fig. 4, the actual measured underground soil reflectance spectra positions by the RTSS are indicated by small and large black dots. The RTSS sounds an alarm at the time of each data acquisition, counts the number of data and displays it on the touch panel screen. We constantly confirmed the alarm sounds and the number of data displayed on the touch panel screen. Simultaneously, to analyze the amount of chemicals in the soil as measured by the RTSS, soil samples were taken from the position of every 11 datasets (approximate center of a grid cell) and a wooden stick was inserted into the soil (Fig. 4). However, the soil sampling positions shifted slightly from our plan because the tractor speed changed under muddy or undulating soil surface conditions. As shown in Fig. 4, the actual positions of the collected soil samples are indicated by large black dots. As shown in Fig. 5, the

Fig. 3. Location of the experimental site including a field layout. (a) Location of Tokachi Plain in Japan. (b) Location of the experimental site in Tokachi Plain. (c) Layout of the experimental fields.

68

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

Fig. 5. Soil sampling method: (a) 72 wooden sticks were inserted in the soil at points predetermined by scanning data number, (b) uniform soil surface smoothed over by the RTSS was dug up, (c) soil samples were collected from in and around the area scanned by the RTSS, (d) soil samples were divided into two sealable plastic bags, one set to be sent to TUAT, the other set to be sent to APCRL.

procedure for soil sampling was as follows: a wooden stick was used to mark each soil sampling position, which corresponded to the position of underground soil reflectance spectra collected by the RTSS (Fig. 5a). After the RTSS traveled through the area, we dug up the soil surface smoothed over by the RTSS (Fig. 5b), and packed the soil samples in

sealable plastic bags (Fig. 5c), taking two sets of soil samples from each point (Fig. 5d). Two sets of 144 fresh soil samples were collected to measure the amount of soil chemicals at the depth of 0.15 m corresponding to the position that the underground soil reflectance spectra was collected by the RTSS.

Table 1 Statistical results of soil chemical analyses on soil properties in calibration dataset. Property

Na

Statistics Min

MC; wt. % SOM; wt. % pH; non EC; mS cm−1 CEC; me 100g−1 C-t; % N-a; mg 100g−1 N-h; mg 100g−1 N-n; mg 100g−1 N-t; % P-a; mg 100g−1 PAC; non a b c

Max

Mean

Range

Skewness Kurtosis Variance

S.D.

C.V.

b

R / Mc

R/ S.D.

Analyzing method

The soil analyzer

Operator

0.14

−0.68

28.08

5.30 24.23 1.06

4.37

Oven dry

DK-610, YAMATO

TUAT

6.59

6.34 −0.20

0.39

1.30

1.14 17.28 0.96

5.56

Muffle furnace

MF-28, YAMATO

TUAT

7.17 0.27

5.72 0.07

2.37 0.24

0.64 1.78

0.30 6.29

0.21 0.001

0.46 8.03 0.41 0.03 49.76 3.50

5.15 7.04

Glass electrode AC bipolar

D-24, HORIBA D-24, HORIBA

TUAT TUAT

5.90

22.62

14.63

16.72 −0.23

−1.01

19.01

4.36 29.79 1.14

3.84

Absorptiometry

144 144

0.81 0.15

3.13 1.54

1.88 0.63

2.32 −0.14 1.39 0.31

−0.31 −0.33

0.22 0.09

0.47 24.84 1.23 0.30 46.83 2.20

4.97 4.70

144

3.40

8.97

5.24

5.57

0.71

0.24

1.07

1.04 19.77 1.06

5.37

144

0.21

4.18

0.70

3.97

4.19

20.76

0.28

0.53 75.82 5.65

7.45

144 144

0.07 25.20

0.24 114.73

0.14 54.23

0.17 −0.18 89.53 0.90

−0.16 0.97

0.001 299.1

0.03 22.67 1.19 17.29 31.89 1.65

5.25 5.18

148.35 23.46 1.20

5.11

144

11.32

34.46

21.87

144

3.88

10.22

144 144

4.81 0.03

144

144 311.0

Number of soil samples. Range. Mean.

1069.0

632.3

23.14

758.0

0.13

−0.26

22,008

QUAATRO, BRAN+LUEBBE Tyurin's Method NC-220F, SUMIGRAPH Absorptiometry QUAATRO, BRAN+LUEBBE Absorptiometry QUAATRO, BRAN+LUEBBE Absorptiometry QUAATRO, BRAN+LUEBBE Kjeldahl method NC-220F, SUMIGRAPH Absorptiometry QUAATRO, BRAN+LUEBBE Absorptiometry QUAATRO, BRAN+LUEBBE

APCRL APCRL APCRL APCRL APCRL APCRL APCRL APCRL

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

69

Table 2 A review of the literature comparing quantitative predictions for the soil properties used in this study. Property

Range (nm)

Multivariate methoda

Ncal|Nvalb

RMSE

R2

RPD

Categoryc

Authors

MC; kg kg−1 MC; % MC MC MC; % MC; % SOM;% SOM SOM; % SOM; % pH (H2O) pH (1:2.5 KCl) pH (1:1 H2O) pH (1:2.5 H2O) EC; μS cm−1 EC; mS cm−1 CEC; c mol(+) kg−1 CEC; c mol(+) kg−1 CEC; c molc kg−1 C–t; g kg−1 C–t C–t; % C–t; % C–t; % N–a (NH4-N) N–h; mg 100g−1 N–n; mg 100g−1 N–n N–t; % N–t; % N–t; % P–a; mg kg−1 P–a; mg 100g−1 P–a; mg 100g−1 PAC

400–2498 400–2400 306–1710 920–1718 350–1700 500–1550 400–2400 920–1718 350–1700 430–2500 400–2400 306–1710 920–1718 200–2000 400–2400 250–2500 400–2498 350–2500 200–2000 400–2498 306–1710 350–1700 500–1550 430–2500 250–2500 400–2500 400–2400 250–2500 350–1700 500–1550 430–2500 400–1100 401–1663 430–2500 1100–2500

PCR (8) SMLR (606,1329,1499) PLS PCR PLS (6) GAWLS SMLR (606,1311,1238) PCR PLS (4) PLSR (7) SMLR (972,1238) PLS PCR PLSR (9) SMLR (456,984,1014) PLSR PCR (8) MARS PLSR (7) PCR (7) PLS PLS (8) GAWLS PLSR (4) PLSR MPLS SMLR (589,1014) PLSR PLS (7) GAWLS PLSR (5) ANN PLSR PLSR (7) MLR

30|119 15|10 348 x-vald 105 x-val 72 x-val 1056 x-val 15|10 106 x-val 72 x-val 122|31 15|10 295 x-val 106 x-val 198 x-val 15|10 277|260 30|119 493|247 197 x-val 30|119 173 x-val 72 x-val 1056 x-val 122|31 – 50|50 15|10 – 72 x-val 1056 x-val 122|31 41 126 x-val 122|31 60|60

0.005 3.111 0.024 2.800 1.689e 7.397 0.559 0.400 0.721e 0.420 0.160 0.215 0.440 0.300 41.72 0.100 3.820 3.800 3.900 7.860 0.268 0.117e 0.724 0.160 – 0.02e 4.741 – 0.010e 0.057 0.024 – 1.202 3.300 –

0.84 0.66 0.89 0.65 0.77e 0.85 0.65 0.80 0.49e 0.94 0.58 0.71 0.62 0.72 0.65 0.71 0.81 0.88 0.87 0.87 0.73 0.95e 0.83 0.89 b0.1 0.94e 0.54 0.42 0.86e 0.84 0.85 0.81 0.73 0.48 0.97e

2.5 – 3.0 – – – – – – 3.7 – 2.1 – 1.9 – 1.8 2.3 – 2.7 2.8 1.9 – – 3.4 b1.0 3.0 – 1.3 – – 2.7 – 1.9 1.7 –

A – A – – – – – – A – A – B – B A – A A B – – A C A – C – – A – B B –

Chang et al. (2001) Imade Anom et al. (2001) Mouazen et al. (2007) Christy (2008) Kodaira et al. (2009) Arakawa et al. (2010) Imade Anom et al. (2001) Christy (2008) Kodaira et al. (2009) Wetterlind et al. (2010) Imade Anom et al. (2001) Mouazen et al. (2007) Christy (2008) D'Acqui et al. (2010) Imade Anom et al. (2001) Islam et al. (2004) Chang et al. (2001) Shepherd and Walsh (2002) D'Acqui et al. (2010) Chang et al. (2001) Mouazen et al. (2007) Kodaira et al. (2009) Arakawa et al. (2010) Wetterlind et al. (2010) Islam et al. (2004) Chodak et al. (2003) Imade Anom et al. (2001) Islam et al. (2004) Kodaira et al. (2009) Arakawa et al. (2010) Wetterlind et al. (2010) Daniel et al. (2003) Maleki et al. (2006) Wetterlind et al. (2010) Matsunaga and Uwasawa (1992)

a Multivariate techniques include stepwise multiple linear regression (SMLR), multiple linear regression (MLR), multivariate adaptive regression splines (MARS), genetic algorithm-based wavelength selection (GAWLS), artificial neural network (ANN), principle components regression (PCR), and partial least-squares regression (PLSR). Shown in brackets are the spectral bands used or the number of bands or number of PCR components or number of PLSR factors used in the predictions. b Ncal |Nval shows the number of samples used in the spectral calibration and the number of factors used in the validation. c Category means the ability of each technique in terms of parameter validation and prediction. A: RPD > 2.0; B: RPD = 1.4–2.0; C: RPD b 1.4 (Chang et al., 2001). d x-val suggests that validation was conducted independently using statistical cross-validation. e Coefficient correlation.

We calibrated the spectrometer with a standard white reference panel (Spectralon, Labsphere Inc). The white reference was measured several times before starting the experiment and during work breaks. 2.4. Soil chemical analysis To measure the amount of chemical components in the soil, two sets of 144 soil samples were collected from both field no. 4 and no. 3 in 2008, and two sets of 72 soil samples were collected from field no.3 in 2009. One hundred and forty-four soil samples were collected in 2008 to develop the spectroscopic calibration models. 72 soil samples of 2009 were collected to confirm accuracy of the predictions and the fine resolution soil maps that we derived from these data. One set was analyzed at our laboratory (TUAT) and the other set was analyzed at the Agricultural Product Chemical Research Laboratory (APCRL: Federation of Tokachi Agricultural Cooperative Association, Hokkaido, Japan). The set of samples sent to TUAT were transported in a refrigerator car at a temperature below 10 °C, and then stored at TUAT in a refrigerator at 5 °C. Soil chemical analyses for MC, SOM, pH and EC were carried out by TUAT using the standard procedures in the Hokkaido area in Japan (Souma and Kikuchi, 1992). The 144 fresh soil samples were crushed and sieved through a 2-mm sieve. The samples were

then stored in sealable plastic bags at 5 °C until the end of chemical analysis. MC, pH and EC were measured using fresh soil. MC was measured using the oven-dry method at 110 °C for 24 h. Soil pH was measured by glass electrode (D-24, HORIBA) using a soil:distilled-water weight ratio of 1:2.5. Soil EC was measured by AC bipolar method (D-24, HORIBA) using a soil:distilled-water weight ratio of 1:5. After shaking for 2 h and equilibration, pH and EC were measured in the supernatant. The soil:distilled-water weight ratio for pH and EC were calculated using each MC result of soil samples. SOM was measured using soil samples sieved through a 1-mm sieve and dried in a muffle furnace at 750 °C for 3 h. Each soil analysis was conducted three times, and the average value was adopted as a dataset for the multivariate statistical analysis. The other set of 216 soil samples was taken to APCRL before the end of the day in which they were collected. Soil chemical analyses for CEC, C-t, N-a, N-h, N-n, N-t, P-a, and PAC were carried out by APCRL using the standard procedures in the Hokkaido area in Japan (Souma and Kikuchi, 1992). The 216 fresh soil samples were dried, crushed and sieved. The C-t was measured by Tyurin's method and N-t was measured by Kjeldahl method. Other soil properties, CEC, N-a, N-h, N-n, P-a, and PAC, were measured by absorptiometry. Statistics of soil chemical values for developing calibration models, soil analysis methods and instruments are given in Table 1.

70

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

All collected underground soil reflectance spectra using the RTSS in the commercial upland field were converted to absorbance with white reference spectra using the standard reflector (Spectralon, Labsphere Inc) and dark reference spectra due to light shielding, and using the Beer-Lambert's law (Williams and Norris, 2001). The calculation formula is as follows:

Absorbance spectra ¼ log10 ðRwhite –Rdark Þ–log10 ðR–Rdark Þ

Rwhite Rdark R

ð1Þ

white reflectance spectra using the standard reflector dark reflectance spectra due to light shielding reflectance spectra of underground soil surface

Absorbance spectra were converted to 5-nm-interval data (original absorbance) by interpolation method using Data Monitor Software (Shibuya Seiki Co., Ltd.). The spectra of original absorbance ranged from 350 to 1700 nm. The next step was pretreatment for original absorbance. Since Vis– NIR spectra are affected by particle size and light scatter, pretreatment of the spectral data can often improve calibration accuracy. Commonly used correction techniques include: smoothing, first derivative and second derivative (Savitzky and Golay, 1964), maximum normalization, standard normal variate and detrending (SNV-D: Barnes et al., 1989) and multiplicative scatter correction (MSC: Martens and Naes, 1989). The selection criteria of any pretreatment technique are the largest coefficient of determination (R2val) and the smallest RMSEval. The final step in the development of twelve soil properties calibration models was PLSR coupled with full (leave-one-out) crossvalidation. The number of latent variables for each twelve soil properties calibration models was determined by examining a plot of residual validation variance against the number of latent variables obtained from PLSR. The latent variable of the first minimum value of residual variance is selected for the Minimum RMSEval and Maximum R 2val on the validation results. Sample outliers are detected by using the residual sample variance plot after the PLSR. Individual samples located far from the zero line of residual variance on the validation views are considered to be outliers and are excluded from the analysis. The final calibration model for each twelve soil properties was re-calculated using PLSR coupled with full cross-validation. Also, if image data similar to Fig. 2c, d, e and f was discovered in the calibration and prediction dataset, it was probably removed from the dataset. In this study, to enhance weak signals and reduce noise and light scatter that would affect the baseline, smoothing and second (2nd) derivative mathematical pretreatment was employed depending on

Fig. 6. Absorbance data for developing the calibration model. (a) Original absorbance, (b) pretreated absorbance using smoothing and 2nd derivative, (c) pretreated absorbance using 2nd derivative.

2.5. Spectral pretreatment and establishment of calibration models The first step in the development of the spectroscopic calibration models is the transformation from soil reflectance to soil absorbance. Table 3 Summary of PLSR results for the twelve soil properties dataset. Property

MC SOM pH EC CEC C–t N–a N–h N–n N–t P–a PAC a b c d

Unit

wt. % wt. % pH unit mS cm−1 me 100g−1 % mg 100g−1 mg 100g−1 mg 100g−1 % mg 100g−1 Non

Range (nm)

500–1600 500–1600 500–1600 500–1600 500–1600 500–1600 500–1600 500–1600 1100–1650 500–1600 500–1600 500–1600

Pre-treatment

2nd Derivative 2nd Derivative Smth and 2nd-Dd 2nd Derivative 2nd Derivative 2nd Derivative 2nd Derivative 2nd Derivative 2nd Derivative 2nd Derivative 2nd Derivative 2nd Derivative

PCa

6 6 6 4 6 5 8 9 5 5 4 6

Nb

130 130 130 130 130 130 130 130 130 130 130 130

Calibration

Categoryc

Prediction

Rcal

2 Rcal

RMSEcal

Rval

0.97 0.96 0.88 0.80 0.96 0.95 0.89 0.89 0.71 0.94 0.87 0.96

0.95 0.92 0.78 0.64 0.92 0.91 0.79 0.77 0.50 0.89 0.76 0.92

1.23 0.30 0.19 0.02 1.25 0.13 0.12 0.44 0.14 0.01 7.46 42.02

0.96 0.95 0.83 0.77 0.94 0.94 0.83 0.81 0.67 0.93 0.85 0.95

S.D.

2 Rval

RMSEval

RPD

This paper

Previous

5.11 1.02 0.37 0.02 4.08 0.42 0.23 0.84 0.15 0.03 14.35 142.75

0.93 0.90 0.69 0.60 0.89 0.89 0.69 0.65 0.45 0.87 0.72 0.90

1.42 0.35 0.23 0.02 1.44 0.15 0.15 0.56 0.15 0.01 8.00 47.94

3.6 2.9 1.6 1.3 2.8 2.9 1.6 1.5 1.0 2.6 1.8 3.0

A A B C A A B B C A B A

A A A B A A C A C A B –

Number of PLSR factors used in the model. Number of samples used in the model. Category of prediction (full cross-validation) ability of PLSR for parameters. A: RPD > 2.0; B: RPD = 1.4–2.0; C: RPD b 1.4 (Chang et al., 2001). Smoothing and 2nd derivative.

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

the parameter conditions. The first derivative is able to eliminate baseline fluctuation. The 2nd derivative preserves the effect of the first derivative. Furthermore, the 2nd derivative preserves some effects as follows: a) Be able to eliminate the background noise by diffuse reflection. b) The peak values of the wavelength are enhanced more than the first derivative. c) The peak value of the wavelength corresponds with the wavelength of original absorbance. Also, the best pretreatment and Vis–NIR wavelength ranges were chosen following a trial and error procedure, using the following criteria: the smallest root mean square error of full cross-validation (RMSEval), the highest coefficient of determination (R2val) close to one.

71

Therefore, the best calibration model for each twelve soil property was re-calculated several times as follows: a) Sample outliers are detected by using the residual sample variance plot after the PLSR. b) Individual sample outliers located far from the zero line of residual variance are considered to be outliers and are excluded from the analysis. In this study, sample outliers were selected for four or five samples at a time. c) A total of 14 sample outliers were selected. If there were several similar outliers, the best calibration model was selected for the Minimum RMSEval and Maximum R 2val in the validation results. d) If the best result on R 2val could not be obtained, other wavelength range (e.g., N-n) for calibration was selected appropriately between 350 and 1700 nm.

Fig. 7. Scatter plot of measured values versus Vis–NIR predicted values using partial least squares regression (PLSR) coupled with full cross-validation datasets for: (a) moisture content, (b) soil organic matter, (c) pH, (d) electrical conductivity, (e) cation exchange capacity, (f) total carbon, (g) ammonium nitrogen, (h) hot water extractable nitrogen, (i) nitrate nitrogen, (j) total nitrogen, (k) available phosphorus, and (l) phosphorus absorptive coefficient.

72

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

Fig. 7 (continued).

Mathematical pretreatment and development of calibration models using PLSR coupled with full cross-validation were performed using the Unscrambler Ver.9.8 software (CAMO Software AS, Oslo, Norway).

reference data set to root mean square error of full cross-validation (RMSEval), as follows:

2.6. Comparison of analysis accuracy

RPD ¼ S:D: RMSEval :

To compare our results with other studies, we re-extracted the results on the Vis–NIR wavelength regions from recent studies. There are summarized as in Table 2. However, there is a difference in several of the conditions for measuring soil reflectance spectra, e.g. dried soil and fresh soil are included as well as field-based and lab-based, and development of calibration model using both pretreatment and multivariate statistics analysis. The properties for comparison of the analysis accuracy were R 2val and residual prediction deviation (RPD) of prediction (same as validation) results. R 2val was obtained by the Unscrambler Ver.9.8 software. RPD is given by the ratio of standard deviation (S.D.) of the

−1

ð2Þ

In this study, RPD is classified according to category, which means the properties of the full cross-validation ability of PLSR in this study. Values of RPD > 2 were considered excellent (symbol A), values b 1.4 (symbol C) were considered unreliable, and values between 1.4 and 2 were almost good (symbol B) (Chang et al., 2001). 2.7. Soil map preparation The predicted soil maps and the measured soil maps for twelve soil properties were prepared using ArcGIS Ver10.0 software (ESRI

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

Inc., USA). The soil maps were interpolated using the inverse distance weighing (IDW) method. The IDW settings were as follows: a) The Power function (the exponent of distance): 2 (default value) b) Cell size: 0.617 (automatically calculated value) c) Search radius type: variable. Shades of color classification were divided into seven segments. 3. Results 3.1. Development of calibration model A total of 144 soil sample statistics for the calibration and full crossvalidation (prediction) sets are listed in Table 1. The statistical results include the minimum, maximum, mean, range and standard deviation for each soil parameter. The soil analysis method, soil analyzer and operator are also listed in the same table.

73

The 144 soil reflectance spectra using the RTSS were converted to absorbance spectra using calculation formula (1). The original absorbance spectra are shown in Fig. 6a. The best pretreatment and Vis–NIR wavelength ranges were chosen as follows. For the pretreatment of pH, the 2nd derivative was applied after the smoothing. The parameter setting of pretreatment for smoothing and the 2nd derivative with Savitzky–Golay for pH was selected as follows: number of smoothing points: 11, and polynomial order: 2. The pretreated absorbance spectra for pH are shown in Fig. 6b. For the pretreatment of other soil properties, only the 2nd derivative was applied. The parameter setting of pretreatment for 2nd derivative with Savitzky–Golay was selected as follows: number of smoothing points: 21, and polynomial order: 2. The pretreated absorbance spectra for the other calibration models are shown in Fig. 6c. The best Vis–NIR wavelength of N-n for modeling was in the range from 1100 to 1650 nm; other soil parameters were in the range from 500 to 1600 nm as shown in Table 3.

Fig. 8. Regression coefficients for: (a) MC, (b) SOM, (c) pH, (d) EC, (e) CEC, (f) C-t, (g) N-a, (h) N-h, (i) N-n, (j) N-t, (k) P-a and (l) PAC.

74

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

Fig. 8 (continued).

PLSR coupled with full cross-validation was used to establish the relationship between pretreated Vis–NIR absorbance data and soil chemical results for each of the 144 datasets. The PLSR results of the calibration and validation sets were obtained as shown in Table 3. A total of 14 sample outliers were detected as shown in Table 3. The excellent calibration model for twelve soil properties carried out MC, SOM, CEC, C-t, N-t and PAC due to the value of RPD was 3.6, 3.0, 2.9, 2.9, 2.6 and 3.0, respectively (Table 3). pH, N-a, N-h and P-a showed good levels of accuracy with RPD values between 1.5 and 1.6 (Table 3). EC and N-n showed a poor level of accuracy with RPD valuesb 1.3 (Table 3). Poor accuracy of RPD with R2val for EC and N-n compared with other statistical values which had gathered in a narrow range on the histogram, e.g., low variance and high skewness with kurtosis (Table 1). Scatter plot of the prediction using a total of 130 datasets was obtained as shown in Fig. 7. Also shown in the figure are several primary regression equations.

The regression coefficients for twelve soil properties due to the number of PLSR factors used in the modeling were obtained as shown in Fig. 8. The size of the regression coefficients (negative or positive) represents the importance of the absorption band in terms of the explanation of variance in soil analysis data. In our case, negative peaks of the regression coefficients are showing positive correlation for soil analysis data. 3.2. Comparison with previous studies As shown in Table 3, the category of RPD for MC, SOM, CEC, C-t, N-a, N-n, N-t and P-a on the prediction models was the same or better than that in previous studies. RPD for pH, EC and N-h had only slightly poorer results compared with previous studies. RPD for PAC could not be found in previous studies. The accuracy of R 2val for Vis–NIR modeling compared with previous studies is as follows: MC, CEC, C-t, N-a and N-t

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79 Table 4 Summary of measurement instruments under measurement conditions from Table 1. Reflectance spectra measurement for calibration modeling

Parameters

Authors

Chang et al. (2001) Chodak et al. (2003) Matsunaga and Uwasawa (1992) Christy (2008)

Instrument

Based

6500, Foss NIR Systems Foss NIR Systems 6250, NEOTEC

Lab Lab Lab

MC, CEC, C–t N–h PAC

NIR–128L–1.7–USB, Control development Inc. FieldSpec FR, ASD Inc. FieldSpec Pro FR, ASD Inc.

Lab

MC, SOM, pH

Lab Lab

CEC SOM, C–t, N–t, P–a pH, CEC

Shepherd and Walsh (2002) Wetterlind et al. (2010)

EC, N–a, N–n P–a MC, pH, C–t MC, SOM, pH, EC, N–n MC, SOM, C–t, N–t MC, C–t, N–t P–a

Islam et al. (2004) Maleki et al. (2006) Mouazen et al. (2007) Imade Anom et al. (2001)

Spectrum GX, PerkinElmer Inc. Cray 500, Varian Inc. Zeiss Corona 45, Zeiss Company MMS 1 and MMS, Carl Zeiss AG

Lab Lab Lab Lab Field Field

Spectroradiometer, StellarNet Inc.

Field Lab

D'Acqui et al. (2010)

Kodaira et al. (2009) Arakawa et al. (2010) Daniel et al. (2003)

75

and N-n at 13.9% and 22.7%, respectively. The prediction error for each parameter on field no. 4 ranged from 0.1 to 1.7%. As shown in Fig. 10, the measured soil maps and the predicted highresolution soil maps using the RTSS for twelve soil properties were drawn again at field no. 3 in November 2009. The high-resolution twelve soil properties soil maps were drawn without adjusting the twelve soil properties calibration models in order to confirm the performance of regression coefficients on the calibration models. Table 5 shows the comparison between mean values measured and predicted for each parameter. In 2009 prediction errors below 10% were given by parameters: MC, SOM, pH, N-h, N-n, and N-t. More than 20% errors were observed for EC and N-a. Prediction error of calibration is commonly rooted in its calculation procedure, properties of reference data collected, and noise of spectral data record. More than 10% prediction errors were obtained for EC and N-n values of field no. 3 in 2008 and EC, CEC, C-t, N-a, P-a, and PAC values of field no. 3 in 2009. Note that the values for coefficient of variation (C.V.), range/mean and range/S.D. corresponding to those parameters in Table 6 are EC (56.4, 3.58, 6.35), N-a (61.8, 2.96, 4.80) and N-n (84.82, 4.72, 5.56), respectively. It indicated that critical values for C.V., range/mean and rang/S.D. appeared when the prediction error is controlled, and the values were 56.4 (C.V.), 2.96 (range/mean) and 4.80 (range/S.D.) in this case. 3.4. A case of decision-making

were the same or better; SOM, pH, EC, N-h, N-t and P-a were only slightly poorer; R2val for PAC could not be found in previous studies. As shown in Table 4, most previous studies developed their Vis–NIR calibration model using lab-based measurement. Field-based measurement appeared in only three studies (Arakawa et al., 2010; Imade Anom et al., 2001; Kodaira et al., 2009). Also, Imade Anom et al. (2001), Maleki et al. (2006), Mouazen et al. (2007), Christy (2008), Kodaira et al. (2009) and Arakawa et al. (2010) used an on-line measurement system working on agricultural fields. Compared to their results, the accuracy of RPD for MC, SOM, C-t and N-t was better in our study. pH, EC, N-n and P-a were only slightly poorer. The main improvement to the RTSS was the change in the speed of the traveling tramline to double (0.56 m s -1) that of the conventional RTSS speed. Compared to the results obtained by Imade Anom et al. (2001), the accuracy of R 2val for MC, SOM and pH was better in our study; EC and N-n were only slightly poorer. Therefore, the analysis accuracy for the calibration model has been improved by a factor of approximately 2. The work speed of the RTSS was obtained as 0.56 m s-1. It became possible to connect the RTSS with several tillage, ridging and fertilization systems (Kusakawa et al., 2003; Sugimoto et al., 2011; Yashiro, 2008) for site-specific soil management in Japan. Also, it was same speed with a soil sensor-based variable-rate fertilization system (Maleki et al., 2008). However, typical speeds for other many field operations (e.g., tillage, seeding, manure and fertilizer application, harvesting) is 2 to 5 times greater than the speed used in this study. There is a need for many improvements to be connected with many precision agriculture applications.

3.3. Twelve soil properties mapping High-resolution twelve soil property soil maps for field no. 3 and no. 4 measured in 2008 are shown in Fig. 9. The measured soil maps of twelve soil properties were drawn using 72 soil samples from each field. The predicted high-resolution twelve soil property soil maps using the RTSS were drawn using 772 spectral data sets from field no. 3, and 802 spectral data sets from field no.4. Table 5 shows the results of comparing the measured and predicted mean values of each parameter. The prediction error for each parameter on field no. 3 ranged from 0.1 to 22.7%. The prediction error was greater than 10% for EC

In February 2008, the predicted high-resolution twelve soil property soil maps along with the measured soil maps were passed to the grower for site-specific soil management on precision farming. The pH map for field no. 3 reminded the grower that several years earlier a large amount of beet lime (CaCO3) was accidentally spilled. The site of the spill was on the south side of the area between field no. 3 and no. 4. The soil pH value was 7.85 in 2007 (Fig. 11a) and 7.17 in 2008 (Fig. 11b). The grower planned to reduce the high pH value in and around the grid cell of high alkaline soil at the end of fiscal 2008. A target pH value of 6.0 was decided. In May 2009, the grower distributed powdered sulfur in the high pH value grid cell using a broadcaster machine (Fig. 11c). The amount of powdered sulfur distributed was approximately 360 kg 10 a −1, calculated based on the manufacturer's instructions. In November 2009, we confirmed a pH value of 5.67 by soil analysis (Fig. 11d). This was the first case in Japan of site-specific soil management for precision farming using high-resolution soil maps with the RTSS. Also, the high-resolution soil maps using the RTSS were accepted by the grower as a decision-making support tool. 4. Discussion and conclusions Spectroscopic calibration models based on Vis–NIR underground soil reflectance spectra collected by the RTSS in commercialized agricultural upland fields were developed using PLSR with full cross-validation. The investigated soil parameters in this study were MC, SOM, pH, EC, CEC, C-t N-a, N-h, N-n, N-t, P-a, and PAC in the alluvial soil. Comparisons of the predicted accuracy using calibration models on a total of 144 samples were made in terms of each value of the PLSR results with previous studies. Accuracy of RPD for MC, SOM, CEC, C-t, N-a, N-n, N-t, P-a, and PAC was almost the same or better than that obtained in previous studies. RPD for pH, EC and N-h showed only slightly poorer results. One reason for this is the low variance value of EC and pH that was concentrated in a narrow range. However, we understand that some problems still must be solved. For example: a) To evaluate accuracy of twelve calibration models on the variations in clay, sand and silt content.

76

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

Fig. 9. High-resolution DSP soil maps for field no. 3 and no. 4 in 2008. The measured maps were drawn using the data from 72 soil samples. The predicted maps were drawn using spectral data as follows: field no. 3 was drawn using 772 datasets; field no. 4 was drawn using 802 datasets.

b) To develop more reliable spectroscopic models using the RTSS, soil samples with Vis–NIR data from over larger areas and using fresh soils are needed. Researchers and growers hope to be able to measure fresh soil directly in the field and obtain several high-resolution soil maps for crop

production of site-specific crop management on precision agriculture. However, most previous studies measured the soil reflectance spectra using dried soil in the lab. In that respect, the results of this study are promising because high-resolution mapping using spectroscopic estimates can be used as a source of information for the grower to make agronomic

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

77

Table 5 Comparison of measured versus predicted mean values of soil properties for each field. Property

MC SOM pH EC CEC C-t N-a N-h N-n N-t P-a PAC

No.4, 2008

No.3, 2008

No.3, 2009

Measured (72 samples)

Predicted (802 data)

Error (%)

Measured (72 samples)

Predicted (772 data)

Error (%)

Measured (72 samples)

Predicted (698 data)

Error (%)

24.73 6.31 5.44 0.07 11.24 1.69 0.47 5.06 0.56 0.13 46.81 592.17

24.89 6.33 5.43 0.07 11.30 1.71 0.47 5.03 0.55 0.13 46.75 597.21

0.62 0.31 0.15 0.21 0.57 1.32 0.14 0.52 1.71 0.75 0.13 0.85

19.00 6.88 6.01 0.07 18.01 2.07 0.80 5.43 0.84 0.16 61.66 672.39

19.15 6.85 6.00 0.06 18.12 2.04 0.74 5.31 0.65 0.15 59.54 682.97

0.80 0.49 0.11 13.86 0.58 1.27 7.43 2.27 22.70 1.29 3.44 1.57

29.37 6.62 6.19 0.04 16.53 1.88 0.53 5.17 0.83 0.15 64.36 697.28

28.90 7.28 6.21 0.08 19.57 2.26 1.03 4.68 0.87 0.16 52.32 831.74

1.60 9.91 0.40 105.93 18.34 19.95 94.76 9.38 4.31 7.72 18.70 19.28

decisions. Also, the reflectance spectra sample point density collected by the RTSS is adequate for use as decision support for variable-rate applications in precision agriculture. These data might also provide useful information for variable rate applications on precision agriculture in the near future.

Acknowledgments This study was part of the research findings obtained from the MAFF sponsored research project known as “Ninaite Pro” (from fiscal 2007 to fiscal 2011; our research project was completed in fiscal

Fig. 10. High-resolution DSP soil maps for field no. 3 in November 2009. The measured maps were drawn using data from 72 soil samples. The predicted maps were drawn using 698 spectral datasets due to non-improvement of calibration models.

78

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79

Table 6 Statistical results of soil chemical analyses on soil properties in each field. Property

Statistics No.4, 2008 N

MC SOM pH EC CEC C-t N-a N-h N-n N-t P-a PAC a b c

a

72 72 72 72 72 72 72 72 72 72 72 72

No.3, 2008 b

Mean

Range

S.D.

C.V.

R /M

24.73 6.31 5.44 0.07 11.24 1.69 0.47 5.06 0.56 0.13 46.80 592.2

19.41 5.43 1.45 0.12 11.20 1.91 1.39 4.22 0.73 0.13 61.10 662.0

4.85 1.25 0.27 0.03 3.08 0.44 0.29 0.92 0.15 0.03 13.73 158.0

19.61 19.76 4.96 43.03 27.42 25.96 61.81 18.20 27.03 24.88 29.33 26.68

0.78 0.86 0.27 1.77 1.00 1.13 2.96 0.83 1.29 0.98 1.31 1.12

c

No.3, 2009

R/S.D.

N

Mean

Range

S.D.

C.V.

R/M

R/S.D.

N

Mean

Range

S.D.

C.V.

R/M

R/S.D.

4.00 4.35 5.39 4.11 3.63 4.36 4.80 4.58 4.79 3.95 4.45 4.19

72 72 72 72 72 72 72 72 72 72 72 72

19.00 6.88 6.01 0.07 18.01 2.07 0.80 5.43 0.84 0.16 61.66 672.4

16.55 6.14 2.32 0.24 10.63 2.11 0.82 5.47 3.97 0.16 81.30 758.0

4.04 0.95 0.43 0.04 2.35 0.42 0.20 1.12 0.71 0.03 17.38 126.9

21.29 13.75 7.20 56.40 13.02 20.22 24.66 20.56 84.82 17.99 28.19 18.87

0.87 0.89 0.39 3.58 0.59 1.02 1.03 1.01 4.72 1.00 1.32 1.13

4.09 6.49 5.35 6.35 4.53 5.05 4.16 4.90 5.56 5.54 4.68 5.97

72 72 72 72 72 72 72 72 72 72 72 72

29.37 6.62 6.19 0.04 16.53 1.88 0.53 5.17 0.83 0.15 64.36 697.3

15.62 3.65 1.20 0.17 9.58 1.70 1.53 4.77 1.67 0.10 74.07 664.0

3.60 0.82 0.29 0.02 2.35 0.38 0.28 0.95 0.32 0.03 15.79 137.9

12.26 12.38 4.64 52.34 14.19 20.33 53.16 18.38 37.87 17.05 24.53 19.77

0.53 0.55 0.19 4.45 0.58 0.90 2.88 0.92 2.00 0.67 1.15 0.95

4.34 4.45 4.18 8.50 4.08 4.45 5.42 5.02 5.28 3.95 4.69 4.82

Number of soil samples. Range. Mean.

Fig. 11. Transition of pH value in field no. 3 due to the grower's decision-making on site-specific soil management. (a) In and around the accident site, the measured pH was 7.85. (b) One year later, the measured pH was 7.17. (c) Powder sulfur was distributed at approximately 360 kg 10 a−1 using a broadcaster. The target pH value was set to 6.0. (d) It was confirmed that the pH value decreased to 5.67.

2009). We would like to thank the Agriculture, Forestry and Fisheries Research Council Secretariat for financial support and the Memuro Upland Farming Research Station of the National Agricultural Research Center for Hokkaido Region for research support. In addition, we wish to thank the Tokoyama farm for allowing us to use their fields and S∙I Seiko Co., Ltd. (current Shibuya Seiki Co., Ltd.) for their cooperation with operating the tractor, and the university students of Tokyo University of Agriculture and Technology for their important help with the extensive soil sampling.

References Adamchuk, V.I., Hummel, J.W., Morgan, M.T., Upadhyaya, S.K., 2004. On-the-go sensors for precision agriculture. Computers and Electronics in Agriculture 44, 71–91. Arakawa, M., Yamashita, Y., Funatsu, K., 2010. Genetic algorithm-based wavelength selection method for spectral calibration. Journal of Chemometrics 25, 10–19. Barnes, R.J., Dhanoa, M.S., Lister, S.J., 1989. Standard normal variate transformation and detrending of near infrared diffuse reflectance. Applied Spectroscopy 43, 772–777. Chang, C.-W., Laird, D.A., Mausbach, M.J., Hurburgh Jr., C.R., 2001. Near-infrared reflectance spectroscopy—principal components regression analysis of soil properties. Soil Science Society of America Journal 65, 480–490. Chodak, M., Khanna, P., Beese, F., 2003. Hot water extractable C and N in relation to microbiological properties of soils under beech forests. Biology and Fertility of Soils 39, 123–130. Christy, C.D., 2008. Real-time measurement of soil attributes using on-the-go near infrared reflectance spectroscopy. Computers and Electronics in Agriculture 61 (1), 10–19.

D'Acqui, L.P., Pucci, A., Janik, L.J., 2010. Soil properties prediction of western Mediterranean islands with similar climatic environments by means of mid-infrared diffuse reflectance spectroscopy. European Journal of Soil Science 61 (6), 865–876. Daniel, K.W., Tripathi, N.K., Honda, K., 2003. Artificial neural network analysis of laboratory and in situ spectra for the estimation of macronutrients in soils of Lop Buri (Thailand). Australian Journal of Soil Research 41 (1), 47–59. Du, Y.P., Liang, Y.Z., Jiang, J.H., Berry, R.J., Ozaki, Y., 2004. Spectral regions selection to improve prediction ability of PLS models by changeable size moving window partial least squares and searching combination moving window partial least squares. Analytica Chimica Acta 501, 183–191. Hara, Y., 1999. Research on precision farming in Hokkaido—studies on the field management system, and farm implement control system using autonomous vehicle for large-scale farming. JSAM Journal 61 (4), 19–23. Imade Anom, S.W., Shibusawa, S., Sasao, A., Hirako, S., 2001. Soil parameters maps in paddy field using the real time soil spectrophotometer. JSAM Journal 63 (3), 51–58. Islam, K., Singh, B., Schwenke, G., McBratney, A., 2004. Evaluation of vertosol soil fertility using ultra-violet, visible and near-infrared reflectance spectroscopy. SuperSoil 2004: Third Australian New Zealand Soils Conference, 5–9 December 2004, University of Sydney, Australia (On CDROM). Jiang, J.H., Berry, R.J., Siesler, H.W., Ozaki, Y., 2002. Wavelength interval selection in multicomponent spectral analysis by moving window partial least-squares regression with applications to mid-infrared and near-infrared spectroscopic data. Analytical Chemistry 74, 3555–3565. Kawamura, S., Arakawa, M., Funatsu, K., 2006. Development of genetic algorithm-based wavelength regional selection technique. Journal of Computer Aided Chemistry 7, 10–17. Kodaira, M., Shibusawa, S., Ninomiya, K., Kato, Y., 2009. Farm mapping techniques for effective soil management in large-scale farming. JSAI Journal 18 (3), 110–121. Kusakawa, T., Matsumaru, T., Aoyagi, S., 2003. Effects of nitrogen fertilizer levels as related to application method under plastic mulch on growth and yield of carrot. JSHS Journal 72 (5), 432–439.

M. Kodaira, S. Shibusawa / Geoderma 199 (2013) 64–79 Maleki, M.R., Van Holm, L., Ramon, H., Merckx, R., De Baerdemaeker, J., Mouazen, A.M., 2006. Phosphorus sensing for fresh soils using visible and near infrared spectroscopy. Biosystems Engineering 95 (3), 425–436. Maleki, M.R., Mouazen, A.M., De Ketelaere, B., Ramon, H., De Baerdemaeker, J., 2008. On-the-go variable-rate phosphorus fertilisation based on a visible and nearinfrared soil sensor. Biosystems Engineering 99, 35–46. Martens, H., Naes, T., 1989. Multivariate Calibration, 2nd edition. John Wiley & Sons, Ltd., Chichester, United Kingdom. Matsunaga, T., Uwasawa, M., 1992. Application of near infrared spectrometry to quantitative analysis of soil physical and chemical properties. JSSSPN Journal 63 (6), 712–714. Mouazen, A.M., De Baerdemaeker, J., Ramon, H., 2005. Towards development of online soil moisture content sensor using a fibre-type NIR spectrophotometer. Soil & Tillage Research 80 (1–2), 171–183. Mouazen, A.M., Maleki, M.R., De Baerdemaeker, J., Ramon, H., 2007. On-line measurement of some selected soil properties using a VIS–NIR sensor. Soil & Tillage Research 93, 13–27. Mouazen, A.M., Kuang, B., De Baerdemaeker, J., Ramon, H., 2010. Comparison among principal component, partial least squares and back propagation neural network analyses for accuracy of measurement of selected soil properties with visible and near infrared spectroscopy. Geoderma 158, 23–31. Norris, K.H., 1989. Introduction—definition of NIRS analysis. In: Marten, G.C., Shenk, J.S., Barton II, F.E. (Eds.), Near Infrared Reflectance Spectroscopy (NIRS): Analysis of Forage Quality. : USDA Agriculture Handbook, No. 643. Agricultural Research Service, Washington DC, p. 6. Savitzky, A., Golay, M.J.E., 1964. Smoothing and differentiation of data by simplified least squares procedures. Analytical Chemistry 36, 1627–1639. Shepherd, K.D., Walsh, M.G., 2002. Development of reflectance spectral libraries for characterization of soil properties. Soil Science Society of America Journal 66 (1), 988–998. Shibusawa, S., Hirako, S., Otomo, A., Li, M., 1999. Real-time underground soil spectrophotometer. JSAM Journal 61 (3), 131–133.

79

Shonk, J.L., Gaultney, L.D., Schulze, D.G., Van Scoyoc, G.E., 1991. Spectroscopic sensing of soil organic matter content. Transactions of ASAE 34, 1978–1984. Souma, S., Kikuchi, K., 1992. Diagnostic Criteria for Soil and Crop Nutrition—Analysis Method (Revised). Agriculture Research Department, Central Agricultural Experiment Station, Hokkaido Research Organization; Agricultural Administration Division, Department of Agriculture, Hokkaido Government, Hokkaido, Japan. Sugimoto, M., Fukami, K., Imazono, S., Inoue, E., 2011. Development of a labor-saving direct-mount Hiller for taro cultivation. JSFWR Journal 46 (1), 15–25. Umeda, H., Shibusawa, S., Okayama, T., Sakuma, D.Y., Kaho, T., Ninomiya, K., 2011. Study of the precision farming with soil maps describing environmental load using a real-time soil sensor. JSAM Journal 73 (1), 37–44. Viscarra Rossel, R.A., Behrens, T., 2010. Using data mining to model and interpret soil diffuse reflectance spectra. Geoderma 158, 46–54. 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, 59–75. Wetterlind, J., Stenberg, B., Söderström, M., 2010. Increased sample point density in farm soil mapping by local calibration of visible and near infrared prediction models. Geoderma 156, 152–160. Williams, P., Norris, K., 2001. Near-infrared Technology in the Agricultural and Food Industries. American Association of Cereal Chemiste, Inc., St. Paul, Minnesota, USA. Wold, H., 1975. Soft modeling by latent variables: the nonlinear iterative partial least squares approach. Perspective in Probability and Statistics, Paper in Honour of M. S. Bartlett. Academic Press, pp. 520–540. Wold, S., Martens, H., Wold, H., 1983. The multivariate calibration method in chemistry solved by the PLS method. In: Ruhe, A., Kagstrom, B. (Eds.), Proc. Conf. Matrix Pencils, Lecture Notes in Mathematics. Springer-Verlag, Heidelberg, pp. 286–293. Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: a basic tool of chemometrics. Chemometrics and Intelligent Laboratory Systems 58, 109–130. Yashiro, M., 2008. Partial mixing and applying technique of fertilizer or chemical in the ridge. JSAM Journal 70 (1), 16–19.