- Email: [email protected]
A prototype measuring system of soil bulk density with combined frequency domain reflectometry and visible and near infrared spectroscopy
Computers and Electronics in Agriculture 151 (2018) 485–491
Contents lists available at ScienceDirect
Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag
Original papers
A prototype measuring system of soil bulk density with combined frequency domain reflectometry and visible and near infrared spectroscopy Raed A. Al-Asadia, Abdul M. Mouazenb, a b
T
⁎
Department of Environmental Health, Al-Qasim Green University, Babil, Iraq Department of Environment, Ghent University, Coupure 653, 9000 Gent, Belgium
A R T I C LE I N FO
A B S T R A C T
Keywords: Bulk density Portable prototype multi-sensor Data fusion Near infrared spectroscopy Frequency domain reflectometry
A combined-penetrometer sensor prototype (CPSP) for the measurement of topsoil bulk density (BD) was developed and tested under field conditions. The prototype consisted of a standard penetrometer, equipped with a near infrared spectrophotometer (NIRS) (1650–2500 nm) to measure gravimetric moisture content (ω) and a frequency domain reflectometry (FDR) to measure volumetric moisture content (θv), while BD was assessed by the combination of both sensors’ data. The CPSP was tested in situ at five arable and two grassland fields of different soil texture classes in Silsoe, Bedfordshire, UK, during the period from August to December 2013. Artificial neural networks (ANN) were used to predict ω and θv based on data fusion of NIRS diffuse reflectance spectra and FDR output voltage (V), and the predicted values were substituted in a model to predict BD. The CPSP showed more accurate BD assessment in grass fields with root mean square error of prediction (RMSEp) of 0.077 g cm−3, compared to arable fields (RMSEp = 0.104 g cm−3). A collective BD model produced for arable and grass fields provided a moderate accuracy with a RMSEp of 0.102 g cm−3. It can be concluded that the new CPSP can be used successfully to measure BD in the topsoil by combining the NIRS and FDR techniques through ANN-data fusion approach.
1. Introduction Soil compaction created by different human and natural factors causes multiple environmental and agronomical problems. It has attracted scientists’ attention to carry out intensive research to understand the occurrence and propose solutions for avoidance and management for more than a century. Soil compaction is mainly occurred by the intensive use of the heavy agriculture machinery, repeated ploughing at the same depth in addition to the trampling of animals. However, soil compaction can also be found under natural conditions without human or animal intervention (Batey, 2009). Many studies indicated increases in soil strength, bulk density (BD), and tillage draught requirement as a result of soil compaction, while decreases in soil total porosity, soil aeration, water infiltration and saturated hydraulic conductivity were evidence of soil compaction (Hamza and Anderson, 2005). Understanding therefore how and to what extent soil compaction may be eliminated seems of vital importance to the future wellbeing of agricultural systems. A key requirement to manage soil compaction is by accurate measurement of associated parameter/s that should be done quickly and cost effectively in the field without the need for laboratory analyses that are time consuming, difficult and costly
⁎
Corresponding author. E-mail address: [email protected] (A.M. Mouazen).
https://doi.org/10.1016/j.compag.2018.06.045 Received 28 December 2017; Received in revised form 26 June 2018; Accepted 27 June 2018 Available online 30 June 2018 0168-1699/ © 2018 Elsevier B.V. All rights reserved.
procedures. In situ measurement of soil compaction is a tricky task to accomplish rapidly, easily and cost effectively, because of the complex nature of agricultural soils (Aragón et al., 2000; Horn et al., 2000; Mouazen and Ramon, 2006). Due to this fact, many scientists have developed various devices for measuring soil compaction. Apart from the traditional Kopercki rings and standard penetrometers commonly used to measure BD and penetration resistance, respectively, new approaches based on multi-sensor and data fusion approach were recently introduced. This was essential to avoid the shortcomings of Kopercki rings method (i.e., time consuming, difficult, and prone to error in dry conditions) and standard penetrometers (i.e., combined effects of moisture content, texture and BD on penetration resistance readings) reported earlier (Mouazen and Ramon, 2006). Therefore, there are now examples of penetrometers equipped with multi-sensors for the measurement of soil compaction. Vaz et al. (2001) reported the development of a combined sensor consisting of a penetrometer and a time domain reflectometry (TDR) to measure soil strength and volumetric moisture content (θv) simultaneously. Peter and Yurui (2004) designed a combined capacitance sensor with a cone penetrometer, to measure θv and penetration
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
resistance, respectively. Hummel et al. (2004) presented a combined probe consisting of a cone penetrometer and a near infrared spectrophotometer (NIRS) to measure penetration resistance and gravimetric moisture content (ω), respectively. Sheng et al. (2011) developed a combined penetrometer sensor kit consisting of two impedance soil moisture sensors, two soil temperature sensors, and an electrical conductivity (EC) sensor to monitor moisture dynamics in the soil. All these combined-penetrometer sensor porotypes (CPSP) were not implemented to predict BD. Quraishi and Mouazen (2013a) introduced a multi-sensor platform, which enabled the assessment of BD from the fusion of data on penetration resistance measured with a load cell and ω, clay content and organic matter (OM) measured with a NIRS sensor. This multi-sensor does not measure θv, necessary for the direct assessment of BD using the following equation (Wijaya et al., 2004): (1)
BD = θ v/ ω
Fig. 1. The combined-penetrometer sensor prototype (CPSP) of a near infrared spectrophotometer (NIRS) and a frequency domain reflectometry (FDR) sensor, for the measurement of soil gravimetric moisture content (ω) and volumetric moisture content (θv), respectively, and assessment of bulk density (BD).
−3
where BD is the soil bulk density in g cm , θv is the volumetric moisture content in cm3 cm−3 and ω is the gravimetric moisture content in g g−1. Therefore, there is a need for a modified penetrometer sensing kit that enables simultaneous measurement of both θv and ω, and then derive BD values using Eq. (1). The aim of this paper was to design and evaluate a CPSP, consisting of a NIRS and a FDR sensor. The developed CPSP kit will be tested for the measurement of the top soil BD under field measurement conditions. 2. Materials and methods 2.1. Development of the combined-penetrometer sensor prototype (CPSP) Al-Asadi and Mouazen (2014) introduced a proof of concept of a new measurement system of BD, based on measurement of θv and ω values with a FDR and NIRS, respectively, which are then substituted into Eq. (1) to assess BD. In a recent work, Mouazen and Al-Asadi (2018) studied the effect of moisture content on the accuracy of BD measurement. These two studies enabled understanding the requirements for designing a new CPSP for the measurement of BD based on a standard penetrometer. The new design to be reported in this work consists of a NIRS, a dielectric (FDA) sensor, a standard penetrometer, a global positioning system (GPS) a battery and a laptop. The NIRS used was Avantes® portable model NIR200-2.6 (Avantes, Eerbeek, the Netherlands), which has a dual stage thermo-electrical Peltier-cooled InGaAs single detector with 256 pixels, with a spectral range of 1650–2500 nm and 7 nm resolution. This spectrometer was connected to a laptop through a high-speed USB2.0 interface and AvaSoft 7.7 software (Avantes, Eerbeek, the Netherlands). The spectrometer was also connected to an assembly of a rod and a 30 degree, 1.26 cm2 base-area cone through two optical fibres (Quraishi and Mouazen, 2013a). An electronically stabilized 20 W halogen lamp is used as a light source, and light was transferred to the soil profile through an illumination fibre connected at one end to the spectrometer and at the other end to a sapphire window (Fig. 1). The diffuse reflected light passing the sapphire window was collected back to the spectrometer by a detecting fibre. The system operates in situ using 24VDC and 20 Watts lead-acid battery as a power source. A 50 channel global positioning system (GPS) (eTrex 60C model, Garmin, USA) was used to record the sampling location. The measurement of θv was done by a new designed FDA sensor, which consists of an electronic circuit generating a 100 MHz electromagnetic sine wave, propagated through the soil body by the central electrode in the form of a copper ring with a 10, 15 and 1.5 mm height, diameter and wall thickness, respectively. The copper ring is insulated from the probe body, which forms two shielding electrodes as they are connected to the electronic circuits’ negative. Each shielding electrode has a cylinder shape with a 13 and 50 mm diameter and height, respectively (Fig. 2). The readout pin of the electronic circuit is connected
Fig. 2. Shows the electromagnetic fringe fields around the dielectric sensors’ electrodes.
to the HH2 meter (Delta-T devices, Cambridge, UK), which at the time of reading acts as a power supply and provides data storage. The probe body also provides protection for the optical fibres, as they run inside its cavity, as described above. As the combined probe inserted into the soil vertically, the surrounding soil in contact with probe electrodes will be affected by the fringe fields of the propagated signal, resulting from the two capacitors (Fig. 2).
2.2. Experimental procedure The CPSP portable measurement system was tested in situ in seven fields with various textures and growing crops (Table 1). The average field particle size distribution (PSD) was measured, using the sieving and sedimentation method (British Standards, 1998). The average soil Table 1 Information of the seven fields used for testing the combined-penetrometer sensor prototype (CPSP). Fields
Nr.
Soil texture
Clay%
Silt%
Sand%
OM%
Crop
Avenue 1 Beechwood Clover hill Orchard Showground
20 20 20 20 20
Sandy loam Cay Clay loam Clay loam Sandy clay loam
17 66 35 33 24
20 11 24 26 17
63 23 41 41 59
3.6 5.8 4.8 4.15 3.34
Barley Wheat Barley Wheat Barley
Avenue 2 Onley
50 50
Sandy loam Clay
29 60
20 30
51 10
2.98 5.4
Grass
OM is organic matter 486
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
Table 2 Sample statistics of the soil samples, used for testing the new combined-penetrometer sensor prototype (CPSP). Values were obtained from laboratory measured volumetric moisture content (θv), in cm3 cm−3, gravimetric moisture content (ω) in g g−1 and bulk density (BD) in g cm−3. Statistics
Maximum Minimum Range Average SD
Arable fields
Grass fields
θv
ω
BD
θv
ω
BD
0.55 0.12 0.43 0.34 0.14
0.41 0.10 0.31 0.26 0.11
1.60 1.08 0.52 1.34 0.11
0.46 0.24 0.22 0.35 0.09
0.36 0.15 0.21 0.25 0.08
1.78 1.20 0.58 1.49 0.24
SD is standard deviation.
accuracy can be achieved (Mouazen et al., 2005). After noise elimination, more noise reduction was also achieved by means of averaging every 10 nm wavelengths to one. A maximum normalisation was followed, which is typically used to get all data to approximately the same scale, or to get a more even distribution of the variances and the average values. Spectra were then subjected to the Savitzky–Golay first derivation (Martens and Naes, 1989), which enabled the transformation of spectra data to the first or higher order derivatives, including a smoothing factor. This method determines how many adjacent variables will be used to estimate the polynomial approximation used for derivatives. A second order polynomial approximation was selected due to its performance in producing more accurate calibration models. The final process of the pre-treatment was smoothing, which was carried out at a 2:2 rate, in order to reduce further the noise in spectra. All pretreatment steps were carried out using the Unscrambler 7.8 software (Camo Inc.; Oslo, Norway). Artificial neural network (ANN) was selected in this study for modelling and data fusion of both sensors’ data, and analyses were performed using Statistica software (StatSoft, USA, 2011). Neural networks are simplified models of the biological structure of human brains (Günaydin, 2009). There are three main layers in the ANN structure, namely, a set of input nodes, one or more layers of hidden nodes and a set of output nodes. In this study, different number of nodes was used in each layer depending on the input data used. The powerful second order Broyden–Fletcher–Goldfarb–Shanno (BFGS) training algorithms were used, with different transfer functions assigned for hidden and output layers, as detailed in Table 3. The number of neurons in the hidden layer is established by training several networks with different number of hidden neurons, and comparing the predicted with measured values. Since a previous study by Al-Asadi and Mouazen (2014) confirmed that data fusion provided the best predictions results for the measurement of θv, ω and BD, this approach was adopted in this study, so that the input data for each ANN were V and NIR spectra and the output were either θv or ω. For modelling of both, the data of the five arable fields (100 samples) and the two grass fields (100 samples), were each divided into training and cross-validation set (75 samples), and validation set (25 samples). The training samples were used to generate two separate ANN calibration models: one for the arable fields and the other one for the grass fields. Additional model named collective calibration model was generated by using 75% of the 200 arable and grass fields’ samples (e.g., 150 samples) for training and the remaining 25% (e.g., 50 samples) were the validation set. Accuracy of calibration models was evaluated using the independent validation sets, and results are reported for the independent validation sets only. By substituting the predicted θv and ω obtained from the ANN analysis into Eqn (1), BD was assessed and compared to the measured values obtained from the oven drying method of soil samples collected at the same positions of the CPSP readings in the field. Model prediction performance was evaluated by means of coefficient of determination
Fig. 3. The combined-penetrometer sensor prototype (CPSP) for the field measurement of soil volumetric moisture content (θv) and gravimetric moisture content (ω) and assessment of bulk density (BD), assembled on a sack barrow.
OM was measured with a TrusSpecCNS spectrometer (LECO Corporation, St. Joseph, MI, USA), using the Dumas combustion method (British Standards, 2000). Table 1 shows the results of laboratory analyses for OM, PSD and soil texture class, classified according to the United States Department of Agriculture (USDA) classification system (Soil Survey Staff, 1999). The light weight and compact size of the CPSP made it easy to move it through the growing crops and carry out the measurements by assembling the system on a sack barrow (Fig. 3). The experiment ran from August 2013 to December 2013, at the Silsoe experimental farm of Cranfield University, Bedfordshire, UK. The penetrometer was used to collect data at 200 points selected randomly in the seven experimental sites. A total of 20 readings per field were collected in the five arable fields, whereas 50 readings per field were collected in the two grass fields (Table 1). Readings were taken after pushing the penetrometer vertically into the soil to 10 cm depth, and then NIR spectra and FDA V were collected together with the GPS readings. A total of 200 assessments of θv, ω and BD in the seven sites were first made using the CPSP, and these assessed values were then validated by comparing with measured values, obtained by drying Kopercki ring collected soil samples at 105 °C for 24 h (British Standards, 2007). Table 2 provides basic statistics of oven drying measured θv, ω and BD of the 200 soil samples, collected during the field testing of the portable CPSP. 2.3. Modelling Before modelling, the NIR spectra has to be pre-treated in order to reduce spurious peaks that do not contain any physical or chemical spectra information and to correct the scatter effects. The same pretreatments were applied on all spectra. The spectra range was first reduced to 1650–2250 nm, to eliminate noise at both edges of the detected wavelength range, so that enhancement of the calibration model 487
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
Table 3 Artificial neural network (ANN) modelling results of the field testing of the combined-penetrometer sensor prototype (CPSP). Index
Network structure
Training R2
Validation R2
Training error
Validation error
Training algorithm
Hidden activation
Output activation
Arable fields Grass fields Collective
170-17-2 170-8-2 170-15-2
1.00 0.99 0.97
0.98 0.99 0.95
0.000008 0.000052 0.000111
0.000458 0.000065 0.000652
BFGS 147 BFGS 113 BFGS 178
Exponential Logistic Exponential
Exponential Exponential Exponential
BFGS: Broyden-Fletcher-Goldfarb-Shanno algorithm; Logistic: logistic sigmoid function; Exponential: negative exponential function.
grass fields), Table 4 shows excellent accuracy of ω prediction.
(R2), root mean square error of prediction (RMSEp) and ratio of prediction deviation (RPD), which was calculated as the ratio of standard deviation of laboratory measured values divided by RMSEp.
3.2. Evaluation of sensor performance for bulk density measurement The prediction of BD in the arable fields revealed lower accuracy results with values of R2 of 0.34 and RMSEp of 0.104 g cm−3 (Table 4), compared to R2 of 0.94 and RMSEp of 0.04 g cm−3 reported by Quraishi and Mouazen (2013a), who used a prototype measuring system that combined a NIRS sensor and a penetrometer. The grass fields’ results show a better accuracy in comparison with the measurements from arable fields’ model, with the validation R2, RMSEp and RPD values of 0.47, 0.077 g cm−3 and 1.36, respectively. The collective model in predicting soil BD revealed comparable results to that of the grass fields’ model with R2 = 0.52 and RMSEp = 0.102 g cm−3. Comparing the scatter plots revealed that BD points are the closest to the 1:1 line for the collective model (Fig. 6C), while BD points of the arable fields’ model are most scattered (Fig. 6A). The smaller intercept of 0.2 g cm−3 and the better slope of 0.84 of the collective model (Fig. 6C), as compared to the other two models indicate better assessment performance, suggesting the use of a large number of soil samples collected from arable and grass soils has improved the prediction accuracy of BD. Literature shows no similar studies about the assessment of BD, as a function of θv and ω measured with a dielectric probe and NIRS, respectively. Therefore, the CPSP introduced in the current study proves to be unique in the assessment of BD, and also the potential the sensor has in the prediction of other soil properties e.g., with NIRS that are important for land management. However, further development is needed so as to improve the assessment accuracy obtained so far. This might concern improvement in the technical specification or calibration procedure followed in the present study.
3. Results 3.1. Evaluation of sensor performance for moisture measurement The relation between the FDA sensor V and θv was somewhat weaker when the measurements were taken in the arable fields compared to those taken in grass fields. However, this is likely to be due to the heterogeneous nature of topsoil layer of arable soils due to repeated tillage. While the R2, RMSEp, and RPD values for arable soils may be lower than grass soils, but they are both considered to be within the excellent prediction level (Table 4), with values of 0.97, 0.024 cm3 cm−3 and 5.80, 1.00, 0.005 cm3 cm−3 and 13.72, and 0.94, 0.039 cm3 cm−3 and 3.67 for arable, grass and the collective models, respectively. Similarly, the best slope and smallest x intercept of linear line indicate predictions in grass fields are in advance of other two models (Fig. 4). Kaleita et al. (2005) reported less successful field calibration of a dielectric sensor for the measurement of θv with a R2 value of 0.77. Similarly, θv result obtained in the present work is more accurate than those reported by Andrade-Sanchez et al. (2001), with R2 of 0.78. The performance of the NIRS for the prediction of ω was rather different compared to the dielectric probe performance for the measurement of θv. Smaller differences in model performance between three models for ω prediction can be observed in Table 4 and Fig. 5. Values of R2, RMSEp and RPD for the prediction of ω are 0.97, 0.019 g g−1 and 5.46, and 0.96, 0.011 g g−1 and 4.73, for arable and grass fields’ soils, respectively, and 0.96, 0.023 g g−1 and 4.56 for the collective model, respectively, which are of similar magnitude to those obtained by Mouazen et al. (2006) and closer to those reported by Quraishi and Mouazen (2013b), who found 0.95, 2.39% and 4.15 values, respectively. This is despite the fact that in the present study spectra were collected from multiple fields (e.g., five arable and two
4. Discussion 4.1. Evaluation of the new combined-penetrometer sensor prototype (CPSP) According to Hemmat and Adamchuk (2008), the current measuring system can be categorized under water content sensors that measures ‘indirectly’ soil compaction. Mosaddeghi et al. (2007) stated that we should not only rely on the strain-related properties as the dependent variables for the assessment of soil compaction, whereas different soil properties can also be considered as a sign of compaction. The CPSP kit introduced in this work showed encouraging accuracy for measuring soil BD under field conditions (Table 4). A NIRS with a short spectral range of 1650–2500 nm was used for the CPSP. The process of combining NIRS and FDR sensors was the most challenging task of this work, as both sensors are very sensitive to any changes made to their sensing probes or heads. Delicate materials have to be used in the sensor manufacture, for example, using thin fibre optics in the limited space inside the steel shaft, upon which the combined probe is assembled and a sapphire material was used in a protective window design (Fig. 1) to ensure that no scratches would affect the probe windows’ transparency. The accuracy of the dielectric probe depends on maintaining good contact between the probe’s electrodes and the soil. For this reason, a penetrating cone is used as one of the shielding electrodes in addition to a second part above the central electrode (Fig. 1). A similar approach
Table 4 Prediction results of soil volumetric moisture content (θv), gravimetric moisture content (ω) and bulk density (BD), obtained by using the combinedpenetrometer sensor prototype (CPSP) in arable fields, grass fields and collective (e.g., both arable and grass fields). Arable fields
Grass fields
Collective
θv R2 RMSEp, cm3 cm−3 RPD
0.97 0.024 5.80
1.00 0.005 13.72
0.94 0.039 3.67
ω R2 RMSEp, g g−1 RPD
0.97 0.019 5.46
0.96 0.011 4.73
0.96 0.023 4.56
BD R2 RMSEp, g cm−3 RPD
0.34 0.104 1.08
0.47 0.077 1.36
0.52 0.102 1.21
R2: coefficient of determination. RMSEp: root mean square error of prediction; and, RPD: residual prediction deviation. 488
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
0.6
0.6
Estimated θv (cm3 cm-3)
A
B
0.4
0.4
0.2
0.2
y = 0.96x + 0.016 R² = 0.97
y = 1.001x - 0.003 R² = 0.995
0.0
0.0 0.0
0.2
0.4
0.6
0.0
0.2
0.4
0.6
0.6
C 0.4
0.2
y = 1.07x - 0.02 R² = 0.94 0.0 0.0
0.2
0.4 3
0.6 -3
Measured șv (cm cm ) Fig. 4. Scatter plots of core sampling measured versus artificial neural networks (AAN) predicted soil volumetric moisture content (θv) using the combinedpenetrometer sensor prototype (CPSP) for field measurements in arable fields (A), grass fields (B) and the collective fields (C).
measuring system has demonstrated that moderate accuracy of BD measurement can be achieved for a wide range of soil types and agricultural practices with less input data needed (e.g., θv and ω only), compared to others reviewed works. For example, Quraishi and Mouazen (2013a) used a similar prototype consisting of a penetrometer and a NIRS for measurement of topsoil BD, reporting RMSEp of 0.04 g cm−3, which is indeed smaller than the RMSEp achieved in this study (0.077 – 0.104 g cm−3). However, large number of variables was needed as input data for the ANN analysis to predict BD by Quraishi and Mouazen (2013a), which included ω, OM, penetration resistance and clay content.
was followed by Peter and Yurui (2004) for the design of a combined capacitance sensor with a cone penetrometer. The Peter and Yurui’ system overestimated penetration resistance compared to the readings of a standard cone penetrometer. They concluded that adding the capacitance sensor electrode and insulator caused additional frictional resistance to soil penetration by the combined sensor, which subsequently led to deterioration in measurement accuracy for the penetration resistance. Such a problem has no effect on the soil moisture related systems to measure soil BD designed and developed in the present study. This is in-line with the literature review, which showed shortcoming of soil strength related measuring systems. The NIRS of the CPSP predicted ω accurately in arable and grassland soils with R2 of 0.97 and 0.96 and RMSEp 0.019 and 0.011 g g−1, respectively (Table 4), compared to a lower accuracy magnitude reported by Hummel et al. (2004), who presented a combined probe consisting of a cone penetrometer and NIRS. They achieved a R2 of 0.90 and the standard error of calibration and prediction of soil ω were 1.97% and 2.38%, respectively. Their combined system has predicted the cone index with R2 = 0.86 compared to the standard cone index, using all the data from clay loam, silt loam and sandy loam soil textures and the entire range of soil moisture content. However, they did not attempt to measure soil BD. In general, the CPSP developed and tested in the present work showed moderate accuracy to estimate BD in the top soil, under various crops, soil types and conditions. The growing crop effect showed that grass fields’ soils are in advance of arable fields soils planted with different crops, but the authors believe that there are no direct connections between the accuracy and the growing crops. Literature indicated that many authors have attempted to provide an accurate soil BD measuring system, due to the importance of such a single soil physical property that has many environmental and economic impacts. The new
4.2. Advantages and practical implementation challenges The advantages of using the new CPSP to measure soil compaction indicated as BD with the implementation of FDR and NIRS sensors are as followed:
• The system is semi- non-invasive, where soil penetration should take
• • 489
place in the vertical direction while collecting data. The FDR sensor measures the dielectric constants of the compound by emitting an electromagnetic signal propagated through the soil body, whereas the NIRS sensor collect the diffused reflectance of electromagnetic wave of light from the soil samples surface down to the depth of 2 mm. The system is relatively small in size and light in weight and can be mounted on a wheelbarrow. It is of a robust design in terms of the penetration rod geometrical structure, which enables smooth penetration of the soil, without blocking of the sapphire window from which NIR spectra are collected. It has an additional advantage by presenting a combined probe
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
0.5
0.5
A
0.4 0.3
0.3
0.2
0.2
Estimated ω (g g-1)
0.1
B
0.4
0.1
y = 0.98x + 0.0034 R² = 0.97
y = 1.03x - 0.01 R² = 0.96
0.0
0.0 0.0
0.1
0.2
0.3
0.4
0.0
0.5
0.1
0.2
0.3
0.4
0.5
0.5
C
0.4 0.3 0.2 0.1
y = 1.1x - 0.0154 R² = 0.96
0.0 0.0
0.1
0.2
0.3
0.4
0.5
-1
Measured Ȧ (g g ) Fig. 5. Scatter plots of core sampling measured versus artificial neural networks (AAN) predicted soil gravimetric moisture content (ω) using the combined-penetrometer sensor prototype (CPSP) for field measurements in arable fields (A), grass fields (B) and the collective fields (C).
containing both sensors, with which readouts can be recorded within few seconds. It is cost effective, efficient and long lasting system for measuring soil BD and has the potential to measure other soil physical and chemical properties using the NIRS. Adequate accuracy for BD assessment with a wide range of soil types and conditions can be obtained, providing accurate calibration models are developed in advance.
main challenge in the manufacturing process of the CPSP, which was particularly due to the difficulty of dealing with the high frequency signals (100 MHz). With more research time and financial support a whole profile measuring probe instead of just the top 10 cm of depth measurements would have been developed. This will be considered for future work.
Despite the above advantageous of the CPSP, few challenges can be discussed. One of the challenges raised during the field measurements was the fact that agricultural soils are not naturally homogeneous and contain stones, gravels and plant residues. These have a significant negative impact on the FDR sensor, relating to the fact that the central electrode is very sensitive to air pockets when they present around it. Stones in the collected soil cores lead to soil moisture estimation error, as the stones do not hold any moisture. Furthermore, dry soil conditions make the measurements difficult and less accurate (Mouazen and AlAsadi, 2018), while the very moist soil status could make it susceptible to block the sapphire window, particularly in clay soils. Technical challenges can be summarised as the necessity for a more compact system, which has the capability of recording the data from both sensors, as well as to process the adapted calibration models and providing the desired output in real time, which may include spatial maps or prediction values. Long life and light weight batteries are essential for the measuring system to be easily mobile and reliable for extended periods of field measurement. However, assembling the system on a quad motorbike could solve most of the above technical issues. Supplying the right coaxial cable for the dielectric sensor was the
A portable combined-penetrometer sensor prototype (CPSP) for the measurements of volumetric moisture content (θv), gravimetric moisture content (ω) and bulk density (BD) in the topsoil was developed and successfully tested under field conditions across a wide range of soil textures and land uses. Results showed the prediction of θv and ω provided excellent results, whereas the BD assessment results were encouraging. The best topsoil BD prediction accuracy was for grass fields’ soils, as the smallest root mean square error of prediction (RMSEp) of 0.077 g cm−3 was recorded. A collective BD model produced for arable and grass fields’ soils provided a moderate accuracy with a RMSEp of 0.102 g cm−3. A narrow wavelength band spectrometer (e.g., 1650–2500 nm) or even mono optical detectors of a certain wavelength associated with OH absorption overtones at 1450 or 1950 nm can also be used to predict ω effectively, which would considerably lower the cost of the new CPSP and would present it as replacement for the high cost traditional laboratory measuring methods.
• •
5. Conclusions
Acknowledgments Authors acknowledge the financial support of the UK Engineering 490
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
2.0
2.0
A
Estimated BD (g cm-3)
1.8
B
1.8
1.6
1.6
1.4
1.4
y = 0.66x + 0.48 R² = 0.344
1.2
1.2
1.0
y = 0.67x + 0.48 R² = 0.47
1.0 1.0
1.2
1.4
1.6
1.8
2.0
1.0
1.2
1.4
1.6
1.8
2.0
2.0
C
1.8 1.6 1.4 1.2
y = 0.84x + 0.2 R² = 0.52
1.0 1.0
1.2
1.4
1.6
1.8
2.0
-3
Measured BD (g cm ) Fig. 6. Scatter plots of core sampling measured versus artificial neural networks (AAN) predicted soil bulk density (BD) using the combined-penetrometer sensor prototype (CPSP) for field measurements in arable fields (A), grass fields (B) and the collective fields (C).
and Physical Science Research Council (EPSRC) and The Douglas Bomford Trust. The corresponding author acknowledges the financial support of the Flemish Scientific Research (FWO) for funding of SiTeMan Odysseus I Project (Nr. G0F9216N).
607–618. Kaleita, A.L., Heitman, J.L., Logsdon, S.D., 2005. Field calibration of the ThetaProbe for der FOR Des Moines lobe soils. Am. Soc. Agric. Eng., Appl. Eng. Agric. 21 (5), 865–870. Martens, H., Naes, T., 1989. Multivariate Calibration, second ed. John Wiley & Sons Ltd, Chichester. Mosaddeghi, M.R., Koolen, A.J., Hajabbasi, M.A., Hemmat, A., Keller, T., 2007. Suitability of pre-compression stress as the real critical stress of unsaturated agricultural soils. Biosyst. Eng. 98, 90–101. Mouazen, A.M., De Baerdemaeker, J., Ramon, H., 2005. Towards development of on-line soil moisture content sensor using a fibre-type NIR spectrophotometer. Soil Tillage Res. 80, 171–183. Mouazen, A.M., Ramon, H., 2006. Development of on-line measurement system of bulk density based on on-line measured draught, depth and soil moisture content. Soil Tillage Res. 86, 218–229. Mouazen, A.M., Karoui, R., De Baerdemaeker, J., Ramon, H., 2006. Characterization of soil water content using measured visible and near infrared spectra. Soil Sci. Soc. Am. J. 70, 1295–1302. Mouazen, A.M., Al-Asadi, R.A., 2018. Influence of soil moisture content on assessment of bulk density with combined frequency domain reflectometry and visible and near infrared spectroscopy under semi field conditions. Soil Tillage Res. 176, 95–103. Peter, S.L., Yurui, S., 2004. Combined sensor for simultaneous investigation of cone index and soil water content. ASAE/CSAE 041046. Quraishi, Z., Mouazen, A.M., 2013a. A prototype sensor for the assessment of soil bulk density. Soil Tillage Res. 134, 97–110. Quraishi, M.Z., Mouazen, A.M., 2013b. Development of a methodology for in situ assessment of topsoil dry bulk density. Soil Tillage Res. 126, 229–237. Sheng, W., Sun, Y., Schulze Lammers, P., Schumann, H., Berg, A., Shi, C., Wang, C., 2011. Observing soil water dynamics under two field conditions by a novel sensor system. J. Hydrol. 409 (2011), 555–560. Soil Survey Staff, 1999. Soil Taxonomy – a basic system of soil classification for making and interpreting soil surveys; second edition. Agricultural Handbook 436; Natural Resources Conservation Service, USDA. Washington DC, USA. Vaz, C.M.P., Bassoi, L.H., Hopmans, J.W., 2001. Contribution of water and bulk density of field soil penetration resistance as measurement by a combined cone penetrometerTDR probe. Soil Tillage Res. 60, 35–42. Wijaya, K., Nishimura, Y., Kato, M., Nakagawa, M., 2004. Field estimation of soil dry bulk density using amplitude domain reflectometry data. J. Jpn. Soc. Soil Phys. 97, 3–12.
References Al-Asadi, R.A., Mouazen, A.M., 2014. Combining frequency domain reflectometry and visible and near infrared spectroscopy for measurement of soil bulk density. Soil Tillage Res. 135, 60–70. Andrade-Sanchez, P., Aguera, J., Upadhyaya, S.K., Jenkins, B.M., Rosa, U.A., Josiah, M., 2001. Evaluation of a dielectric based moisture and salinity sensor for in-situ applications. Paper No. 01-1010, ASAE, St. Joseph, Michigan. Aragón, A., García, M.G., Filgueira, R.R., Pachepsky, Y.A., 2000. Maximum compactability of Argentine soils from the Proctor test; the relationship with organic carbon and water content. Soil Tillage Res. 56, 197–204. Batey, T., 2009. Soil compaction and soil management – a review. Soil Use Land Manage. 25 (4), 335–345. British Standards, 1998. Soil Quality: BS 7755: Section 5.4: 1998. Part 5: Physical Methods. Section 5.4: Determination of Particle Size Distribution in Mineral Soil Material – Method by Sieving and Sedimentation. British Standards Institution, UK. British Standards, 2000. Soil Improvers and Growing Media: BS EN 13039:2000. Determination of Organic Matter Content and Ash. British Standards Institution, UK. British Standards, 2007. Soil Improvers and Growing Media: BS EN 13040:2007. Sample Preparation for Chemical and Physical Tests, Determination of Dry Matter Content, Moisture Content and Laboratory Compacted Bulk Density. British Standards Institution, UK. Günaydin, O., 2009. Estimation of soil compaction parameters by using statistical analyses and artificial neural networks. Environ. Geol. 57, 203–215. Hamza, M.A., Anderson, W.K., 2005. Soil compaction in cropping systems: a review of the nature, causes and possible solutions. Soil Tillage Res. 82 (2), 121–145. Hemmat, A., Adamchuk, V.I., 2008. Sensor systems for measuring soil compaction: review and analysis. Comput. Electron. Agric. 63, 89–103. Horn, R., van den Akker, J.J.H., Arvidsson, J., 2000. Subsoil Compaction: Distribution, Processes and Consequences. Deutsche Bibliothek, Germany. Hummel, J.W., Ahmad, I.S., Newman, S.C., Sudduth, K.A., Drummond, S.T., 2004. Simultaneous soil moisture and cone index measurement. Am. Soc. Agric. Eng. 47 (3),
491
Contents lists available at ScienceDirect
Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag
Original papers
A prototype measuring system of soil bulk density with combined frequency domain reflectometry and visible and near infrared spectroscopy Raed A. Al-Asadia, Abdul M. Mouazenb, a b
T
⁎
Department of Environmental Health, Al-Qasim Green University, Babil, Iraq Department of Environment, Ghent University, Coupure 653, 9000 Gent, Belgium
A R T I C LE I N FO
A B S T R A C T
Keywords: Bulk density Portable prototype multi-sensor Data fusion Near infrared spectroscopy Frequency domain reflectometry
A combined-penetrometer sensor prototype (CPSP) for the measurement of topsoil bulk density (BD) was developed and tested under field conditions. The prototype consisted of a standard penetrometer, equipped with a near infrared spectrophotometer (NIRS) (1650–2500 nm) to measure gravimetric moisture content (ω) and a frequency domain reflectometry (FDR) to measure volumetric moisture content (θv), while BD was assessed by the combination of both sensors’ data. The CPSP was tested in situ at five arable and two grassland fields of different soil texture classes in Silsoe, Bedfordshire, UK, during the period from August to December 2013. Artificial neural networks (ANN) were used to predict ω and θv based on data fusion of NIRS diffuse reflectance spectra and FDR output voltage (V), and the predicted values were substituted in a model to predict BD. The CPSP showed more accurate BD assessment in grass fields with root mean square error of prediction (RMSEp) of 0.077 g cm−3, compared to arable fields (RMSEp = 0.104 g cm−3). A collective BD model produced for arable and grass fields provided a moderate accuracy with a RMSEp of 0.102 g cm−3. It can be concluded that the new CPSP can be used successfully to measure BD in the topsoil by combining the NIRS and FDR techniques through ANN-data fusion approach.
1. Introduction Soil compaction created by different human and natural factors causes multiple environmental and agronomical problems. It has attracted scientists’ attention to carry out intensive research to understand the occurrence and propose solutions for avoidance and management for more than a century. Soil compaction is mainly occurred by the intensive use of the heavy agriculture machinery, repeated ploughing at the same depth in addition to the trampling of animals. However, soil compaction can also be found under natural conditions without human or animal intervention (Batey, 2009). Many studies indicated increases in soil strength, bulk density (BD), and tillage draught requirement as a result of soil compaction, while decreases in soil total porosity, soil aeration, water infiltration and saturated hydraulic conductivity were evidence of soil compaction (Hamza and Anderson, 2005). Understanding therefore how and to what extent soil compaction may be eliminated seems of vital importance to the future wellbeing of agricultural systems. A key requirement to manage soil compaction is by accurate measurement of associated parameter/s that should be done quickly and cost effectively in the field without the need for laboratory analyses that are time consuming, difficult and costly
⁎
Corresponding author. E-mail address: [email protected] (A.M. Mouazen).
https://doi.org/10.1016/j.compag.2018.06.045 Received 28 December 2017; Received in revised form 26 June 2018; Accepted 27 June 2018 Available online 30 June 2018 0168-1699/ © 2018 Elsevier B.V. All rights reserved.
procedures. In situ measurement of soil compaction is a tricky task to accomplish rapidly, easily and cost effectively, because of the complex nature of agricultural soils (Aragón et al., 2000; Horn et al., 2000; Mouazen and Ramon, 2006). Due to this fact, many scientists have developed various devices for measuring soil compaction. Apart from the traditional Kopercki rings and standard penetrometers commonly used to measure BD and penetration resistance, respectively, new approaches based on multi-sensor and data fusion approach were recently introduced. This was essential to avoid the shortcomings of Kopercki rings method (i.e., time consuming, difficult, and prone to error in dry conditions) and standard penetrometers (i.e., combined effects of moisture content, texture and BD on penetration resistance readings) reported earlier (Mouazen and Ramon, 2006). Therefore, there are now examples of penetrometers equipped with multi-sensors for the measurement of soil compaction. Vaz et al. (2001) reported the development of a combined sensor consisting of a penetrometer and a time domain reflectometry (TDR) to measure soil strength and volumetric moisture content (θv) simultaneously. Peter and Yurui (2004) designed a combined capacitance sensor with a cone penetrometer, to measure θv and penetration
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
resistance, respectively. Hummel et al. (2004) presented a combined probe consisting of a cone penetrometer and a near infrared spectrophotometer (NIRS) to measure penetration resistance and gravimetric moisture content (ω), respectively. Sheng et al. (2011) developed a combined penetrometer sensor kit consisting of two impedance soil moisture sensors, two soil temperature sensors, and an electrical conductivity (EC) sensor to monitor moisture dynamics in the soil. All these combined-penetrometer sensor porotypes (CPSP) were not implemented to predict BD. Quraishi and Mouazen (2013a) introduced a multi-sensor platform, which enabled the assessment of BD from the fusion of data on penetration resistance measured with a load cell and ω, clay content and organic matter (OM) measured with a NIRS sensor. This multi-sensor does not measure θv, necessary for the direct assessment of BD using the following equation (Wijaya et al., 2004): (1)
BD = θ v/ ω
Fig. 1. The combined-penetrometer sensor prototype (CPSP) of a near infrared spectrophotometer (NIRS) and a frequency domain reflectometry (FDR) sensor, for the measurement of soil gravimetric moisture content (ω) and volumetric moisture content (θv), respectively, and assessment of bulk density (BD).
−3
where BD is the soil bulk density in g cm , θv is the volumetric moisture content in cm3 cm−3 and ω is the gravimetric moisture content in g g−1. Therefore, there is a need for a modified penetrometer sensing kit that enables simultaneous measurement of both θv and ω, and then derive BD values using Eq. (1). The aim of this paper was to design and evaluate a CPSP, consisting of a NIRS and a FDR sensor. The developed CPSP kit will be tested for the measurement of the top soil BD under field measurement conditions. 2. Materials and methods 2.1. Development of the combined-penetrometer sensor prototype (CPSP) Al-Asadi and Mouazen (2014) introduced a proof of concept of a new measurement system of BD, based on measurement of θv and ω values with a FDR and NIRS, respectively, which are then substituted into Eq. (1) to assess BD. In a recent work, Mouazen and Al-Asadi (2018) studied the effect of moisture content on the accuracy of BD measurement. These two studies enabled understanding the requirements for designing a new CPSP for the measurement of BD based on a standard penetrometer. The new design to be reported in this work consists of a NIRS, a dielectric (FDA) sensor, a standard penetrometer, a global positioning system (GPS) a battery and a laptop. The NIRS used was Avantes® portable model NIR200-2.6 (Avantes, Eerbeek, the Netherlands), which has a dual stage thermo-electrical Peltier-cooled InGaAs single detector with 256 pixels, with a spectral range of 1650–2500 nm and 7 nm resolution. This spectrometer was connected to a laptop through a high-speed USB2.0 interface and AvaSoft 7.7 software (Avantes, Eerbeek, the Netherlands). The spectrometer was also connected to an assembly of a rod and a 30 degree, 1.26 cm2 base-area cone through two optical fibres (Quraishi and Mouazen, 2013a). An electronically stabilized 20 W halogen lamp is used as a light source, and light was transferred to the soil profile through an illumination fibre connected at one end to the spectrometer and at the other end to a sapphire window (Fig. 1). The diffuse reflected light passing the sapphire window was collected back to the spectrometer by a detecting fibre. The system operates in situ using 24VDC and 20 Watts lead-acid battery as a power source. A 50 channel global positioning system (GPS) (eTrex 60C model, Garmin, USA) was used to record the sampling location. The measurement of θv was done by a new designed FDA sensor, which consists of an electronic circuit generating a 100 MHz electromagnetic sine wave, propagated through the soil body by the central electrode in the form of a copper ring with a 10, 15 and 1.5 mm height, diameter and wall thickness, respectively. The copper ring is insulated from the probe body, which forms two shielding electrodes as they are connected to the electronic circuits’ negative. Each shielding electrode has a cylinder shape with a 13 and 50 mm diameter and height, respectively (Fig. 2). The readout pin of the electronic circuit is connected
Fig. 2. Shows the electromagnetic fringe fields around the dielectric sensors’ electrodes.
to the HH2 meter (Delta-T devices, Cambridge, UK), which at the time of reading acts as a power supply and provides data storage. The probe body also provides protection for the optical fibres, as they run inside its cavity, as described above. As the combined probe inserted into the soil vertically, the surrounding soil in contact with probe electrodes will be affected by the fringe fields of the propagated signal, resulting from the two capacitors (Fig. 2).
2.2. Experimental procedure The CPSP portable measurement system was tested in situ in seven fields with various textures and growing crops (Table 1). The average field particle size distribution (PSD) was measured, using the sieving and sedimentation method (British Standards, 1998). The average soil Table 1 Information of the seven fields used for testing the combined-penetrometer sensor prototype (CPSP). Fields
Nr.
Soil texture
Clay%
Silt%
Sand%
OM%
Crop
Avenue 1 Beechwood Clover hill Orchard Showground
20 20 20 20 20
Sandy loam Cay Clay loam Clay loam Sandy clay loam
17 66 35 33 24
20 11 24 26 17
63 23 41 41 59
3.6 5.8 4.8 4.15 3.34
Barley Wheat Barley Wheat Barley
Avenue 2 Onley
50 50
Sandy loam Clay
29 60
20 30
51 10
2.98 5.4
Grass
OM is organic matter 486
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
Table 2 Sample statistics of the soil samples, used for testing the new combined-penetrometer sensor prototype (CPSP). Values were obtained from laboratory measured volumetric moisture content (θv), in cm3 cm−3, gravimetric moisture content (ω) in g g−1 and bulk density (BD) in g cm−3. Statistics
Maximum Minimum Range Average SD
Arable fields
Grass fields
θv
ω
BD
θv
ω
BD
0.55 0.12 0.43 0.34 0.14
0.41 0.10 0.31 0.26 0.11
1.60 1.08 0.52 1.34 0.11
0.46 0.24 0.22 0.35 0.09
0.36 0.15 0.21 0.25 0.08
1.78 1.20 0.58 1.49 0.24
SD is standard deviation.
accuracy can be achieved (Mouazen et al., 2005). After noise elimination, more noise reduction was also achieved by means of averaging every 10 nm wavelengths to one. A maximum normalisation was followed, which is typically used to get all data to approximately the same scale, or to get a more even distribution of the variances and the average values. Spectra were then subjected to the Savitzky–Golay first derivation (Martens and Naes, 1989), which enabled the transformation of spectra data to the first or higher order derivatives, including a smoothing factor. This method determines how many adjacent variables will be used to estimate the polynomial approximation used for derivatives. A second order polynomial approximation was selected due to its performance in producing more accurate calibration models. The final process of the pre-treatment was smoothing, which was carried out at a 2:2 rate, in order to reduce further the noise in spectra. All pretreatment steps were carried out using the Unscrambler 7.8 software (Camo Inc.; Oslo, Norway). Artificial neural network (ANN) was selected in this study for modelling and data fusion of both sensors’ data, and analyses were performed using Statistica software (StatSoft, USA, 2011). Neural networks are simplified models of the biological structure of human brains (Günaydin, 2009). There are three main layers in the ANN structure, namely, a set of input nodes, one or more layers of hidden nodes and a set of output nodes. In this study, different number of nodes was used in each layer depending on the input data used. The powerful second order Broyden–Fletcher–Goldfarb–Shanno (BFGS) training algorithms were used, with different transfer functions assigned for hidden and output layers, as detailed in Table 3. The number of neurons in the hidden layer is established by training several networks with different number of hidden neurons, and comparing the predicted with measured values. Since a previous study by Al-Asadi and Mouazen (2014) confirmed that data fusion provided the best predictions results for the measurement of θv, ω and BD, this approach was adopted in this study, so that the input data for each ANN were V and NIR spectra and the output were either θv or ω. For modelling of both, the data of the five arable fields (100 samples) and the two grass fields (100 samples), were each divided into training and cross-validation set (75 samples), and validation set (25 samples). The training samples were used to generate two separate ANN calibration models: one for the arable fields and the other one for the grass fields. Additional model named collective calibration model was generated by using 75% of the 200 arable and grass fields’ samples (e.g., 150 samples) for training and the remaining 25% (e.g., 50 samples) were the validation set. Accuracy of calibration models was evaluated using the independent validation sets, and results are reported for the independent validation sets only. By substituting the predicted θv and ω obtained from the ANN analysis into Eqn (1), BD was assessed and compared to the measured values obtained from the oven drying method of soil samples collected at the same positions of the CPSP readings in the field. Model prediction performance was evaluated by means of coefficient of determination
Fig. 3. The combined-penetrometer sensor prototype (CPSP) for the field measurement of soil volumetric moisture content (θv) and gravimetric moisture content (ω) and assessment of bulk density (BD), assembled on a sack barrow.
OM was measured with a TrusSpecCNS spectrometer (LECO Corporation, St. Joseph, MI, USA), using the Dumas combustion method (British Standards, 2000). Table 1 shows the results of laboratory analyses for OM, PSD and soil texture class, classified according to the United States Department of Agriculture (USDA) classification system (Soil Survey Staff, 1999). The light weight and compact size of the CPSP made it easy to move it through the growing crops and carry out the measurements by assembling the system on a sack barrow (Fig. 3). The experiment ran from August 2013 to December 2013, at the Silsoe experimental farm of Cranfield University, Bedfordshire, UK. The penetrometer was used to collect data at 200 points selected randomly in the seven experimental sites. A total of 20 readings per field were collected in the five arable fields, whereas 50 readings per field were collected in the two grass fields (Table 1). Readings were taken after pushing the penetrometer vertically into the soil to 10 cm depth, and then NIR spectra and FDA V were collected together with the GPS readings. A total of 200 assessments of θv, ω and BD in the seven sites were first made using the CPSP, and these assessed values were then validated by comparing with measured values, obtained by drying Kopercki ring collected soil samples at 105 °C for 24 h (British Standards, 2007). Table 2 provides basic statistics of oven drying measured θv, ω and BD of the 200 soil samples, collected during the field testing of the portable CPSP. 2.3. Modelling Before modelling, the NIR spectra has to be pre-treated in order to reduce spurious peaks that do not contain any physical or chemical spectra information and to correct the scatter effects. The same pretreatments were applied on all spectra. The spectra range was first reduced to 1650–2250 nm, to eliminate noise at both edges of the detected wavelength range, so that enhancement of the calibration model 487
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
Table 3 Artificial neural network (ANN) modelling results of the field testing of the combined-penetrometer sensor prototype (CPSP). Index
Network structure
Training R2
Validation R2
Training error
Validation error
Training algorithm
Hidden activation
Output activation
Arable fields Grass fields Collective
170-17-2 170-8-2 170-15-2
1.00 0.99 0.97
0.98 0.99 0.95
0.000008 0.000052 0.000111
0.000458 0.000065 0.000652
BFGS 147 BFGS 113 BFGS 178
Exponential Logistic Exponential
Exponential Exponential Exponential
BFGS: Broyden-Fletcher-Goldfarb-Shanno algorithm; Logistic: logistic sigmoid function; Exponential: negative exponential function.
grass fields), Table 4 shows excellent accuracy of ω prediction.
(R2), root mean square error of prediction (RMSEp) and ratio of prediction deviation (RPD), which was calculated as the ratio of standard deviation of laboratory measured values divided by RMSEp.
3.2. Evaluation of sensor performance for bulk density measurement The prediction of BD in the arable fields revealed lower accuracy results with values of R2 of 0.34 and RMSEp of 0.104 g cm−3 (Table 4), compared to R2 of 0.94 and RMSEp of 0.04 g cm−3 reported by Quraishi and Mouazen (2013a), who used a prototype measuring system that combined a NIRS sensor and a penetrometer. The grass fields’ results show a better accuracy in comparison with the measurements from arable fields’ model, with the validation R2, RMSEp and RPD values of 0.47, 0.077 g cm−3 and 1.36, respectively. The collective model in predicting soil BD revealed comparable results to that of the grass fields’ model with R2 = 0.52 and RMSEp = 0.102 g cm−3. Comparing the scatter plots revealed that BD points are the closest to the 1:1 line for the collective model (Fig. 6C), while BD points of the arable fields’ model are most scattered (Fig. 6A). The smaller intercept of 0.2 g cm−3 and the better slope of 0.84 of the collective model (Fig. 6C), as compared to the other two models indicate better assessment performance, suggesting the use of a large number of soil samples collected from arable and grass soils has improved the prediction accuracy of BD. Literature shows no similar studies about the assessment of BD, as a function of θv and ω measured with a dielectric probe and NIRS, respectively. Therefore, the CPSP introduced in the current study proves to be unique in the assessment of BD, and also the potential the sensor has in the prediction of other soil properties e.g., with NIRS that are important for land management. However, further development is needed so as to improve the assessment accuracy obtained so far. This might concern improvement in the technical specification or calibration procedure followed in the present study.
3. Results 3.1. Evaluation of sensor performance for moisture measurement The relation between the FDA sensor V and θv was somewhat weaker when the measurements were taken in the arable fields compared to those taken in grass fields. However, this is likely to be due to the heterogeneous nature of topsoil layer of arable soils due to repeated tillage. While the R2, RMSEp, and RPD values for arable soils may be lower than grass soils, but they are both considered to be within the excellent prediction level (Table 4), with values of 0.97, 0.024 cm3 cm−3 and 5.80, 1.00, 0.005 cm3 cm−3 and 13.72, and 0.94, 0.039 cm3 cm−3 and 3.67 for arable, grass and the collective models, respectively. Similarly, the best slope and smallest x intercept of linear line indicate predictions in grass fields are in advance of other two models (Fig. 4). Kaleita et al. (2005) reported less successful field calibration of a dielectric sensor for the measurement of θv with a R2 value of 0.77. Similarly, θv result obtained in the present work is more accurate than those reported by Andrade-Sanchez et al. (2001), with R2 of 0.78. The performance of the NIRS for the prediction of ω was rather different compared to the dielectric probe performance for the measurement of θv. Smaller differences in model performance between three models for ω prediction can be observed in Table 4 and Fig. 5. Values of R2, RMSEp and RPD for the prediction of ω are 0.97, 0.019 g g−1 and 5.46, and 0.96, 0.011 g g−1 and 4.73, for arable and grass fields’ soils, respectively, and 0.96, 0.023 g g−1 and 4.56 for the collective model, respectively, which are of similar magnitude to those obtained by Mouazen et al. (2006) and closer to those reported by Quraishi and Mouazen (2013b), who found 0.95, 2.39% and 4.15 values, respectively. This is despite the fact that in the present study spectra were collected from multiple fields (e.g., five arable and two
4. Discussion 4.1. Evaluation of the new combined-penetrometer sensor prototype (CPSP) According to Hemmat and Adamchuk (2008), the current measuring system can be categorized under water content sensors that measures ‘indirectly’ soil compaction. Mosaddeghi et al. (2007) stated that we should not only rely on the strain-related properties as the dependent variables for the assessment of soil compaction, whereas different soil properties can also be considered as a sign of compaction. The CPSP kit introduced in this work showed encouraging accuracy for measuring soil BD under field conditions (Table 4). A NIRS with a short spectral range of 1650–2500 nm was used for the CPSP. The process of combining NIRS and FDR sensors was the most challenging task of this work, as both sensors are very sensitive to any changes made to their sensing probes or heads. Delicate materials have to be used in the sensor manufacture, for example, using thin fibre optics in the limited space inside the steel shaft, upon which the combined probe is assembled and a sapphire material was used in a protective window design (Fig. 1) to ensure that no scratches would affect the probe windows’ transparency. The accuracy of the dielectric probe depends on maintaining good contact between the probe’s electrodes and the soil. For this reason, a penetrating cone is used as one of the shielding electrodes in addition to a second part above the central electrode (Fig. 1). A similar approach
Table 4 Prediction results of soil volumetric moisture content (θv), gravimetric moisture content (ω) and bulk density (BD), obtained by using the combinedpenetrometer sensor prototype (CPSP) in arable fields, grass fields and collective (e.g., both arable and grass fields). Arable fields
Grass fields
Collective
θv R2 RMSEp, cm3 cm−3 RPD
0.97 0.024 5.80
1.00 0.005 13.72
0.94 0.039 3.67
ω R2 RMSEp, g g−1 RPD
0.97 0.019 5.46
0.96 0.011 4.73
0.96 0.023 4.56
BD R2 RMSEp, g cm−3 RPD
0.34 0.104 1.08
0.47 0.077 1.36
0.52 0.102 1.21
R2: coefficient of determination. RMSEp: root mean square error of prediction; and, RPD: residual prediction deviation. 488
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
0.6
0.6
Estimated θv (cm3 cm-3)
A
B
0.4
0.4
0.2
0.2
y = 0.96x + 0.016 R² = 0.97
y = 1.001x - 0.003 R² = 0.995
0.0
0.0 0.0
0.2
0.4
0.6
0.0
0.2
0.4
0.6
0.6
C 0.4
0.2
y = 1.07x - 0.02 R² = 0.94 0.0 0.0
0.2
0.4 3
0.6 -3
Measured șv (cm cm ) Fig. 4. Scatter plots of core sampling measured versus artificial neural networks (AAN) predicted soil volumetric moisture content (θv) using the combinedpenetrometer sensor prototype (CPSP) for field measurements in arable fields (A), grass fields (B) and the collective fields (C).
measuring system has demonstrated that moderate accuracy of BD measurement can be achieved for a wide range of soil types and agricultural practices with less input data needed (e.g., θv and ω only), compared to others reviewed works. For example, Quraishi and Mouazen (2013a) used a similar prototype consisting of a penetrometer and a NIRS for measurement of topsoil BD, reporting RMSEp of 0.04 g cm−3, which is indeed smaller than the RMSEp achieved in this study (0.077 – 0.104 g cm−3). However, large number of variables was needed as input data for the ANN analysis to predict BD by Quraishi and Mouazen (2013a), which included ω, OM, penetration resistance and clay content.
was followed by Peter and Yurui (2004) for the design of a combined capacitance sensor with a cone penetrometer. The Peter and Yurui’ system overestimated penetration resistance compared to the readings of a standard cone penetrometer. They concluded that adding the capacitance sensor electrode and insulator caused additional frictional resistance to soil penetration by the combined sensor, which subsequently led to deterioration in measurement accuracy for the penetration resistance. Such a problem has no effect on the soil moisture related systems to measure soil BD designed and developed in the present study. This is in-line with the literature review, which showed shortcoming of soil strength related measuring systems. The NIRS of the CPSP predicted ω accurately in arable and grassland soils with R2 of 0.97 and 0.96 and RMSEp 0.019 and 0.011 g g−1, respectively (Table 4), compared to a lower accuracy magnitude reported by Hummel et al. (2004), who presented a combined probe consisting of a cone penetrometer and NIRS. They achieved a R2 of 0.90 and the standard error of calibration and prediction of soil ω were 1.97% and 2.38%, respectively. Their combined system has predicted the cone index with R2 = 0.86 compared to the standard cone index, using all the data from clay loam, silt loam and sandy loam soil textures and the entire range of soil moisture content. However, they did not attempt to measure soil BD. In general, the CPSP developed and tested in the present work showed moderate accuracy to estimate BD in the top soil, under various crops, soil types and conditions. The growing crop effect showed that grass fields’ soils are in advance of arable fields soils planted with different crops, but the authors believe that there are no direct connections between the accuracy and the growing crops. Literature indicated that many authors have attempted to provide an accurate soil BD measuring system, due to the importance of such a single soil physical property that has many environmental and economic impacts. The new
4.2. Advantages and practical implementation challenges The advantages of using the new CPSP to measure soil compaction indicated as BD with the implementation of FDR and NIRS sensors are as followed:
• The system is semi- non-invasive, where soil penetration should take
• • 489
place in the vertical direction while collecting data. The FDR sensor measures the dielectric constants of the compound by emitting an electromagnetic signal propagated through the soil body, whereas the NIRS sensor collect the diffused reflectance of electromagnetic wave of light from the soil samples surface down to the depth of 2 mm. The system is relatively small in size and light in weight and can be mounted on a wheelbarrow. It is of a robust design in terms of the penetration rod geometrical structure, which enables smooth penetration of the soil, without blocking of the sapphire window from which NIR spectra are collected. It has an additional advantage by presenting a combined probe
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
0.5
0.5
A
0.4 0.3
0.3
0.2
0.2
Estimated ω (g g-1)
0.1
B
0.4
0.1
y = 0.98x + 0.0034 R² = 0.97
y = 1.03x - 0.01 R² = 0.96
0.0
0.0 0.0
0.1
0.2
0.3
0.4
0.0
0.5
0.1
0.2
0.3
0.4
0.5
0.5
C
0.4 0.3 0.2 0.1
y = 1.1x - 0.0154 R² = 0.96
0.0 0.0
0.1
0.2
0.3
0.4
0.5
-1
Measured Ȧ (g g ) Fig. 5. Scatter plots of core sampling measured versus artificial neural networks (AAN) predicted soil gravimetric moisture content (ω) using the combined-penetrometer sensor prototype (CPSP) for field measurements in arable fields (A), grass fields (B) and the collective fields (C).
containing both sensors, with which readouts can be recorded within few seconds. It is cost effective, efficient and long lasting system for measuring soil BD and has the potential to measure other soil physical and chemical properties using the NIRS. Adequate accuracy for BD assessment with a wide range of soil types and conditions can be obtained, providing accurate calibration models are developed in advance.
main challenge in the manufacturing process of the CPSP, which was particularly due to the difficulty of dealing with the high frequency signals (100 MHz). With more research time and financial support a whole profile measuring probe instead of just the top 10 cm of depth measurements would have been developed. This will be considered for future work.
Despite the above advantageous of the CPSP, few challenges can be discussed. One of the challenges raised during the field measurements was the fact that agricultural soils are not naturally homogeneous and contain stones, gravels and plant residues. These have a significant negative impact on the FDR sensor, relating to the fact that the central electrode is very sensitive to air pockets when they present around it. Stones in the collected soil cores lead to soil moisture estimation error, as the stones do not hold any moisture. Furthermore, dry soil conditions make the measurements difficult and less accurate (Mouazen and AlAsadi, 2018), while the very moist soil status could make it susceptible to block the sapphire window, particularly in clay soils. Technical challenges can be summarised as the necessity for a more compact system, which has the capability of recording the data from both sensors, as well as to process the adapted calibration models and providing the desired output in real time, which may include spatial maps or prediction values. Long life and light weight batteries are essential for the measuring system to be easily mobile and reliable for extended periods of field measurement. However, assembling the system on a quad motorbike could solve most of the above technical issues. Supplying the right coaxial cable for the dielectric sensor was the
A portable combined-penetrometer sensor prototype (CPSP) for the measurements of volumetric moisture content (θv), gravimetric moisture content (ω) and bulk density (BD) in the topsoil was developed and successfully tested under field conditions across a wide range of soil textures and land uses. Results showed the prediction of θv and ω provided excellent results, whereas the BD assessment results were encouraging. The best topsoil BD prediction accuracy was for grass fields’ soils, as the smallest root mean square error of prediction (RMSEp) of 0.077 g cm−3 was recorded. A collective BD model produced for arable and grass fields’ soils provided a moderate accuracy with a RMSEp of 0.102 g cm−3. A narrow wavelength band spectrometer (e.g., 1650–2500 nm) or even mono optical detectors of a certain wavelength associated with OH absorption overtones at 1450 or 1950 nm can also be used to predict ω effectively, which would considerably lower the cost of the new CPSP and would present it as replacement for the high cost traditional laboratory measuring methods.
• •
5. Conclusions
Acknowledgments Authors acknowledge the financial support of the UK Engineering 490
Computers and Electronics in Agriculture 151 (2018) 485–491
R.A. Al-Asadi, A.M. Mouazen
2.0
2.0
A
Estimated BD (g cm-3)
1.8
B
1.8
1.6
1.6
1.4
1.4
y = 0.66x + 0.48 R² = 0.344
1.2
1.2
1.0
y = 0.67x + 0.48 R² = 0.47
1.0 1.0
1.2
1.4
1.6
1.8
2.0
1.0
1.2
1.4
1.6
1.8
2.0
2.0
C
1.8 1.6 1.4 1.2
y = 0.84x + 0.2 R² = 0.52
1.0 1.0
1.2
1.4
1.6
1.8
2.0
-3
Measured BD (g cm ) Fig. 6. Scatter plots of core sampling measured versus artificial neural networks (AAN) predicted soil bulk density (BD) using the combined-penetrometer sensor prototype (CPSP) for field measurements in arable fields (A), grass fields (B) and the collective fields (C).
and Physical Science Research Council (EPSRC) and The Douglas Bomford Trust. The corresponding author acknowledges the financial support of the Flemish Scientific Research (FWO) for funding of SiTeMan Odysseus I Project (Nr. G0F9216N).
607–618. Kaleita, A.L., Heitman, J.L., Logsdon, S.D., 2005. Field calibration of the ThetaProbe for der FOR Des Moines lobe soils. Am. Soc. Agric. Eng., Appl. Eng. Agric. 21 (5), 865–870. Martens, H., Naes, T., 1989. Multivariate Calibration, second ed. John Wiley & Sons Ltd, Chichester. Mosaddeghi, M.R., Koolen, A.J., Hajabbasi, M.A., Hemmat, A., Keller, T., 2007. Suitability of pre-compression stress as the real critical stress of unsaturated agricultural soils. Biosyst. Eng. 98, 90–101. Mouazen, A.M., De Baerdemaeker, J., Ramon, H., 2005. Towards development of on-line soil moisture content sensor using a fibre-type NIR spectrophotometer. Soil Tillage Res. 80, 171–183. Mouazen, A.M., Ramon, H., 2006. Development of on-line measurement system of bulk density based on on-line measured draught, depth and soil moisture content. Soil Tillage Res. 86, 218–229. Mouazen, A.M., Karoui, R., De Baerdemaeker, J., Ramon, H., 2006. Characterization of soil water content using measured visible and near infrared spectra. Soil Sci. Soc. Am. J. 70, 1295–1302. Mouazen, A.M., Al-Asadi, R.A., 2018. Influence of soil moisture content on assessment of bulk density with combined frequency domain reflectometry and visible and near infrared spectroscopy under semi field conditions. Soil Tillage Res. 176, 95–103. Peter, S.L., Yurui, S., 2004. Combined sensor for simultaneous investigation of cone index and soil water content. ASAE/CSAE 041046. Quraishi, Z., Mouazen, A.M., 2013a. A prototype sensor for the assessment of soil bulk density. Soil Tillage Res. 134, 97–110. Quraishi, M.Z., Mouazen, A.M., 2013b. Development of a methodology for in situ assessment of topsoil dry bulk density. Soil Tillage Res. 126, 229–237. Sheng, W., Sun, Y., Schulze Lammers, P., Schumann, H., Berg, A., Shi, C., Wang, C., 2011. Observing soil water dynamics under two field conditions by a novel sensor system. J. Hydrol. 409 (2011), 555–560. Soil Survey Staff, 1999. Soil Taxonomy – a basic system of soil classification for making and interpreting soil surveys; second edition. Agricultural Handbook 436; Natural Resources Conservation Service, USDA. Washington DC, USA. Vaz, C.M.P., Bassoi, L.H., Hopmans, J.W., 2001. Contribution of water and bulk density of field soil penetration resistance as measurement by a combined cone penetrometerTDR probe. Soil Tillage Res. 60, 35–42. Wijaya, K., Nishimura, Y., Kato, M., Nakagawa, M., 2004. Field estimation of soil dry bulk density using amplitude domain reflectometry data. J. Jpn. Soc. Soil Phys. 97, 3–12.
References Al-Asadi, R.A., Mouazen, A.M., 2014. Combining frequency domain reflectometry and visible and near infrared spectroscopy for measurement of soil bulk density. Soil Tillage Res. 135, 60–70. Andrade-Sanchez, P., Aguera, J., Upadhyaya, S.K., Jenkins, B.M., Rosa, U.A., Josiah, M., 2001. Evaluation of a dielectric based moisture and salinity sensor for in-situ applications. Paper No. 01-1010, ASAE, St. Joseph, Michigan. Aragón, A., García, M.G., Filgueira, R.R., Pachepsky, Y.A., 2000. Maximum compactability of Argentine soils from the Proctor test; the relationship with organic carbon and water content. Soil Tillage Res. 56, 197–204. Batey, T., 2009. Soil compaction and soil management – a review. Soil Use Land Manage. 25 (4), 335–345. British Standards, 1998. Soil Quality: BS 7755: Section 5.4: 1998. Part 5: Physical Methods. Section 5.4: Determination of Particle Size Distribution in Mineral Soil Material – Method by Sieving and Sedimentation. British Standards Institution, UK. British Standards, 2000. Soil Improvers and Growing Media: BS EN 13039:2000. Determination of Organic Matter Content and Ash. British Standards Institution, UK. British Standards, 2007. Soil Improvers and Growing Media: BS EN 13040:2007. Sample Preparation for Chemical and Physical Tests, Determination of Dry Matter Content, Moisture Content and Laboratory Compacted Bulk Density. British Standards Institution, UK. Günaydin, O., 2009. Estimation of soil compaction parameters by using statistical analyses and artificial neural networks. Environ. Geol. 57, 203–215. Hamza, M.A., Anderson, W.K., 2005. Soil compaction in cropping systems: a review of the nature, causes and possible solutions. Soil Tillage Res. 82 (2), 121–145. Hemmat, A., Adamchuk, V.I., 2008. Sensor systems for measuring soil compaction: review and analysis. Comput. Electron. Agric. 63, 89–103. Horn, R., van den Akker, J.J.H., Arvidsson, J., 2000. Subsoil Compaction: Distribution, Processes and Consequences. Deutsche Bibliothek, Germany. Hummel, J.W., Ahmad, I.S., Newman, S.C., Sudduth, K.A., Drummond, S.T., 2004. Simultaneous soil moisture and cone index measurement. Am. Soc. Agric. Eng. 47 (3),
491