Journal of Hydrology 530 (2015) 657–666

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage:

General calibration of TDR to assess the moisture of tropical soils
using artificial neural networks
Sidney Sara Zanetti ⇑, Roberto Avelino Cecílio, Vitor Heringer Silva, Estevão Giacomin Alves
Federal University of Espírito Santo, Forest and Wood Sciences Department, PO Box 16, 29.500-000 Alegre, ES, Brazil

a r t i c l e i n f o s u m m a r y

Article history: Determinations of soil moisture are important for various agricultural, environmental and hydrological
Received 23 June 2015 applications, and accurate assessments are required. Artificial neural networks (ANNs) were applied in
Received in revised form 17 September the present study to conduct general calibrations of time-domain reflectometry (TDR) probes using the
physical characteristics of soil to estimate the moisture. The ANNs were trained and tested using data
Accepted 12 October 2015
from five different soils. All of the combinations of physical properties, including the bulk density and
This manuscript was handled by sand, silt, clay and organic matter contents, were tested, and the inclusion of at least one of those network
Konstantine P. Georgakakos, Editor-in-Chief, variables along with the apparent dielectric constant (Ka), which was assessed using the TDR device, were
with the assistance of Emmanouil N. sufficient to calibrate all five of the soils simultaneously. The ANN selected for the general calibration has
Anagnostou, Associate Editor a hidden layer with 13 neurons and tan-sigmoid-type transfer function. The analysis of the statistical
indexes values indicates that the ANNs were slightly better than the third-order polynomial equations
Keywords: (Topp-like equations), which were specifically fitted to each soil. The tests were conducted to assess
Soil water content the performance of the general calibrations that were applied to estimate the moisture of the soils
Soil moisture excluded from the training process, although the ANNs have such a potential; the most representative
Neural networks variables in descending order of importance were as follows: organic matter, sand, clay, and bulk density.
Time domain reflectometry
The soil silt content failed to stand out in this analysis and showed a lower performance. Based on the
results, the organic matter content was the preferred variable for use along with the Ka in the ANNs
Tropical soils
applied to the general calibration of TDR (RMSE ranging from 0.0126 to 0.0237 g/g and r2 ranging from
0.9083 to 0.9891). The sand content was also considered an advantageous variable because it was more
easily assessed. The variables clay content and bulk density or the combination of several variables may
also be used when available. As found by previous studies, the TDR calibration using ANN were better to
sandy soils.
Ó 2015 Elsevier B.V. All rights reserved.

1. Introduction The use of the TDR method has expanded and provided a rele-
vant contribution for studies on the soil–water–plant–atmosphere
There are various methods for assessing soil water content, relationship, enabling assessments of the soil moisture (Calamita
including the gravimetric method, which is considered to be the et al., 2012; Graeff et al., 2010; Zehe et al., 2010) and its electrical
standard method, and methods that use soil physical properties conductivity (Moret-Fernández et al., 2012; Persson and Uvo,
related to moisture to indirectly estimate the soil water content, 2003; Skierucha et al., 2012), estimations of crop evapotranspira-
including temperature, neutron thermalisation, electrical resis- tion (Schelde et al., 2011; Ward and Dunin, 2001) and analyses
tance, capacitance, spectrometry and time-domain reflectometry of the movement of water and solutes in soils (Kulasekera et al.,
– TDR (Altendorf et al., 1999; Anderson and Croft, 2009; 2011; Souza and Folegatti, 2010; Thomsen et al., 2000; Timlin
Antonucci et al., 2011; Calamita et al., 2012; Francesca et al., and Pachepsky, 2002).
2010; Imhoff et al., 2007; Noborio, 2001; Souza and Matsura, The TDR method has advantages over other methods that
2002). include improved accuracy, lack of soil sample destruction, real-
time assessment of moisture and potential for continuity and
automation of data collection. However, a key issue when using
⇑ Corresponding author. Tel.: +55 28 3558 2527. the TDR method is the difficulty of developing calibration equa-
E-mail addresses: (S.S. Zanetti), tions for the different soil types of interest (Coelho et al., 2003;
(R.A. Cecílio), (V.H. Silva), Imhoff et al., 2007; Staub et al., 2010; Zehe et al., 2010).
(E.G. Alves).
0022-1694/Ó 2015 Elsevier B.V. All rights reserved.

estimate moisture without having to conduct specific calibration The calibration consisted of mathematically correlating the experiments for each type of soil (Arsoy et al. Alluvial soil (soil 4). physical and chemical three-litre volume) were used in the experiment. samples were collected in the field to characterise the studied rithms and methods (Arsoy et al.S. Tomasella et al. ibrate a TDR device to assess the moisture of different types of 3 h ¼ a þ bK a þ cK 2a þ dK a þ eðBulk dry densityÞ þ f Clay tropical soils found in Brazil.. 2. a promising research collected from each soil.. LM: Levenberg–Marquardt algorithm.. This behaviour resembles that of a typical temperate climate sandy soil. Various types of mathematical models may be used in an important role as binding agents of negatively charged clay TDR calibrations. Three samples were ANN’s architecture affects its performance. tion using ANNs since the pioneer Persson et al. Materials and methods Persson et al. Persson apparent dielectric constant (Ka). (2002) Ka. including the bulk density and clay and Based on the relevance of this topic and because no similar organic matter contents.. different 3). medium texture Latosol (soil 2). Guaçuí and applied ANNs to calibrate the TDR using other soil properties in Jerônimo Monteiro in the state of Espírito Santo. As an example. although Other researchers (Jacobsen and Schjønning. A TDR100 device with three-rod CS 610 probes (Campbell Sci- only very few studies have been conducted to provide TDR calibra. Table 2 outlines the descriptive characteristics of the studied soils. (2001) used artificial neural networks (ANNs) as a pedotransfer function (PTF) to model the relation between soil The present study consisted of fitting mathematical models moisture and Ka and produced in a better performance than that based on ANNs to calibrate a TDR device using five types of tropical of other tested equations.. . specific gravity. clay. 2007. mately 30 cm long.. 1997). Soil sampling and TDR measurements The soil density and clay content were the most frequently used input variables alongside Ka. Stangl et al. The bottom sur- differences between temperate and tropical soils might explain face of the containers had previously been perforated to enable Table 1 Studies found in the literature on TDR calibration using artificial neural networks. Nitosols and estimated when the apparent dielectric constant (Ka) of the system related soils). resulting in a general calibration for several types was also calibrated using the fitted third-order polynomial equa- of soils. organic matter 4-9-1 Sigmoid GD Denmark Namdar-Khojasteh et al. 2009). training epochs taken on training and etry (texture) and organic matter content were assessed in the lab- the ANN’s numbers of neurons and layers. et al. However.. Aracruz. and sandy soil (soil 5). tent. Zanetti et al. (2013) Ka.. GD: gradient descent algorithm (backpropagation).. including the water con- Despite the good performance of ANNs to predict soil moisture. Tomasella and Hodnett. Using ANNs associated with soil physical characteristics tions (Topp-like equations) for comparison purposes (Topp et al. the pre- of third-order polynomial equation (Eq. physical characteristics. Author Inputs Network architecture Transfer function Training algorithm Country Persson et al. Soil moisture (h) may be inadequate for tropical soils (especially Latosols. Kaiser et al. Minasny and Hartemink. (1)) proposed by Topp et al. (1980) is one of the most well. Once the conducted at depths ranging from 0 to 10 cm. Furthermore. creating stable micro-aggregates within the size range (Eq. 2011. Persson et al. 2013). (2001) paper. 2000. clay + silt 4-11-1 Tan-sigmoid LM Turkey Ka: apparent dielectric constant. 2014) strongly encourages treating these soils separately from Fifteen plastic containers of known mass that were approxi- the temperate climate ones. 2013. Brazil. probes. ical temperate clayey soils. Table 1 outlines the papers that have described this kind of studies. Two types of area consists in the evaluation and unification of training algo. GDM: GD with momentum. the sent study was conducted to fit the models based on ANNs to cal- improvement in fit was considered small. (2002) also soils collected in the municipalities of Alegre. Li to assess the bulk density according to the volumetric ring method.. UT. many methods reported by Brazilian Corporation of Agricultural researchers (van den Berg et al.. Research (EMBRAPA.658 S. 1997) have proposed calibrating the TDR by using the soil character of Latosols. 2002). structure of Latosols and Nitosols. of silt to fine sand. Nagare et al. bulk density. oratory using the disturbed samples according to the standard Studying the behaviour of tropical soils moisture. 1993. kaolinitic tropical soils (about known and used (Comegna et al.1. in addition to Ka. and the soil physical properties. enables the development of general calibrations that adequately 1980). however. clay 3-10-1 None informed GDM Iran Arsoy et al. The undisturbed methods were used Hodnett and Tomasella. Logan. 1997. bulk density. 15 cm wide and 7 cm deep (approximately According to Tomasella and Hodnett (2004). Babangida et al. Soil sampling was training algorithms and activation functions were used. The TDR addition to Ka. 2002. The pronounced differences between temperate has been established based on calibration equations (pedotransfer and tropical soils are usually explained by the micro-aggregated function) developed in the laboratory or in the field (Jones et al. The third-order polynomial regression equation minerals. entific. The granulom- to data standardisation. þ gðOrganic matterÞ ð2Þ 2. (2001) Ka 1-8-1 Sigmoid GD Sweden Persson et al. Silva and the water contents are comparatively higher because of the clayey Kay. where Fe and Al oxides play 2002). deriving specific tropical soil PTFs. especially concerning to soil moisture behaviour: low available water capacity and almost 80% of the 3 h ¼ a þ bK a þ cK 2a þ dK a ð1Þ plant available water between 10 and 100 kPa.. USA) was used. Subsequently. as well as the procedure soils: disturbed samples and undisturbed samples. 2013. (2)). all of the studies used The following five soil types were used: a clayey red-yellow multilayer perceptron (MLP) feedforward networks with only one Latosol (soil 1). 2004. 60% in clay) show ‘‘unusual” properties when compared with typ- Loiskandl et al.. 2011. 2010. which was measured using TDR et al. (2010) Ka.. representing three replicates. through an adaptation study has been found in the literature using tropical soils. 2010. / Journal of Hydrology 530 (2015) 657–666 The TDR method is based on the effect of moisture on the why PTFs derived for soils of temperate climate appeared to be microwave propagation velocity in soils. bulk density. Cambisol (soil hidden layer containing from 8 to 11 neurons.

7 1.S. i¼1 tained with the upper surface exposed so that water loss from where y is the neuron output value. (4)) and prevent loss of soil. vided into two subsets: a training subset (85% of data) and valida- The TDR device was calibrated by training the ANNs and fitting tion subset (15% of data). 2 The soil samples were broken down and packed in the plastic f ðxÞ ¼ 1 ð5Þ 1 þ e2x containers.441 0. The disturbed soil samples were used because the type function f(x) = x. which are shown in Eq. the containers were main. tan-sigmoid (Eq. 1999) .029 3  44 = 132 5 0. 1. by Eq.79 4 46.62 (g/g). and periodically measured after each resting period.321 0.91 1.0 13. xmax  xmin dard oven (gravimetric) method. the neuron shown in Fig. The mass of soil used to fill each container was the mass required to reach a bulk density near the value found in The network output layer was configured with one neuron and the field. The samples were left stand- the synaptic weight. respectively. Model of the neuron used in the artificial neural networks.021 3  54 = 162 4 0. 1999) was used as the criterion to terminate the training. with two replications (520 observations) used which enables us to identify the beginning of overfitting through for the model fitting/training (2/3 of the data) and one (260 obser. (6): The containers were partly immersed in water trays for a period ! X n of 48 h to saturate the soil samples from the bottom up through the y¼f wi xi þ b ð6Þ container holes. ues of the data. xi is the neuron input and output data and b is ing for several hours inside the laboratory to evaporate part of the the bias. on the soil moisture for a given value of Ka and has no effect on Mathematically. and xmin and xmax are the minimum and maximum val- based on the measured wet and dry masses.0 0. Zanetti et al. The first subset was used to train the the third-order polynomial equation parameters by associating the network. Artificial neural networks Feedforward multi-layer perceptron (MLP) ANNs with only one drainage of the excess water and closed using a fine mesh screen to hidden layer of neurons were used.017 3  44 = 132 Fig. (3): Table 3 Maximum and minimum soil moisture and number of observations from the moisture curves assessed using three replicates of the five soils studied. This proce- The data were also linearly standardised between 1 and +1 using dure was conducted daily until the soil reached the moisture at Eq. which is shown in Fig. 1 cal and longitudinal medium section of each container to prevent it f ðxÞ ¼ ð4Þ 1 þ ex from moving in the soil during handling. All of the neurons were configured based on the of sampling (disturbed or undisturbed) has no significant effect model introduced by Haykin (1999).59 1. / Journal of Hydrology 530 (2015) 657–666 659 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Table 2 Pn 2 Descriptive characteristics of the five soils studied.2. The data were split (Haykin.0 1. tion to conduct the cross-validation. Following saturation.7 42. The network training was con- ical parameters. hygroscopic equilibrium with the atmospheric air.3 1. 1 may be described the TDR calibration according to Santos et al. water.0 10. The dry mass of soil within the containers was assessed after drying. respectively. The resulting calibrations were evaluated using the soil mois- ture assessed by the standard gravimetric method as reference.54 the soil (g/g) and Ei is the soil moisture estimated in the calibration 2 56. Thus. (2010). Oi is the moisture found in 1 35. i¼1 ðOi  Ei Þ RMSE ¼ ð3Þ n Soil Sand (%) Silt (%) Clay (%) Bulk density Organic matter (%) (g/cm3) where n is the number of observations. The early-stopping training method using the three replicates of each soil sample. 1.0 29. and the sigmoid (Eq.021 3  64 = 192 2 0. Source: adapted from (Haykin.0 2. (5)) transfer functions were tested. effects of their order on the generalisation ability of the networks.55 2.   A sample of soil within each container was removed after the xo  xmin xnorm ¼ 2 1 ð7Þ last measurement to assess the residual moisture using the stan. which were derived from the subset of fitting/.32 2. S. Soil Maximum moisture Minimum moisture Observation number (g/g) (g/g) 1 0.41 1. the training data were subdi- vations) used for the performance/validation test (1/3 of the data). the soil water content where xnorm and xo are the standardised and observed values. coefficient of determination (r2) of the simple linear regression and root mean square error (RMSE). and the training section was interrupted after each itera- soil moisture values with their respective Ka values and soil phys. (3)) of the validation subset was training data. The mass of the container-soil-probe sets and Ka values The training data were previously randomised to avoid negative were assessed using a precision scale and TDR probes.7 58. f(x) is the transfer function.0 1.0 7. 3 47. (7).0 41. A TDR probe of known mass was horizontally fixed to the verti. tinued when the RMSE (Eq.3 36.40 2. wi is the soil would occur by evaporation. For that purpose.301 0.313 0.3 6. Table 3 outlines the maximum and minimum soil moisture and The ANNs were trained under supervision using the algorithm the number of observations from the moisture curves assessed Levenberg–Marquardt.227 0.02 5 69. cross-validation.036 3  54 = 162 3 0. into two subsets. at the time of each measurement was assessed by subtraction respectively.

clay and organic matter contents in the soil. Root mean square error (RMSE. nation using the best results. The general network trained with only Ka showed an r2 value that 3. totalling 38 different options. all of the RMSE val. collinearity between them. which is the value with the error closest to the global minimum because that procedure tends to minimise the effect of random initialisation of weights and biases of networks.. The silt content showed the lowest . (5)) and the value of k equal to 9 performances between them. ously with similar results. The value 100 was arbitrarily Fig. The networks trained using only Ka as 2010. what leads this research to training each ANN 100 times. which had similar sigmoid transfer function (Eq. at least one variable in the network along with Ka was sufficient the neural networks were trained using four types of soil. Arsoy et al.S. the value of k equal to 13 with a tan-sigmoid transfer along with the Ka value. the training was terminated because the network was essentially assimilating noise (experimental random error) contained in the training data. Furthermore. As it can be seen. 8 and 11 neurons in the hidden layer and concluded that the great- acteristic on the improvement of the network fit. all 31 possible combinations of those vari. (RMSE). Namdar-Khojasteh et al. otherwise. input variable (network 1-13-1) and using Ka along with a second ues from k values equal to 6 in the present study (Fig. networks were trained and tested again for each input data combi. So. clay with a sigmoid function (Eq. The inclusion of at least one network variable along with Ka pro- A single optimal k value may not exist or may be difficult to assess duced improved results in the analysis of the general values of with accuracy because different k values are found in each com. and the six input variables were as follows: Ka. The value of k equal to 13 with a tan.. 2013.660 S. with different from the training process. each value shown for each soil in results found for each training procedure. these values are shown Fig. 2) stabilised variable (network 2-13-1). and so forth.9727 or the potential application of general calibrations to other soil types 0. silt. Pearson linear correlation coefficients (r) between ing. training the the table was generated using a network trained (100 times) with same network several times may generate the most consistent the data from the other four soils. / Journal of Hydrology 530 (2015) 657–666 smaller than its value in the previous iterations. This procedure was conducted five times. 2. All of the combinations of input data mance. assess the performance of general calibrations conducted with The performance of ANNs’ training is affected by the random the ANNs when applied to estimate the moisture of soils excluded initialisation of free parameters (weights and bias). with different k values and transfer functions by using the grouped Table 5 shows the results from the cross-validation used to test data from the five studied soils. The integer values of k were tested. This study aimed to identify general RMSE value of 0. silt. results. 2002) selected k values ranging from 9 to the input variable (1-k-1) and Ka along with one of the variables (2. the bulk density and sand. which causes a loss in its generalisation ability according to Haykin (1999). (4)) provided the smallest errors and organic matter contents were used as the input for the ANNs. As previously mentioned. This study also The general calibrations developed using the ANNs showed aimed to assess the generalisation ability of the networks that similar performances to those found with the specific calibrations were applied to estimate the moisture of soils that were not used (third-order polynomial equations). thus. which is highlighted in bold. the inclusion of in the training process by conducting a new cross-validation. so its The smallest values of RMSE were found in the network 6-k-1 use is not indicated. every two. Results and discussion was 18% lower and RMSE value that was 165% higher than the val- ues obtained with the fitted third-order polynomial equation. and they ranged from 2 to 20 neurons for each trans- fer function (sigmoid and tan-sigmoid). and a different except for the network that was only trained with Ka showed a soil was used for testing each time. Therefore. One hundred works despite the uncertainties mentioned above.9728 when calculated using the grouped test data from the five by using the results of cross-validation.0160 g/g. 11. sand content. clay content and bulk density. est number of neurons tested provided the smallest RMSE in train- Furthermore. Each network was trained and tested 100 times to obtain a network with the best possible fit for each k value. Zanetti et al. 2. The slightly below 0. RMSE and r2 generated in the test of ANNs using only Ka as the puter experiment conducted. function was adopted in the present study in all of the tested net- ables were tested one by one. Persson et al. nomial equations and ANNs using Ka with 31 combinations of The networks were tested to establish a general calibration that input variables trained for all of the soils simultaneously to would represent the five soils simultaneously by assessing the test develop a general model of TDR calibrations. samples from the same soils used in the training. and the to develop a general calibration for all five of the soils simultane- test sample form the fifth soil was used for testing the perfor. which consequently resulted in increased accuracy in the TDR the variables were calculated to identify the possible existence of calibration. 2. Thus. equal to 6 have the potential to reproduce those results because the general RMSE value matches the values shown in Fig. bulk density and the sand. The best input variables for the ANNs Table 4 shows the calibration results using the third-order poly- were assessed based on the tests and correlation coefficients. g/g) of the artificial neural networks with chosen.. the networks tested with k values organic matter content promoted an improved performance that equal to or above 6 have a relatively similar potential. regardless of had smaller RMSE and greater r2 values and was followed by the the activation function used. For this purpose. which would obviate the soils. (2013) tested the performance of networks with k-1) were evaluated to identify the isolated effect of each soil char. different numbers of neurons in the hidden layer (k) using grouped test data from The selection of the best input variables was conducted after the five studied soils. Other authors (Arsoy et al. assessing the best number (k) of neurons of the hidden layer of the network. thus. the networks tested with k values need to perform specific calibrations.0160 g/g and r2 value equal to 0. Several tests with 6-k-1 networks were conducted to select the ANNs with the best performance by optimising the number of neu- rons (k) of the hidden layer and its transfer function.

0171 0. it is preferred because it showed the best results in the et al.9009 0. posed using the silt content in addition to the clay content. In other similar studies (Arsoy available.9948 0.0047 0.0065 0. OM. OM.9716 0.9472 0.0183 0.9946 0. Table 6 outlines the Pearson coefficients of linear correlation all of the tested variables showed a marked improvement in the between the tested input variables for the five studied soils.0047 0.9719 0.9949 0.0184 0.0046 0. 2002).0046 0.0064 0.9125 0.9472 0.9949 0. silt 0.0064 0.0046 0. The lowest performance was generated using the silt con.0184 0.9948 0.9125 0.0183 0.0064 0.0171 0. clay.0160 Ka.9951 0.0171 0.0221 0.9718 0.9948 0.9946 0.0183 0.9718 0. content.0160 Ka.0045 0.0183 0.9130 0.9128 0. clay and organic matter.9727 0.0220 0.9728 0. and fit when used alongside Ka in the ANNs.9717 0.0064 0. BD. OM.9947 0.0064 0.0064 0.9127 0. satisfactory. silt 0.9474 0.9472 0. BD. sand.9718 0.0063 0. which are highlighted in bold in Table 5: organic the ANNs because the variables contain redundant information.9728 0.9728 0.9728 0. the bulk density and clay content were the most fre.0160 Ka.0187 0. silt 0.0221 0. sand.0160 Ka.0065 0.9716 0.0166 Ka 0.9126 0.9129 0. using a tan-sigmoid transfer function and one neuron in the out- All of the combinations of the tested input variables showed put.9728 0. during selection. clay.9128 0.0221 0.9950 0.0045 0. clay.9728 0.0064 0.9471 0.0065 0. clay. clay.9728 0.0221 0.0171 0.9718 0.9946 0.0221 0.0220 0. these easily assessed. silt 0.0046 0.0252 g/g and smallest r2 of 0. sand 0.9717 0.0065 0.0396 0.0065 0.0160 Ka.9717 0. The r2 values included in Table 4. promising results except for the networks that were only trained Figs.0063 0.9716 0.0045 0.9472 0.0221 0.0064 0.9127 0.0184 0. silt 0.0184 0.0064 0.9471 0.9444 0.0171 0.0160 Ka.0160 Ka.0160 Ka.9472 0.0160 Ka. clay.0221 0.9727 0.0183 0.9128 0. sand 0.9946 0.9947 0. silt 0.0160 Ka. sand and silt contents.9472 0.9947 0.0221 0.0183 0. clay.9949 0.9717 0.9655 0.0046 0.0171 0.9727 0. present study. silt 0.0171 0.0160 Ka.0160 Ka.9950 0.9471 0.9128 0.9948 0.0171 0.0160 Ka. Zanetti et al.9949 0.9476 0.0221 0.0047 0.9717 0.0171 0.9945 0.0172 0.0184 0.9467 0.0160 Ka.9718 0. a Third-order polynomial equations adjusted for each specific soil.0160 Ka. albeit with a relatively weaker intensity.9128 0.0171 0.0064 0. (2013) pro. OM. OM.0221 0.9124 0. OM.9717 0.0065 0.9718 0. silt 0.9718 0.0160 Ka.0184 0.9717 0.0047 0.0064 0.0221 0.0171 0.0171 0.9107 0. The sand content (fine sand + coarse sand) have The silt content may be used as an input variable for the ANNs an advantage among the studied variables because they are more intended for use in general calibrations of the TDR.9950 0. 6 shows the scatter plots of the recorded and estimated soil some soil physical properties (listed in Table 5) and measure Ka moisture resulting from the cross-validation using Ka and organic with TDR probes.0220 0..0171 0.9945 0.0171 0.9707 0.9717 0.9718 0.9471 0.9468 0. sand 0. 4 and 5 show the scatter plots of the recorded and esti- with Ka.9946 0.9727 0. BD.0222 0.9945 0.9728 0. the capability to be used to dispersions.9717 0.9948 0.0183 0. relative performance.0160 Ka.0183 0.0670 0. OM.9472 0.9128 0. clay. BD.0184 0. OM.0220 0.9472 0. BD.0440 Ka.0045 0.0171 0.9128 0. S. users doesn’t need to calibrate TDR for each soil type.0063 0. Although the performances were different.9728 0. clay 0.0050 0.0183 0.9728 0.0171 0. BD.0183 0.0160 Ka.0045 0. however.9717 0. which is shown in Fig.0220 0. OM.0046 0.0220 0. The networks .9727 0. OM 0.9471 0.0046 0.9728 0.9125 0.0184 0.0045 0.0171 0. BD 0.9472 0.9472 0. on their availability. 2013. Arsoy et al. These cross-validation results indicates a very impor.9727 0.9718 0.9947 0.9728 0. sand 0.0171 0.0063 0.9944 0. silt 0.0065 0.0184 0.9947 0.7934 0.9846 0.9949 0. sand 0.9947 0.9727 0.0046 0.9948 0. sand.9947 0.9471 0.0064 0.9728 0.9945 0.0160 Ka.9716 0.0066 0.9124 0.9470 0.0221 0.0220 0.9124 0.9472 0. However.0171 0.0220 0.0184 0.9472 0. BD.9951 0. only know Fig.0171 0. OM: organic matter content. however.9946 0.0184 0.9940 0.9946 0.0221 0. silt 0. OM.9950 0.9942 0. 3. however.9946 0.9728 0.0184 0.0220 0.9718 0.0160 Ka. organic matter) was adopted for the general calibration of TDR quently used variables.0160 BD: bulk dry density.S.9675 0. including producing the greatest general RMSE of tion for each soil and general ANN trained for all of the soils 0.0046 0..9948 0. sand. BD.0220 0. sand 0.0047 0.9935 0. In such The combination of the most efficient input variables (Ka and studies.0047 0.9472 0. which confirmed the smaller RMSE and values greater predict the water content in a totally independent soil type.9947 0. BD.9728 0.9127 0.0183 0.0160 Ka.9126 0. and a trend of larger errors occurred in the other combina.9901 0.9718 0. silt 0. sand. clay.0160 Ka.0220 0.9719 0. mated soil moisture using the fitted third-order polynomial equa- mances. clay. Namdar-Khojasteh et al.0160 Ka.0071 0.0046 0.0160 Ka.9127 0.9728 0.0160 Ka.9472 0.9946 0.0046 0.9127 0.9473 0.0221 0. OM.0047 0.9126 0.0220 0.0064 0.0183 0.9948 0.0220 0.0473 0.9728 0. BD.9948 0.0171 0. g/g) between the soil moisture recorded and estimated using neural networks with different combinations of input data for the five different soils. (2010).0172 0. silt 0.0184 0.0184 0.0171 0.0049 0. it shows that a stronger collinearity occurred between the vari- A comparison of the results from all of the tested combinations ables sand. OM. The general calibration generated an ANN that had slightly smaller tant characteristic of the tested ANNs.9717 0. sand.0064 0. BD. Such results indicate the pos- allowed us to identify the following input variables with the best sibility of considering only one of those variables as an input for performance.0064 0.0064 0. tent.0184 0. clay 0. bulk density and sand The variables clay and bulk density were also linearly correlated. sand.9949 0.0183 0. The availability of an input variable should also be considered tions of input data that contained silt. Persson et al.9126 0.9128 0.0170 0.9947 0.9344.0183 0. sand.0171 0.0160 Ka. The other variables may also be used depending the use of silt was not indicated.0180 0.0171 0.0184 0.9948 0.0160 Ka.9943 0.9949 0.9728 0.9945 0. sand 0. silt 0. OM 0.9126 0.0065 0.0047 0.0049 0.9473 0.9472 0.9127 0. clay.0047 0.0171 0.9947 0.9946 0.9727 0. silt 0.0184 0.0064 0..0221 0.0171 0. when the variable organic matter is ANNs may produce greater errors.9128 0. for the purpose of data presentation.9728 0. BD.9948 0.0197 0. sand 0. were considered relatively simultaneously using the Ka and organic matter as input variables. clay.0064 0. matter content. 2010. silt 0.0047 0.9951 0.9947 0.9717 0.0220 0.9472 0.0220 0. BD.9728 0.9947 0. clay.9728 0. clay.9948 0.0171 0.9728 0.9948 0.9471 0.9728 0. OM 0.9472 0.0160 Ka.9830 0.0171 0. BD. matter as the input variables of the general ANNs.9947 0.9946 0.0160 Ka.0160 Ka.9718 0. OM 0. Inputs Soil 1 Soil 2 Soil 3 Soil 4 Soil 5 General r2 RMSE r2 RMSE r2 RMSE r2 RMSE r2 RMSE r2 RMSE Ka (polynomial equations)a 0.9717 0.9129 0.0064 0.0220 0. and the organic matter content were only using an ANN with only one hidden layer and thirteen neurons indicated in the study by Namdar-Khojasteh et al.9472 0. BD.9945 0.0171 0.0046 0. / Journal of Hydrology 530 (2015) 657–666 661 Table 4 Coefficient of determination (r2) and root mean square error (RMSE.9125 0. clay. even the variables that showed a lower perfor.9948 0. silt 0.9128 0.

9669 0.9658 0.9872 0. sand.0172 0. clay 0.0232 0. Silt 1 0.9598 0.0251 0.0334 0.9682 0. OM.6958 0.9481 0. sand.0205 Ka.9816 0. 3.7460** eral calibration of the TDR.0186 0.9891 0.0184 0.9925 0.9428 0.8817 0.9595 0.0174 0.9577 0.014.0216 Ka.0202 Ka.0259 0.9198 0.0199 0.0216 0.0071 0.9916 0.9083 0.9209 0. BD.6993*** 0. OM. Namdar-Khojasteh et al.0220 0.9906 0.0140 0.2503 errors (RMSE).0157 0.9452 0.9675 0.0252 0.9502 0. In clayey soils occurs an increase in the resolution of the TDR measuring systems.0129 0.9901 0.0248 0.0135 0.9707 0.9545 0. calibration. OM.0246 Ka.9449 0.9049 0.0190 0.9726 0.0094 0.0227 BD: bulk dry density.0156 0.9083 0.9465 0. OM 0.9417 0. This is still a subject of considerable debate which must be more investigated (Namdar-Khojasteh et al. silt 0.0217 Ka.0181 0. silt 0.0227 0.0195 0. which used four soils to train the neural networks and one soil for the test (cross-validation).0212 Ka.9489* 0.0226 0.0487 0.0246 0.9893 0.3830 0.0063 0. silt 0.0249 Ka. silt 0.9456 0.9224 0.9669 0.9569 0. OM.9884 0.9902 0.9452 0. Inputs Soil 1 Soil 2 Soil 3 Soil 4 Soil 5 General r2 RMSE r2 RMSE r2 RMSE r2 RMSE r2 RMSE r2 RMSE a Ka (polynomial equations) 0.662 S.0213 Ka. / Journal of Hydrology 530 (2015) 657–666 Table 5 Coefficient of determination (r2) and root mean square error (RMSE.0192 0.0237 0. sand 0.0091 0.9040 0. OM. BD.0177 0.9433 0.9379 0.0141 0.9605 0.0222 0.9892 0.9023 0. OM.0349 0.9463 0. clay.0259 0.9477 0.9546 0.0225 Ka.9746 0.0121 0.9815 0.0182 0.0224 0.9526 0.9726 0.9407 0.0221 Ka. Bittelli et al.9342 0. despite having clay content similar to soil 3.0284 0.9470 0.0241 0.0166 0.9368 0.9463 0.9497 0.9451 0.0214 Ka.S.0157 0.9452 0.9438 0.0190 0.9123 0.9579 0.0234 0.0268 0.9515 0.0073 0.0257 0.9733 0.9568 0.0117 0.6000 0.0146 0.9563 0.9328 0.9688 0. silt 0.9065 0. OM. sand.0182 0.0342 0. (2013) reported that a dis- * p-value of the t test equal to 0.0311 0. (2008) presented a methodology to correct TDR- based measurements in conductive soils (clayey soils). sand.9440 0. Those results confirm the potential of using ANNs for the gen- Sand 1 0.0186 0.0193 0.0177 0.0267 0.0233 0.8936 0. Namdar- matter Khojasteh et al.0227 Ka.9859 0.0264 0.0217 Ka.0109 0.9397 0.9431 0. BD 0. Zanetti et al. by using the TDR-measured electrical conductivity of the bulk material as indicator of dielectric losses. sand 0.0187 0.0200 0.0142 0.9723 0.9537 0. the electrical conductivity wasn’t measured in this study and.0190 0.9471 0. silt 0.0203 Ka.9618 0.7461** All of the calibrations obtained for soils 1 and 3 showed greater Bulk density 1 0.9644 0.9350 0.0342 0.9908 0. therefore.9664 0.9512 0. BD. (2012) and Arsoy et al.0245 0.0237 Ka.9407 0. BD.9448 0. a Third-order polynomial equations adjusted for each specific soil.9905 0.0240 0.0201 0.9766 0. silt 0.9896 0.0161 0. clay. Future researches may include these correction Fig.0236 0.0230 0. clay.0243 Ka.9841 0. OM 0. BD.0202 0. were applied in this case to estimate the moisture of soils excluded from the training process.0229 0. OM.9606 0.9565 0.0261 0. silt 0.9525 0.0247 0.9530 0.6745 0.0211 0.9089 0.0186 0.9504 0.0238 0.9593 0. Namdar- .0215 0.0114 0.0145 0.0227 0.9738 0.0438 0. clay.0230 Ka.8971 0.0222 Ka.0214 Ka. bound water polarization and dielectric losses on the TDR mea- surement.0293 0. BD. silt 0.9137 0.9544 0.9814 0.9445 0.9404 0. OM 0.0222 Ka. OM: organic matter content.0252 Ka.8744 0.9344 0.9636 0.9844 0. Architecture of the neural network used for the general calibration of the method proposed by Bittelli et al.0107 0.0231 0.0184 0.9135 0.8873 0.9468 0. sand.3505 Clay 1 0.0536 Ka.8926 0. clay.9109 0. sand 0.9076 0.0116 0.0211 0.0131 0. clay.9573 0. clay.9805 0.0240 0.9478 0.9051 0. silt 0.0175 0.0229 0.0235 0.0193 Ka.8989 0. clay.0305 0.9615 0.9393 0.0196 Ka. OM.9934 0. However.9746 0.9669 0.9461 0.9945 0. BD.0252 0.9085 0.9623 0. (2009).9504 0.9451 0.9065 0.0136 0.0167 0.0246 Ka.0142 0.0243 0.9488 0.0126 0.0239 0.8972 0. silt 0.9563 0.9330 0. g/g) between the recorded and estimated soil moisture with different combinations of input data.148.9901 0. BD.9123 0.0312 0.9555 0.0310 0. 2013). so the plot of each soil was generated Sand Silt Clay Bulk Organic using a trained network with all of the data from the other four density matter soils.9588 0.0209 Ka.0175 0.0175 0.0229 0.9560 0.0239 0.0202 Ka.0308 0. sand.9782 0.0357 0. OM.0232 0.9383 0.0178 0.9848 0.0272 0.0075 0. Table 6 Pearson coefficients of linear correlation between the data of the five studied soils.9786 0.9539 0. TDR using the apparent dielectric constant (Ka) and soil organic matter content Soil 4 showed the smallest error in soil moisture estimated by (OM) as input data. 2010). clay. due to effects of *** p-value of the t test equal to 0.0110 0.9777 0.9433 0.9823 0.9875 0.0197 0.0186 0.9463 0. sand 0.0103 0.0123 0. silt 0. OM 0.0291 0.9373 0. sand 0.0329 0. BD.9383 0.9528 0.9028 0.9841 0.0106 0.8999 0. The cause for this deviation has not been completely identified (Arsoy et al.9764 0.9323 0. (2012) noted that high clay contents may also cause the underestimation and overestimation of moisture using TDR when the soil has low and high moisture.0240 0.9043 0.9440 0. which was most likely because they have higher Organic 1 clay contents.189.9393 0. sand 0.9197 0. (2008).4624 0.8992 0.9554 0.8861 0. sand.0184 0..0207 0.0270 0. silt 0.0191 0.0235 0.0238 0.0232 0. BD.9909 0.9560 0.0148 0.9710 0.0191 Ka.0166 0.9473 0.9866 0.0414 0.0204 0.9353 0. silt 0.9426 0.9312 0. this correction wasn’t performed.0172 0..0166 Ka 0.0094 0.0089 0.0270 0.0240 Ka.9408 0.9107 0.9015 0.9060 0.0245 0.9524 0. BD. BD.0236 0.0233 0.0216 Ka. clay. OM. Bittelli et al.0086 0.8913 0.9831 0.0230 0.0887 0.9529 0.0146 0.9078 0. clay.0159 0.0236 0.9467 0.9222 0.0117 0.0232 0. respectively.9041 0. BD. BD. Stangl et al.0250 0. (2008).0163 0.0092 0.9557 0.0217 0. crepancy between the TDR calibration models and recorded mois- ** p-value of the t test equal to 0.9087 0. sand 0.9535 0. silt 0.8973 0.0227 0. ture of the soils with high clay contents may occur.0100 0.9935 0.0231 0. clay 0.0225 Ka. sand.8889 0.0160 0. clay.0253 0.9634 0.0250 0.0154 0.9072 0. OM.0177 0.9860 0.0252 0. sand 0. clay.9435 0.9416 0.3167 0.0247 0. BD. silt 0.9079 0.0241 0.0151 0. clay.

resulted in different impacts on the soil dielec.. 2013. 2008) as calibrations parame- ture behaviour. In initialization of the free parameters (weights and bias). ibration indexes to clayey soils were worse than the ones of sandy tric constant. . The statistical cal- different minerals. The present study failed to identify the types of soil soils. the types of clay of soils 3 and 4 were sup. however. Zanetti et al. with differ- the present paper different soils from tropical areas with different ent results generated in each training procedure of the ANNs. Former papers (Arsoy et al. observed on the present paper. However.. (2012) noted that similar clay contents.. S. its results can be considered it wasn’t measured as already mentioned. (2010). Dispersion between the recorded and estimated soil moisture using the third-order polynomial equation fitted for each soil. / Journal of Hydrology 530 (2015) 657–666 663 Fig. Thus. It leads to the necessity of conducting new studies to better clay minerals. ters may be explored on future researches. The that future researches may include electrical conductivity mea. Khojasteh et al. 2013) and soil surements to better explain the differences on clayey soils mois. albeit with textures (from sandy to clayey) were analysed. known the limitations of calibration of these soils. Differences electrical con. Besides the limitations ductivity had the potential to explained this behaviour. Once again it is believed promising on TDR calibration of tropical soils with ANNs. electrical conductivity (Bittelli et al. 4. as also pointed posedly different based on the results. inclusion of particles specific surface (Arsoy et al.S. out by Namdar-Khojasteh et al.. 2002) showed It is noteworthy that the ANN training is affected by the random promising results on TDR calibration using ANN to sandy soils. Persson et al.

Conclusions silt. even when testing the networks. the combination was sufficient to conduct a general ical characteristics to estimate the soil moisture in the present calibration of all five soils simultaneously with similar results. Data from five different soils were experimentally generated The results of the cross-validation indicate that the ANNs have in the laboratory with disturbed samples and used for training and the potential to provide general calibrations of the TDR. In that case. despite conducting 100 simulations for each effect of the random initialisation of the network parameters on network tested.664 S. applied to soils excluded from the training process. study. All of the combinations of the variables bulk density and sand. clay and organic matter contents were tested. which in this . the Ka value. Other similar computer experiments might indicate a dif- dom variations and may show differences from other similar com. / Journal of Hydrology 530 (2015) 657–666 Fig. function. the training process. Dispersion between the recorded and estimated soil moisture using the general ANN trained for all of the soils simultaneously with Ka and organic matter as the input variables. 5. the The ANN selected for the general calibration of the TDR has a evaluation of the input variables individually along with Ka enables hidden layer with thirteen neurons and uses a tan-sigmoid transfer the most representative variables to be identified. the results generated in the present study were subjected to ran.S. ferent network configuration with better results because of the puter experiments. and when at least one of those network variables was considered along with ANNs were applied to the general calibration of TDR using phys. 4. Zanetti et al.

. http://dx..3390/s111110114. Muhammad. http://dx. The sand con- Arsoy.11. 2011. Zanetti et al.. Geoderma 195–196.019. The evaluation of the results from all of the tested combinations Altendorf.9083 to 0. S. V. bulk density and silt content. Elliott. 8– available. P. M. F. 691–699. clay. Enhancing TDR based water content tent was also indicated as an advantageous variable because it is measurements by ANN in sandy soils. K. Rimatori. S. Development and validation of a neural network model for soil water content prediction with showed that the input variables with the best performance were as comparison to regression techniques. Ozgur. Prog. Mater. and silt contents. or combinations of several variables may also be used when Babangida.. 2014.W.doi. K.0126 to active infrared thermal methods for in-field applications. sand doi.. 10114–10128.. Remote sensing of soil surface properties. Pallottino.scientific..13031/2013. 6.2012.. Sensors 11 (11)..geoderma. Askari. 457–473. Yilmaz. Scatter plots of the recorded and estimated soil moisture using the ANNs and the variables Ka and organic 2009. bulk density and sand content. E.M.8. Stevens. with a RMSE ranging from 0.. 33 (4).org/10.9891.567. P. Appl. Menesatti. E. F. Trans. Geogr. input variable along with Ka for use in the ANNs applied to the gen. The variables clay content and bulk density doi. study were as follows (in descending order): organic matter. R. the organic matter content was considered the preferred Antonucci. Phys. 567.doi. C..T. http://dx.. Development of a rapid soil water content detection technique using eral calibration of 0.L. 133–144. Papetti. Croft. Costa.W. Comparison of soil water retention functions for humid tropical soils. M.1177/0309133309346644. 2013. S. sand References and clay contents. C. R. Mech. Keskin. The plot of each soil was generated using a trained network with the data from the other four soils.4028/www. 1999. ASAE 42 (3). Stone.M. 13. Giorgi.. results. Based on the / Journal of Hydrology 530 (2015) 657–666 665 Fig. follows (in descending order): organic matter content....doi.. http://dx. more easily assessed. http://dx.1016/j.13230.. C.0237 g/g and an r2 ranging from 0.L. N.. M.

2000. R. 2002. 2010. Sci.002. D. L. J. Geoderma 143. 16 (1). Souza.. P. Field use and calibration of a TDR- Jacobsen.P.. Schjønning. E. M. org/10... Dielectric constant Thomsen. W. water-retention parameters for temperate and tropical soils: a new water. 2008. M. Creutzfeldt. Gué J. J. 1997. T. 2011. http://dx. texture. Reinhart. Becker. C. M.x. J. Søgaard. 213–237. 16 (3)..1016/ probe for measurements of soil solution electrical conductivity. 2010... R.2478/v10098-010-0011-3.R. R. 1694(00)00305-X. Osvaldo. Agríc. measurement. 2001. Earth Syst... 269–311. Sokol. S.D. 151 (2–4). C. Soil Sci.1016/0022-1694 j. van Reeuwijk. W. Soc. time domain reflectometry: a review. 136 (2).F.. 2010.06.. 52–62..666 S. 877–883. Laboratory in south-western Australia. J. O. Calibrating Topp. Staub.. Pisa. Ciênc. A. W.. Soil Till. Piscitelli. 165– 10.1061/(ASCE)0733-9437(2010) 136:2(81). W..doi. Thomsen.. A laboratory calibration of time domain based probe for monitoring water content in a high-clay landslide soil in reflectometry for soil water measurement including effects of bulk density and Austria. Mater.. 161–180.003. 141–153.1590/ 10.. experiments.wasman.04. 13545– P. Rossato. http://dx. 64 (1).. 125.2010. van den Berg. J. 448– S0016-7061(97)00045-1.P. Reichert. Solo. G. Sombroek.. Hydrol. functions for the estimation of moisture retention characteristics of Ferralsols Moret-Fernández.018. Schjønning. 06832010000300006. 31 (3)... http://dx. http://dx. http://dx.. undisturbed peat samples. M.P. 2002.2007.01. Wilczek. 93–100.. B. Vadose Zone J. 2010.. 236 (3–4). jhydrol. F. Water Resour. Evaluation of dielectric constant by clay mineral and soil physico-chemical properties..1016/S0166-2481(04)30021-8. Bronstert. J.2136/sssaj1997. http://dx. 7 (2).doi.X. 46 (3). Zanetti et al.B. C. 2012. 155–180. B. / Journal of Hydrology 530 (2015) 657–666 Bittelli. Davis. 249–256. Eng.doi. 1993. M..B. http://dx. H. Ringgaard. A. Souza. http://dx. http://dx. Zonta. Basile... Soil Sci.3390/s121013545. 299–308. M..2012. Discuss. 61 (3).G. Zehe. Buchan...L. J. 34 (3). A new TDR and related soils. Am. M. J. Measurement of soil water content and electrical conductivity by content measurements in conductive soils. http:// L.M. E.. C. V. Martinez.P. http://dx... Gourc. B. Kay. 6 (1). 212.346. 62 (4).. Hodnett. G.... andic features.. Pedotransfer Sci. 101–112... Sivakumar.. Hydrol. p. 63–68. Res.doi. von Bertoldi. 2011. Schlaeger. Ward.doi..1016/j. Review of state of the art water content. P. Paola. Revi. bioreactor cells using TDR and neutron probes.. O.... P. Englund.. Sensors 12 (10). Graeff. A. 133–142.. http://dx. G. M. Persson.. K. in Portuguese.doi. 67 (1). E. Lamorski. T.042. G. soil moisture measurements. J. M. 1997. Parkin. D. 2002. R.. http://dx..doi. 27 (6).-P.. Francesca.. 2000. Ledo. S0103-90162010000100002. A. Delhi.E.E. Namdar-Khojasteh. Berndtsson.1590/S0100-06832010000200003. R. F. 1999. Sławin´ski. V.B.V. Peñ Persson.. C. 2001. 2012.doi..G. Res. Eng.1016/j. Multi-wire TDR probe evaluation to monitor soil Jonnalagadda. Omid.. Timlin. D.doi. Bauer. 215–220. Hydrol. Coelho... Estimating the least limiting water range of soils from measurements in homogenous and heterogeneous media with laboratory properties and management.A.1016/j. 50 (2).R.doi.. Influence of type of sampling on soil methods for measuring water in landfills. Agric. Wraith.2011. 252–258.. Chen.F. central Italy test-sites. Bras. in Portuguese. Long-term moisture measurements in large-scale principles and applications.01. D. C.J. Y. Using neural networks for Ambient. Hydrol. 2009. Santos.. M. Marked differences between van Genuchten soil dx.03. doi. Folegatti. Shorafa. S. Townsend..06.. Mod.E.. Hydrol. S. Waste Manage. Sommella. comparison of instrumentation performances. 2010. Geoderma 108 solute distribution patterns. temperature and electrical conductivity. Brocca. Stangl.001. sensors to characterize tillage effects on preferential flow. Himmelbauer.H.. 180 (1–3).5194/hess-14-873-2010. Eng. J. http://dx.. D. F. G. R. K.doi. Kaiser. 7 (2)..... 30. Rev. J.jhydrol.T.1016/S0022- Bras..earscirev.. Estimating soil solution electrical conductivity from Comegna. Melone. 389–398. M.. Szypłowska. http://dx. http://dx.1590/ (3–4).W.doi. 2011. http://dx.1016/j. 106 (1–2). http://dx.. R.. Hydromech. B. http://dx. 10 (1). 2003.022.H.. 114– content: measurements in coaxial transmission lines. M. A. Hazard.0063. M. Matsura. Sci.C.doi.1016/S0022-1694(02)00387-6. Shorafa. 66 (5). López. J.. P.geoderma. 147–157. Hydrol. Am.. http://dx. Loiskandl.641327x..1016/j.doi.. Morgner. A.. 78–83. B. Rio de Janeiro. R..D.1590/S0100..doi. 23–31. D. Ambient. A. A. Water Manage. 449. Earth Syst.. M. Using soil water content domain reflectometry probe and a reflection simulation model.. Y. R. Skierucha. Dunin. Appl. moisture dynamics and subsurface wetness control on runoff generation in a Namdar-Khojasteh. Klamt. S. Manual of Soil Analysis Methods.S. 649–658. Soil j.doi. E. L. 261–271. neural networks. CNPS. 2012. Persson. http://dx. calibration of time domain reflectometry to determine moisture content in doi. Moramarco. M. http://dx. M.doi. Growing season evapotranspiration from duplex soils Nagare. Bras. 2007. Buchan. Soil Sci. Herbst. Laurent.G. Graeff. Agric. A. 10.11.. 2007. properties of Fengqiu County soils in the North China Plain using pedo-transfer 327–338.. Caldas. Electromagnetic determination of soil water electromagnetic short soil water sensors. Solo. based soil moisture monitoring system with simultaneous measurement of soil Pearson Education. Minella.doi. doi..M. V. Lapenna. 14. 167 (1). Kintzuger. Hydrol. 2013. 4 (10). 2012. http://dx. Severino. org/10. A. networks for evaluation soil water content with time domain reflectometry. Irrig. Estimating soil hydraulic estimation of soil water retention in Brazilian Soils. W. Minasny. 1–8. Sci. http://dx. J. R... Imhoff.. Tomasella. R. 81– Comparing evapotranspiration rates estimated from atmospheric flux and TDR 89. P. Zhu. A. J.2136/sssaj2002. 1980. Spatial and temporal characterization of water and retention pedo-transfer functions developed for tropical soils. Correction of TDR-based soil water Noborio..1016/S0016-7061(02)00105-2. Omid. http://dx. Vicente. Dragonetti.2009.R. http://dx. R. Hydrol. Uvo.doi..005. J. S. Drain. dielectric constant and TDR probe calibration. Bauer. Hydrol.2006. M.. EMBRAPA. Soil Sci. J.doi. Soc. Ciênc. 133. Sci. Hansen. N. Sivakumar. K....2136/sssaj2000. 73–79.G. J.2012. 10 (2).doi. A TDR- Zehe. Number and spacing Predicting the dielectric constant–water content relationship using artificial between wave guide rods for measurement of soil water content with TDR.. Dev. Hydrol. Am.. Jones. 873– Bras. Zhang. 2010. http://dx.L. Quinton. Comput. Application of TDR to water level obtained from TDR and volumetric moisture of soils in southern Brazil.2136/vzj2010.. 141–159.0060. http://dx. M. http://dx..2013.H.doi. (93)90233-Y. second ed. Soil Sci. R.. Morra. 34 (2).5897/AJAR10. 2004. Rev.. G. R. A. Plot and field scale soil 10. J.04. T.1111/j.5194/hessd-7-269-2010. M.1016/S0168-1699(00)00184-8.B..1590/S1415-43662002000100012.G. B. Perrone. Morgner. Pachepsky. Salvatorelli. Electron. Soil moisture measurements: Schelde.C. Soc.. Agric. Tomasella.009.doi. Coppola.J. D. 574–582.doi. Geoderma 150 (1–2). 172. Neural Networks: A Comprehensive Foundation. 505–515. time domain reflectometry measurements using neural networks. J. Novak. 76–84. Friborg.. http://dx. 415–429.. Hodnett. Earth. B. Afr. M. Annan. http://dx.jhazmat.doi. G. 58 (2). M.2136/ Vadose Zone J. C. http://dx... Calamita...doi. Jacobsen..da S.. Infiltration measurement using a vertical time- Kulasekera.geoderma. J. 2001.1016/j. 2011. J.. Benbelkacem. Schincariol. Han. J. in p. Rev.doi. Yazdani. Pedotransfer functions for the Li. Agric. dos Santos. F.03615995006100030023x.F.. 273 Dielectric properties of a tilled sandy volcanic-vesuvian soil with moderate (1–4). 10.M. 1424–1429.2011. K.1016/S0378-3774(01)00092-0. Agríc.1424.1016/ Tomasella.. Berndtsson. A quality assessment of spatial TDR soil moisture Silva. Oxarango. Bronstert. 7 (1). Loiskandl... White. J.. http://dx. 2002. http://dx. A.2006. 454–455. J. 729–745. Gawande. M. A.... P.doi. A. Stefano. M. second ed. D. C. Predicting soil properties in the tropics. Geoderma 78 (3–4).1365-2389. Process.024.1016/j... Eur. in Portuguese. Hartemink.R. Pedotransfer functions for tropical soils. 683–696. Electrical resistivity and TDR methods for soil moisture estimation in calibration of time-domain reflectometry P. 2010. http://dx. Aragüés.. W.513. . Hayashi. 170–176.V.. A. 09–15. 2010.. A. Time domain reflectometry measurement Bayard. Geoderma 138 (3–4). Using of artificial neural headwater in the Ore Mountains.doi.10. Sci. Rev.1029/WR016i003p00574. 823..