Impaginato 209 Adv. Hort. Sci., 2023 37(2): 209­219 DOI: 10.36253/ahsc­13591 Prediction of chamomile essential oil yield (Matricaria chamomilla L.) by physicochemical characteristics of soil N. Khakipour 1 (*), A. Mohammadi Torkashvand 2, A. Ahmadi 3, W. Weisany 2 1 Department of Soil Science, Savadkooh Branch, Islamic Azad University, Savadkooh, Iran. 2 Department of Horticulture, Science and Research Branch, Islamic Azad University, Tehran, Iran. 3 Department of Horticulture, University of Tabriz, Tabriz, Iran. Key words: Artificial Neural Network (ANN), calcium carbonate equivalent (CCE), multilayer perceptron, nitrogen. Abstract: The purpose of this study was to predict the percentage and yield of chamomile essential oils using the artificial neural network system based on some soil physicochemical properties. Several habitats of chamomile cultiva­ tion were investigated and 100 soil samples were shipped to the greenhouse. The maximum and minimum of pH, EC, K, OM (organic matter), CCE (calcium carbonate equivalent), and clay in soils were 8.75­7.94, 1.6­1.0, 381­135, 2.30­ 0.22, 69­16, and 55.6­32.0, respectively. Growth indices, essential oil percent­ age, and yield were measured. Artificial neural network modeling was carried out to predict the essential oil concentration and yield using three groups of soil properties as a predictor: 1­ nitrogen (N), phosphorus (P), potassium (K), and clay; 2­ pH, EC, organic matter (OM) and clay; 3­ CCE, clay, silt, sand, N, P, K, OM, pH, and EC. So, three pedotransfer functions (PTFs) were developed using the multi­layer perceptron (MPL) with Levenberg­Marquardt training algorithm for estimating chamomile essential oil content. Results evaluation of the accu­ racy and reliability of showed that, the third PTF (PTF3) which developed by all independent variables had the highest accuracy and reliability. Results also showed that, it is possible to predict the concentration and yield of chamomile essential oil based on soil physicochemical properties. This issue is important in terms of land suitability, identify areas susceptible to chamomile cultivation and planning for essential oil yields. 1. Introduction Optimum nutrition is a major condition for improving the quality and quantity of the crops, and it is affected by the soil environment (Barker and Pilbeam, 2007; Hargreaves et al., 2008; Ashoorzadeh et al., 2016; Ajili et al., 2018; Tofighi Alikhani, 2021). Peng et al. (2012) considered the role of soil characteristics to be highly effective in crops yield. Obviously, the production of organic materials in leaves without the presence of mineral (*) Corresponding author: nazanin_kh_43713@yahoo.com Citation: KHAKIPOUR N., MOHAMMADI TORKASHVAND A., AHMADI A., WEISANY W., 2023 ­ Prediction of chamomile essential oil yield (Matricaria chamo‐ milla L.) by physicochemical characteristics of soil. ­ Adv. Hort. Sci., 37(2): 209­219. Copyright: © 2023 Khakipour N., Mohammadi Torkashvand A., Ahmadi A., Weisany W. This is an open access, peer reviewed article published by Firenze University Press (http://www.fupress.net/index.php/ahs/) and distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Competing Interests: The authors declare no competing interests. Received for publication 24 August 2022 Accepted for publication 1 March 2023 AHS Advances in Horticultural Science https://doi.org/10.36253/ahsc-13591 http://www.fupress.net/index.php/ahs/ http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ http://creativecommons.org/licenses/by/4.0/ Adv. Hort. Sci., 2023 37(2): 209­219 210 elements in the process of photosynthesis is not pos­ sible. Each of the macronutrients plays a special role in the metabolism of plant growth. Bernier et al. (1981) concluded that flowering of plants was under the control of nutritional status, and in this regard, the balance between the elements that plant takes from air and soil is very important. Plant mineral compounds are one of the factors affecting the quali­ ty of crops (Prasad and Spiers, 1991). Considering the importance of developing the cultivation of medicinal plants and using their products as natural ingredients compatible with human health, it is necessary to use different cultivation and nutrition methods that increase the essential oil and effective compounds of medicinal plants (Ajili et al., 2019; Savitikadi et al., 2020). Chamomile (Matricaria chamomilla L.) is an annual plant, aromatic, 20­40 cm high that grows wildly on fields and side roads (Omid Beigi, 2004). The plant now has a large dispersal in Europe, Western Asia, North Africa, North and South America, and Australia. The extensive cultivation of this plant is carried out in countries such as Hungary, Germany, Egypt, Czech Republic, Slovakia, and India (Omid Beigi, 2004). In Iran, different species of the genus Matricaria grow in different parts of the country. Chamomile flowers are used to treat stomach, flatu­ lence, and skin lesions. In most western countries, they are used as appetizers and digestible foods. The active ingredient of chamomile has been mentioned for many medicinal properties such as sedative, anti­ spasmodic, stimulating white blood cells and strengthening the body’s defense system, and anti­ bacterial gram­positive and anti­allergic (Salamon, 1992). Accordingly, chamomile is used in many coun­ tries as dry flowers and essential oils in the pharma­ ceutical, food, cosmetic and sanitary industries. In recent years, it has also become one of the most popular pharmaceutical plants in the world (Gheedi Jashni et al., 2015). Upadhyay and Patra (2011) inves­ tigated the effects of calcium and magnesium on chamomile growth and yield and stated that the effect of magnesium on the growth and essential oil of chamomile was higher than that of calcium. One of the main issues in producing agricultural and garden crops is a lack of ability to forecast pro­ duction/yield using accessible and easily measured indicators (Mohammadi Torkashvand et al., 2020). Various factors affect the yield and essential oil con­ tent of the plant, including nutrition and physico­ chemical properties of the soil (El­Gohary et al., 2015; Belal et al., 2016; Radkowski and Radkowska, 2018; Mohammadi Torkashvand et al., 2020). For example, it is hypothesized that one can estimate the yield of a product based on the concentration of nutrients in a leaf (Lahiji et al., 2018; Mohammadi Torkashvand et al., 2020), fruit (Mohammadi Torkashvand et al., 2019) or soil characteristics (Rahmani Khalili et al., 2020; Tashakori et al., 2020). In this case, it will be possible to plan fertilization or choosing the soil susceptible and suitable for planti­ ng, or the farmer has an estimate of his income and, accordingly, plans its costs for the future programs. There are various predictive methods for estimat­ ing several natural variables, among which more transfer functions are used. Different regression methods have been widely used to derive transition­ al functions (Vereecen et al., 1992; Sepaskhah et al., 2000; Marashi et al., 2017; Eslami et al., 2019). These methods consider the relationship between the input data and the data to be predicted to be predefined. Since the soil and plant are natural and heteroge­ neous systems, it is difficult to establish a connection between their properties. Therefore, in these sys­ tems, artificial neural networks (ANN) operate more efficiently than regression methods. Numerous stud­ ies have been carried out to estimate soil variables through artificial neural networks (Zhou et al., 2008; Bocco et al., 2010; Gago et al., 2010; Parvizi et al., 2010; Mokhtari Karchegani et al., 2011; Besalatpour et al., 2013; Dai et al., 2014; Moghimi et al., 2014; Aitkenhead et al., 2015; Marashi et al., 2017; Khanbabakhani et al., 2019; Marashi et al., 2019). Also, some studies have been conducted to predict crop yield by remote sensing, stochastic, artificial neural network (ANN) and simulation models (Bannayan and Crout, 1999; O’Neal et al., 2002; Bartoszek, 2014; Farjam et al., 2014; Domínguez et al., 2015; Emamgholizadeh et al., 2015; Dias and Sentelhas, 2017; Mohammadi Torkashvand et al., 2017; Niedbała, 2019; Mohammadi Torkashvand et al., 2019) based on weather, soil and growth charac­ teristics as input data. Mohammadi Torkashvand et al. (2017) estimated the storage life of kiwifruit based on chemical characteristics of fruits, including the amount of nutrients by analyzing the neural net­ work (NN), identifying it as a superior method in comparison to multiple regression. A similar study was carried out to predict kiwifruit yield based on leaf nutrients by ANN (Mohammadi Torkashvand et al., 2020). Tashakori et al. (2020) evaluated the effi­ ciency of artificial neural network (ANN), multiple lin­ Khakipour et al. ‐ Forecasting chamomile essential oil yield through soil physicochemical 211 ear regression (MLR), and adaptive neuro­fuzzy infer­ ence system (ANFIS) in terms of saffron yield estima­ tion by soil properties in some lands of Golestan province, Iran. According to the results, ANN showed the highest accuracy (R2= 0.58­0.89) in estimating saf­ fron yield as compared to MLR (R2= 0.41­0.47) and ANFIS (R2= 0.41­0.69) models. Poorghadir et al. (2021) concluded that the yield and percentage of essence influenced by the soil properties and nutrition. The purpose of this study was to investigate the importance of soil characteristics on the concentra­ tion and amount of chamomile essential oils and their estimation with respect to some important physicochemical properties of the soil, and investiga­ tion of feasibility of using artificial neural networks for estimating the concentration and amount of chamomile essential oils. 2. Materials and Methods Soil experiments Several habitats or areas of Chamomile cultivation were surveyed in Kermanshah and Hamadan provinces, West of Iran. From 20 areas, 100 soil sam­ ples (five of each area) were taken from soil depths of 0­30 cm and transferred to IAU, Science and Research Branch, Tehran, Iran. The environmental characteristics of the sampling areas, in particular the topographic and climatic characteristics, were simi­ lar. Soil samples were shipped to the laboratory and air­dried and clods were broken down in small parti­ cles with a plastic hammer; then they were passed through a sieve of 2 millimeters (Klute, 1986). Afterward, 0.5 kg of each soil was used for laboratory analysis and the rest was used for greenhouse experi­ ments. The soil samples were analyzed for phospho­ rus, nitrogen and potassium nutrients, pH, Electrical Conductivity (EC), texture, and organic matter. Soil pH and EC were measured in saturated soil extract. Soil texture was determined by hydrometric method and the amount of calcium carbonate equivalent (CCE) was measured using titration method (Paye et al., 1948). The Kjeldahl method was used to measure nitrogen (Goos, 1995). Soil samples were extracted by Soltanpour and Schwab method (1977) and the concentration of available potassium and phosphorus were measured by flame emission and spectropho­ tometry methods (Emami, 1996). Organic matter was measured by Walkley and Black (1934) method. Some statistical data of soils are seen in Table 1. Greenhouse experiment In a completely randomized design, 100 different soil samples were sprayed in a plot (box) with dimen­ sions of 30 to 35 cm and 25 cm in depth, and 20 seeds were planted in each plot. After germination and early growth of the plant, in the quadruple stage, the number of plants was reduced to 10 in each plot. During the growing season, field operations included irrigation, weed control and pest control for the plots were done alike for all the boxes. After full flowering, the flowers were harvested at a maximum of five centimeters of length. The flowers were immediately dried at 60°C with an electronic dryer. In addition to the dry flower yield per plot in kg ha­1, the concentration of essential oil for each soil was obtained. The essential oil content of the sam­ ples was determined by the Kelevenger apparatus by the water distillation method and expressed as g/100 g of dry flowers. The essential oil yield was expressed Table 1 ­ Statistics of data set for estimation of essential oil percentage and essential oil yield OM= Organic Matter; CCE= Calcium carbonate equivalent. Soil properties pH EC (dS/m) N (%) P (mg/kg) K (mg/kg) OM (%) CCE** (%) Sand (%) Silt (%) Clay (%) Essential oil (%) Essential oil yield (kg/ha) Max 8.75 1.60 1.30 81.00 381.00 2.30 69.00 39.40 42.40 55.60 1.57 7.37 Min 7.94 1.00 0.25 8.00 135.00 0.22 16.00 14.50 21.00 32.00 0.01 0.28 Average 8.13 1.28 0.73 22.10 226.10 1.43 33.40 24.11 30.57 45.32 0.69 3.71 Median 8.06 1.25 0.78 16.50 244.00 1.46 28.50 23.20 30.25 46.75 0.76 4.13 Standard 0.22 0.15 0.37 18.93 66.28 0.50 13.83 5.69 4.80 6.16 0.51 2.61 Kurtosis 1.72 ­0.61 ­1.72 5.37 ­0.70 0.34 ­0.06 1.32 ­0.30 ­0.95 ­1.23 ­1.72 Skewness 1.67 0.15 0.02 2.61 0.28 ­0.69 0.92 1.02 0.26 ­0.38 0.26 ­0.12 Adv. Hort. Sci., 2023 37(2): 209­219 212 as kg ha­1 in dry flower yield. Artificial neural network One of the best known rules is multilayer percep­ tron (MLP) learning rule. MLP is a feed forward net­ work in which information flow from input side and pass through the hidden layers to the output layer to produce outputs. In this research, MLP rule and Levenberg­Marquart back propagation algorithm was used for training the artificial neural networks. The Tangent axon function was used as an activation function, which is a nonlinear function. Pourhaghi et al. (2013) also used the Tangent axon functions for predicting the input flows by ANN. NeuroSolutions 5.05 (NeuroDimension, Inc., Gainesville, FL, USA) software was used to design the artificial neural net­ work. The data used in training, validation and test were 60, 20 and 20% of the total data respectively. The training data are used for network education and training. Evaluation data are not used in network training, but this data are used to compare different models and to determine the most suitable network and PTFs. The analysis of the neural network with three sub­ series of variables as input variables was performed to estimate or predict the essential oil percentage and essential oil yield: 1 ­ In first step, total nitrogen, available phosphorus and potassium and clay content of soils, which are the most important factor in soil fertility, were selected as predictors and PTF1 was developed. 2 ­ In the second step, pH, EC, organic matter and clay, which are found in the most common and important measurements in standard soil analy­ ses, were selected as predictors for estimation desired variables, and PTF2 was developed. 3 ­ In third step all of the measured soil properties (N, P, K, OM, CCE, pH, EC, Sand, Silt and Clay) were included as predictor for developing PTF3. Therefore, three PTFs (PTF1, PTF2 and PTF3) were developed using ANNs and their efficiency was com­ pared with each other to find the best and most suit­ able PTF. The number of input layer nods was chosen as the number of input parameters for each desired issue. The number of hidden layers determines the com­ plexity of the grid, and the reason for this complexity is that as the number of hidden layers increases, the number of connections between the nerve layers increases, which leads to network complexity. The number of output layer neurons is equal to the num­ ber of output parameters of the desired problem. After training the network with the data of the training and validation series, the precision and accu­ racy of the generated models were evaluated using the test series data. In order to evaluate network accuracy, the coeffi­ cient of determination (R²) and square error squared (RMSE) were used. (1) (2) In which: yi, ȳi, ŷ, respectively, the measured depen­ dent variable, its mean and the estimated dependent variable, and N is the number of observations. Other criteria used to evaluate the precision of transition functions were the Geometric Mean Error Ratio (GMER) and Geometric Standard Deviation of error ratio (GSDER): (3) (4) Geometric mean of error ratio (GMER) represents the degree of conformance between measured and estimated values. If the GMER is equal to one, it rep­ resents a complete fit between measured and pre­ dicted values. If the GMER is greater than one, it indi­ cates that the predicted values are greater than the measured values, and the GMER less than one is an indication of lower estimated values than the mea­ sured values. Geometric standard deviation of error ratio (GSDER) is a measure of data diffusion. If it is close to one, it shows less diffusion and the differ­ ence between one and the other represents the devi­ ation of most estimates from the measured data. Khakipour et al. ‐ Forecasting chamomile essential oil yield through soil physicochemical 213 3. Results and Discussion Correlation between variables Table 2 shows the correlation between the vari­ ables studied and the percentage and yield of the essential oils. The results presented in this table could be important to find input data series to the neural network. In our study, the percentage and yield of essential oil showed a positive and significant correlation with organic matter, N, P and K contents and clay. Jat and Ahaheat (2006) showed that the use of bio­fertilizers containing nitrogen, phosphorus and potassium increased the growth and amount of essential oil of the fennel plant. Phosphorus plays an important role in seed, flowers germination, vegeta­ tive growth, the acceleration of ripening and the completion of metabolic processes of fruits (Bennett, 1993; Malakouti and Shahabi, 2000; Malakouti et al., 2008). It is also involved in controlling enzymatic reactions and regulating metabolic pathways (Rejali, 2005; Miransari et al., 2007). Potassium affects the amount and quality of herbal essential oils due to its effect on metabolic pathways and enzymatic activity (Pacheco et al., 2008). In the case of potassium defi­ ciency, the quality of some products, especially fruits and vegetables, decreases (Egilla et al., 2005; Mohiti et al., 2011). Potassium deficiency during plant growth leads to a decrease in photosynthetic rate and chlorophyll content (Gerardeaux et al., 2010), activation of enzymes, and reduced growth and yield (Kanai et al., 2007). Potassium interacts with almost all essential elements. Furthermore, a synergistic role of K with either N or P has been already noted (Barker and Pilbeam, 2007). Nurzynska­Wierdak (2013), Cecílio Filho et al. (2015) and Chrysargyris et al. (2017 a) evaluated the impact of different potassium levels (275, 300, 325, 350 and 375 mg/L) on the morphological and bio­ chemical characteristics of spearmint (Mentha spica‐ ta L.). The results showed that the potassium in 325 mg/L treatment could be appropriate for spearmint cultivation and production for essential oil uses. In the same study, Chrysargyris et al. (2017 b) found that the lavender grown in 300 mg L­1 of K was appro­ priate for the essential oil uses/production while the 325 mg L­1 of K were more appropriate for lavender cultivation for fresh and dry matter uses. Due to the significant correlation of essential oils with organic matter, N, P and K contents and clay (Table 2), it was determined three series of data as input data of ANN that is observed in Table 3. Table 3 shows the number of hidden layers, and the number of nodes in the hidden layers in the three input series. Estimation (prediction) of essential oil Figure 1 shows the distribution of actual values (measured) and the estimation or prediction of the percentage of essential oil of chamomile and their conformity in three series of input data. R2 was between the measured values and the estimated essential oils as observed in Table 4. As seen in the first transfer function (PTF1), R2 was 82.71% in the test data and 78.56% in the training data series. The GMER value indicated that almost the network in the test data was not under or over­estimating, and its Table 2 ­ Correlation between the input variables of the neural network and the amount of essential oil (g/100 g of dry flowers) and essential oil yield Variable pH EC N P K OM CCE Sand Silt Clay Essential oil Essential oil yield pH 1 EC 0.511** 1 N ­0.101 0.152 1 P 0.806** 0.546** ­0.382* 1 K ­0.390* ­0.419** 0.099 ­0.323* 1 OM 0.414** 0.675** 0.362* 0.443** ­0.046 1 CCE ­0.267 0.015 0.001 ­0.287 0.044 ­0.544** 1 Sand ­0.032 ­0.339* ­0.439** ­0.029 0.268 ­0.577** 0.648** 1 Silt ­0.202 0.219 0.269 ­0.069 ­0.201 0.409** ­0.258 ­0.408** 1 Clay 0.188 0.178 0.243 0.084 ­0.12 0.276 ­0.465** ­0.711** ­0.353* 1 Essential oil 0.092 ­0.024 0.391* 0.278 0.434** 0.355* ­0.181 0.155 0.153 ­0.277 1 Essential oil yield 0.235 0.163 0.401* 0.423** 0.382* 0.449** ­0.101 0.226 0.246 ­0.421** 0.919** 1 https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! 214 Adv. Hort. Sci., 2023 37(2): 209­219 values were roughly one, but for the training data series the model overestimated. The R2 value in the training data of the second series was 18.06% more than the first transfer function. In the test data, the R2 value increased by 6.83%. The GMER value indicat­ ed that the model underestimated, and this is espe­ cially evident in the test data series. According to the results of Table 4, when each of the nine variables was considered as inputs to the network, the distribution of estimated and measured values (Fig. 1) was reduced around the 1:1 axis and R2 in training data is up to 96.62% and the test rose to 94.78%. Another important aspect is the more accu­ rate estimation of the percentage of chamomile essential oil in both the training and test data, so that the GMER value in both data series was near one, indicating an accurate estimate and a lack or low esti­ mate. The value of GSDER also showed that the low­ est non­conformance of predicted essential oils with the measured values in the data of the training and test series in the third transfer function was observed. Based on these results, increasing vari­ ables from 4 (PTF1 model) to 9 (PTF3 model) decreased error and increased R2 of ANN­model. Mohammadi Torkashvand et al. (2020) employed an artificial neural network (ANN) to evaluate the kiwi yield of Hayward cultivar based on the concentration Fig. 1 ­ Measured values (actual) and estimated concentration of essential oils in diagram 1: 1 test data and their confor­ mance. Table 3 ­ Input data for constructing a neural network in three different transfer functions and the characteristics of neural networks made Transition function Model inputs No. of hidden layers No. of hidden layer nodes 1 No. of hidden layer nodes 2 No. of hidden layer nodes 3 Type of transfer function Type of target function PTF1 N, P, K and clay 3 4 2 1 Tangent axon Levenberg­Marquardt algorithm PTF2 pH, EC, Organic matter 3 4 4 2 Tangent axon Levenberg­Marquardt algorithm PTF3 All variables 1 10 ­ ­ Tangent axon Levenberg­Marquardt algorithm Table 4 ­ Determination coefficient (R2), error (RMSE), GMER and GSDER in two sets of training and test data in predicting essential oil concentration (g/100 g) Transition function Data series R2 RMSE GMER GSDER PTF1 training 0.7856 0.226 1.27 2.37 test 0.8271 0.201 1.04 1.24 PTF2 training 0.9662 0.072 0.91 2.27 test 0.8954 0.237 0.83 1.57 PTF3 training 0.9562 0.023 0.98 1.22 test 0.9478 0.086 1.02 1.32 Khakipour et al. ‐ Forecasting chamomile essential oil yield through soil physicochemical 215 of nitrogen, potassium, calcium, and magnesium in leaves. They concluded that the maximum R2 and the lowest root mean square error were obtained when all nutrients and related ratios were considered as input variables. Mohammadi Torkashvand et al. (2017) tested and compared the performance of an artificial neural network in predicting the firmness of six­month stored kiwifruit with different input datasets. Reversely, they showed that the best answer was obtained using ANN with a RMSE of 0.539 and a correlation coefficient of 0.850 (R2=0.724) when the nitrogen and calcium (N/Ca ratio) were input data (two variables). Prediction of 6­month fruit firmness using nutrient concentrations and their rations (8 variables) datasets resulted in the lowest R­value and the highest error (Mohammadi Torkashvand et al., 2017). Figure 2 shows the dispersion of the measured yield values and the estimation of essential oil in the test series in chart 1:1 and in the 20 samples. According to the results of Table 5, the value of R2 in the test data series in the first transfer function was 91.25%, but less than 80% in the training data. The other thing is the high amount of dispersion, non­ conformance of the estimated and measured and over­estimation data in training. Therefore, in the test series, the accuracy has increased, and in addi­ tion to reducing the dispersion and increasing the conformance, the prediction accuracy has significant­ ly increased, because GMER was approximately one. Forecasting models of plant yield are prognostic tools that can be an important element in precision agri­ culture (Shearer et al., 2000; Dias and Sentelhas, 2017; Prasad et al., 2017; Mohammadi Torkashvand et al., 2017, 2020) and the principal factor in deci­ sion­making systems (Park et al., 2005). Akbar et al. (2018) used a model based on artificial neural net­ work to predict essential oil yield in turmeric (Curcuma longa L.). The data of essential oil, soil and environmental factors were collected from 131 turmeric germplasms in 8 agro­climatic regions of Odisha. Results showed that multilayer­feed­forward neural networks was the most reasonable model to use with R2 value of 0.88. Niazian et al. (2018) calcu­ lated a root mean square error (RMSE) of 0.192 and R2 of 0.901 in predicting essential oil content of Ajowan by using the artificial neural network when the number of rays, pedicels, and flowers per umbel­ let, and a number of umbellets in an umbel, inputted variables. Bahmani et al. (2018) utilized an artificial neural network modeling to predict kinetics of essen­ tial oils extraction from tarragon (Artemisia dracun‐ culus L.) using ultrasound pre­treatment with Clevenger. Based on results, the best prediction per­ formance was belonged to 3­7­1 ANN architecture (0.0008 normalized mean squared error and R2 were respectively 0.0008 and 0.99) which means that it is possible to predict the extraction yield of essential oils with an acceptable precision. Tashakori et al. (2020) research showed that a model with organic matter, phosphorus, potassium, and calcium carbon­ ate as in dependent variables, was the best model (R2 = 0.87) in estimating saffron yield. If the focus is on the test data, in the second transfer function, the accuracy of the model (R2 = 83.23%) was less than the other two functions and its accuracy is much lower than the other two, so that the deviation of the measured and estimated essen­ tial oil yield data was 1.45 (GSDER) and the model had an over­estimation of 26% (GMER=1.26). The important point is that, like the content of essential oil (g/100 g dry matter), the highest accuracy and precision of the neural network model was obtained in predicting the essential oil yield in the third trans­ Fig. 2 ­ Measured (actual) and estimated essential oil yield in diagram 1: 1 test data and their conformance. https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/neural-networks https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/essential-oils https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/essential-oils https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/tarragon https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/artemisia-dracunculus https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/artemisia-dracunculus https://www.sciencedirect.com/topics/agricultural-and-biological-sciences/pretreatment Adv. Hort. Sci., 2023 37(2): 209­219 216 fer function, in which all soil variables (nine variables) were considered as input variables of the network. In the test data series, the lowest error (RMSE = 0.088) and the most consistent measured and estimated data on the essential oil yield were obtained in the third transfer function, although the fitted model had a lower estimation than the first function. Of course, it should be noted that in view of the great difference between the first and third functions in training data, the third function had a higher credibility in general. 4. Conclusions In general, according to the results, the third transfer function (9 variables as input variables of the network) was the most accurate for estimation of the essential oil concentration and for the estimation of essential oil yields. They had the highest R2 and the lowest RMSE values. Also, the estimated values of these functions were the most consistent with the observed values and the least deviation from the 1:1 line. As the R2, RMSE, GMER and GSDER values of the proposed model for estimating essential oil percent for test series data were 94.78, 0.86, 1.02 and 1.32, respectively, and to evaluate the essential oil yield in the test series data respectively, equal to 91.51, 0.608, 0.92 and 1.20 respectively. Therefore, the results showed that with high accuracy and precision, it is possible to predict the concentration and yield of chamomile essential oil based on soil physicochemi­ cal properties. This issue is important in terms of land suitability, making possible to identify areas suscepti­ ble to chamomile cultivation and to plan for essential oil yields. It is suggested to model the prediction of the percentage and yield of chamomile essential oil with an artificial neural network with other charac­ teristics of the soil alone or in combination with these characteristics and compare the results. It is also suggested that other models, such as neuro fuzzy should be evaluated for estimating the concen­ tration and essential oil of chamomile as well as other medicinal plant species. References AITKENHEAD M.J., DONNELLY D., SUTHERLAND L., MILLER D.G., COULL M.C., BLACK H.I.J., 2015 ­ Predicting Scottish topsoil organic matter content from color and environmental factors. ­ European J. Soil Sci., 66: 112­ 120. AKBAR A., KUANAR A., PATNAIK J., MISHRA A., NAYAK S., 2018 ­ Application of Artificial Neural Network model‐ ing for optimization and prediction of essential oil yield in turmeric (Curcuma longa L.). ­ Computers Electronics Agric., 48: 160­178. ASHOORZADEH H., TORKASHVAND A.M., KHOMAMI A.M., 2016 ­ Choose a planting substrate and fertilization method to achieve optimal growth of Araucaria excel­ sa. ­ J. Ornam. Plants, 6: 201­215. BAHMANI L., ABOONAJMI M., ARABHOOSEINI A., MIR­ SAEEDGHAZI H., 2018 ­ ANN modeling of extraction kinetics of essential oil from tarragon using ultrasound pre‐treatment. ­ Engineering Agric., Environ. Food, 11(1): 25­29. BANNAYAN M., CROUT N.M.J., 1999 ­ A stochastic model‐ ling approach for real‐time forecasting of winter wheat yield. ­ Field Crops Research, 62: 85­95. BARKER A.V., PILBEAM D.J., 2007 ­ Handbook of plant nutrition. ­ CRC Press, Boca Raton, FL, USA, pp. 774. BARTOSZEK K., 2014 ­ Usefulness of MODIS data for assessment of the growth and development of winter oilseed rape. ­ Zemdirbyste­Agriculture, 101: 445­452. BELAL B.E.A., EL­KENAWY M.A., UWAKIEM M.K., 2016 ­ Foliar application of some amino acids and vitamins to improve growth, physical and chemical properties of flame seedless grapevines. ­ Egyptian J. Hortic., 43: 123­136. BERNIER G., KINET J.M., SACHS R.M., 1981 ­ The physilogy of flowering. ­ CRC Press, Boca Raton, FL, USA, Vol. 1, pp. 149, Vol. pp. 231. BENNETT W., 1993 ­ Plant nutrient utilization and diagnos‐ tic plant symptoms. ­ In: BENNETT W.F. (ed.) Nutrient deficiencies and toxicities in crop plants. APS Press, St Paul, Minnesota. BESALATPOUR A.A., AYOUBI S., HAJABBASI M.A., MOSAD­ DEGHI M.R., SCHULIN R., 2013 ­ Estimating wet soil aggregate stability from easily available properties in a highly mountainous watershed. ­ Catena, 111: 72­79. BOCCO M., WILLINGTON E., ARIAS M., 2010 ­ Comparison of regression and neural networks models to estimate solar radiation. ­ Chilean J. Agric. Res., 70: 428­435. CECÍLIO FILHO L.A.B., FELTRIM A.L., MENDOZA CORTEZ J.W., GONSALVES M.V., PAVANI L.C., BARBOSA J.C., 2015 ­ Nitrogen and potassium application by fertiga‐ tion at different watermelon planting densities. ­ J. Soil Sci. Plant Nutr., 15(4): 928­937. CHRYSARGYRIS A., XYLIA P., BOTSARIS G., TZORTZAKIS N., 2017 a ­ Antioxidant and antibacterial activities, miner‐ al and essential oil composition of spearmint (Mentha spicata L.) affected by the potassium levels. ­ Industrial Crops Products, 103: 202­212. CHRYSARGYRIS A., XYLIA P., BOTSARIS G., TZORTZAKIS N., 2017 b ­ Optimization of potassium fertilization/nutri‐ tion for growth, physiological development, essential oil composition and antioxidant activity of Lavandula https://www.sciencedirect.com/science/journal/01681699 https://www.sciencedirect.com/science/journal/01681699 https://www.sciencedirect.com/science/journal/01681699 https://www.sciencedirect.com/science/journal/18818366 https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/journal/09266690 https://www.sciencedirect.com/science/journal/09266690 https://www.sciencedirect.com/science/journal/09266690 https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! https://www.sciencedirect.com/science/article/abs/pii/S0926669017302315#! Khakipour et al. ‐ Forecasting chamomile essential oil yield through soil physicochemical 217 angustifolia Mill. ­ J. Soil Sci. Plant Anal., 17(2): 291­ 306. DAI F., ZHOU Q., LV Z., WANG X. LIU G., 2014 ­ Spatial pre‐ diction of soil organic matter content integrating artifi‐ cial neural network and ordinary kriging in Tibetan Plateau. ­ Ecological Indicators, 45: 184­194. DIAS H.B., SENTELHAS P.C., 2017 ­ Evaluation of three sug‐ arcane simulation models and their ensemble for yield estimation in commercially managed fields. ­ Field Crops Research, 213: 174­185. DOMÍNGUEZ J.A., KUMHÁLOVÁ J., NOVÁK P., 2015 ­ Winter oilseed rape and winter wheat growth predic‐ tion using remote sensing methods. ­ Plant, Soil and Envir., 61: 410­416. EGILLA J.N., DAVIES F.T., BOUTTON T.W., 2005 ­ Drought stress influences leaf water content, photosynthesis, and water‐use efficiency of Hibiscus Rosa­sinensis at three potassium concentrations. ­ Photosynthetica, 43: 135­140. EL­GOHARY A., EL GENDY A., HENDAWY S., EL­SHERBENY S., HUSSEIN M., GENEVA M., 2015 ­ Herbage yield, essential oil content and composition of summer savory (Satureja hortensis L.) as affected by sowing date and foliar nutrition. ­ Genetics Plant Physiol., 5(2): 170­178. EMAMGHOLIZADEH S., PARSAEIAN M., BARADARAN M., 2015 ­ Seed yield prediction of sesame using artificial neural network. ‐ European J. Agron., 68: 89­96. EMAMI A.S., 1996 ­ Methods of plant analysis, Volume II. ­ Soil and Water Research Institute, Technical Journal No. 982, Karaj, Iran (In Persian). ESLAMI M., SHADFAR S., MOHAMMADI TORKASHVAND A., PAZIRA E., 2019 ­ Assessment of density area and LNRF models in landslide hazard zonation (case study: Alamout watershed, Qazvin Province, Iran). ­ Acta Ecologica Sinica, 39: 173­180. FARJAM A., OMID M., AKRAM A., FAZEL NIARI Z., 2014 ­ A neural network based modeling and sensitivity analysis of energy inputs for predicting seed and grain corn yields. ­ J. Agric. Sci. Technol., 16: 767­778. GAGO J., MARTÍNEZ­NÚÑEZ L., LANDÍN M., GALLEGO P.P., 2010 ­ Artificial neural networks as an alternative to the traditional statistical methodology in plant research. ­ J. Plant Physiol., 167: 23­27. GERARDEAUX E., JORDAN­MEILLE L., CONSTANTIN J., PEL­ LERIN S., DINGKUHN M., 2010 ­ Changes in plant mor‐ phology and dry matter partitioning caused by potassi‐ um deficiency in Gossypium hirsutum (L.). ­ Environ. Exp. Bot., 67: 451­459. GHEEDI JASHNI M., MOUSAVI NIK S.M., AMINI FAR J., 2015 ­ The role of drought stress and phosphorus and zinc fertilizers on the elemental characteristics and function of German chamomile essences. ­ J. Hortic. Sci., 29(4): 642­651. (In Persian). GOOS R.G., 1995 ­ A laboratory exercise to demonstrate nitrogen mineralization and immobilization. ­ J. Natural Resources Life Sci. Educ., 24: 68­70. HARGREAVES J.C., ADL M.S., WARMAN P.R., 2008 ­ A review of the use of composted municipal solid waste in agriculture. ­ Agric., Ecosystems Environ., 123: 1­14. JAT R.S., AHAHEAT I.P.S., 2006 ­ Direct and residual effects of vermicomposts, biofertilizer and phosphorus on soil nutrient dynamics and productivity of chickpea‐fodder maize sequence. ­ J. Sustain. Agric., 28(1): 41­54. KANAI S., OHKURA K., ADU­GYAMFI J.J., MOHAPATRA P.K., NGUYEN N.T., SANEOKA H., FUJITA K., 2007 ­ Depression of sink activity precedes the inhibition of biomass production in tomato plants subjected to potassium deficiency stress. ­ J. Exper. Bot., 58: 2917­ 2928. KHANBABAKHANI E., MOHAMMADI TORKASHVAND A., MAHMOODI M.A., 2020 ­ The possibility of preparing soil texture class map by artificial neural networks, inverse distance weighting, and geostatistical methods in Gavoshan dam basin, Kurdistan Province, Iran. ­ Arabian J. Geosciences, 13: 237. KLUTE A., 1986 ­ Methods of soil analysis: Part 1. Physical and mineralogical methods. ­ SSSA Soil Science Society of America, Inc., Madison, WI, USA, pp. 1188. LAHIJI A.A., MOHAMMADI TORKASHVAND A., MEHNATKESH A., NAVIDI M., 2018 ­ Status of macro and micro nutrients of olive orchard in northern Iran. ­ Asian J. Water, Envir. Pollution, 15(4): 143­148. MALAKOUTI M.C., SHAHABI A., 2000 ­ Chloro‐calcium spraying is an indispensable necessity in improving the qualitative characteristics of apple. ­ Deputy Directorate for the Promotion of TAT, Ministry of Agriculture, Tehran (In Persian). MALAKOUTI M.J., KARIMIAN N., KESHAVARZ P., 2008 ­ Determination methods for nutritional deficiencies and recommendations for fertilizer. ­ Office for the Publishing of Scientific Works, Tehran, Iran (In Persian). MARASHI M., MOHAMMADI TORKASHVAND A., AHMADI A., ESFANDYARI M., 2017 ­ Estimation of soil aggregate stability indices using artificial neural network �and multiple linear regression models. ­ Spanish J. Soil Sci., 7: 122­132. MARASHI M., MOHAMMADI TORKASHVAND A., AHMADI A., ESFANDYARI M., 2019 ­ Adaptive neuro‐fuzzy infer‐ ence system: estimation of soil aggregates stability. ­ Acta Ecologica Sinica, 39: 95­101. MIRANSARI M., BAHRAMI H., REJALI F., MALAKOUTI M., TORABI H., 2007 ­ Using arbuscular mycorrhiza to reduce the stressful effects of soil compaction on corn (Zea mays L.) growth. ­ Soil Biol. Biochem., 39: 2014­ 2026. MOGHIMI S., MAHDIAN M.H., PARVIZI Y., MASIHABADI M.H., 2014 ­ Estimating effects of terrain attributes on local soil organic carbon content in a semi‐arid pasture‐ land. ­ J. Biodiversity Environ. Sci., 5(2): 67­106. MOHAMMADI TORKASHVAND A., AHMADI A., GÓMEZ P.A., MAGHOUMI M., 2019 ­ Using artificial neural net‐ work in determining postharvest life of kiwifruit. ­ J. Sci. https://link.springer.com/article/10.1007/s12517-020-5134-1#auth-1 https://link.springer.com/article/10.1007/s12517-020-5134-1#auth-2 https://link.springer.com/article/10.1007/s12517-020-5134-1#auth-3 https://link.springer.com/article/10.1007/s12517-020-5134-1#auth-3 Adv. Hort. Sci., 2023 37(2): 209­219 218 Food Agric., 99(13): 5918­5925. MOHAMMADI TORKASHVAND A., AHMADI A., NIKRAVESH N., 2017 ­ Prediction of kiwifruit firmness using fruit mineral nutrient concentration by artificial neural net‐ work (ANN) and multiple linear regressions (MLR). ‐ J. Integrative Agric., 16(7): 1634­1644. MOHAMMADI TORKASHVAND A., AHMADIPOUR A., KHANEGHAH A.M., 2020 ­ Estimation of kiwifruit yield by leaf nutrients concentration and artificial neural net‐ work. ‐ J. Agric. Sci., 158(3): 185­193. MOHITI M., ARDALAN M.M., MOHAMMADI TORKASH­ VAND A., SHOKRI VAHED H., 2011 ­ The efficiency of potassium fertilization methods on the growth of rice (Oryza sativa L.) under salinity stress. ­ African J. Biotech., 10: 15946­15952. MOKHTARI KARCHEGANI P., AYOUBI S.H., HONARJU N., JALALIAN A., 2011 ­ Predicting soil organic matter by artificial neural network in landscape scale using remotely sensed data and topographic attributes. ­ Geophysical Res. Abstracts 13, article no. EGU2011­ 1075. NIAZIAN M., SADAT NOORI S., ABDIPOUR M., 2018 ­ Artificial neural network and multiple regression analy‐ sis models to predict essential oil content of ajowan (Carum copticum L.). ­ J. Appl. Res. Medicinal Aromatic Plants, 9: 124­131. NIEDBAŁA G., 2019 ­ Simple model based on artificial neur‐ al network for early prediction and simulation winter rapeseed yield. ­ J. Integrative Agric., 18: 54­61. NURZYNSKA­WIERDAK R., 2013 ­ Does mineral fertilization modify essential oil content and chemical composition in medicinal plants? ­ Acta Scientiarum Polonorum Hortoru, 12: 3­16. O’NEAL M.R., ENGEL B.A., ESS D.R., FRANKENBERGER J.R., 2002 ­ Neural network prediction of maize yield using alternative data coding algorithms. ­ Biosystems Engineering, 83: 31­46. OMID BEIGI R., 2004 ­ Production and processing of medic‐ inal plants, Volume 3. ­ Astan Quds Razavi Press, Mashhad, Iran, pp. 397. (In Persian). PACHECO C., CALOURO F., VIEIRA S., SANTOS F., NEVES N., CURADO F., FRANCO J., RODRIGUES S., ANTUNES D., 2008 ­ Influence of nitrogen and potassium on yield, fruit quality and mineral composition of kiwifruit. ­ Intern. J. Energy Environ., 2: 9­15. PARK S.J., HWANG C.S., VLEK P.L.G., 2005 ­ Comparison of adaptive techniques to predict crop yield response under varying soil and land management conditions. ­ Agricultural Systems, 85: 59­81. PARVIZI Y., GORJI M., OMID M., MAHDIAN M.H., AMINI M., 2010 ­ Determination of soil organic carbon vari‐ ability of rainfed crop land in semi‐arid region (neural network approach). ­ Modern Appl. Sci., 4: 25­33. PAYE A.L., MILLER R.H., KEENEY D.R., 1948 ­ Method of soil analysis, Part 2 Chemical and Microbiological Properties. ­ SSSA, Soil Science Society of America Inc., Madison, WI, USA, pp. 1159. PENG J., ZHANG Y.Z., PANG X.A., 2012 ­ Hyperspectral fea‐ tures of soil organic matter content in South Xinjiang. ­ Arid Land Geography, 55(5): 740­746. POORGHADIR M., MOHAMMADI TORKASHVAND A., MIR­ JALILI S.A., MORADI P., 2021 ­ Interactions of amino acids (proline and phenylalanine) and biostimulants (salicylic acid and chitosan) on the growth and essential oil components of savory (Satureja hortensis L.). ­ Biocatalysis Agric. Biotechnoly, 30: 101815. POURHAGHI A., ALI A.M.A., RADMANESH F., PODEH H.T., SOLGI A., 2013 ­ Predicting the input flow into the Dez dam reservoir using the optimized neural network by genetic algorithm. ­ Intern. J. Engineering, 2(6): 231­ 236. PRASAD A., PRAKASH O., MEHROTRA S., KHAN F., MATHUR A.K., MATHUR A., 2017 ­ Artificial neural network‐ based model for the prediction of optimal growth and culture conditions for maximum biomass accumulation in multiple shoot cultures of Centella asiatica. ­ Protoplasma, 254: 335­341. PRASAD M., SPIERS T.M., 1991 ­ The effect of nutrition on the storage quality of kiwifruit (a review). ­ Acta Horticulturae, 297: 579­585. RADKOWSKI A., RADKOWSKA I., 2018 ­ Influence of foliar fertilization with amino acid preparations on morpho‐ logical traits and seed yield of timothy. ­ Plant, Soil Environ., 64: 209­213. RAHMANI KHALILI M., ESMAEIL ASADI M., MOHAMMADI TORKASHVAND A., PAZIRA E., 2020 ­ Regression analy‐ sis for yield comparison of saffron as affected by physic‐ ochemical properties of the soil. Case study in Northeast of Iran. ­ Agric. Res., 9: 568­576. REJALI F., 2005 ­ A glimpse over the coexistence of micror‐ rizha “basics and applications”. ­ Technical J. No. 468. Soil and Water Research Institute. Senate publications. Tehran, Iran (In Persian). SALAMON I., 1992 ­ Chamomile: A medicinal plant. ­ Herb, Spice, Medicinal Plant Digest, 10(1): 345­354. SAVITIKADI P., JOGAM P., KHAN ROHELA G., ELLENDULA R., SANDHYA D., RAO ALLINI V., ABBAGANI S., 2020 ­ Direct regeneration and genetic fidelity analysis of regenerated plants of Andrographis echioides (L.). An important medicinal plant. ­ Industrial Crops Products, 155: 112766. SEPASKHAH A.R., MOOSAVI S.A.A., BOERSMA L., 2000 ­ Evaluation of fractal dimensions for analysis of aggre‐ gate stability. ­ Iran Agric. Res., 19: 99­114. In Persian with English abstract. SHEARER J.R., BURKS T.F., FULTON J.P., HIGGINS S.F., 2000 ­ Yield prediction using a neural network classifier trained using soil landscape features and soil fertility data. Annual International Meeting, Midwest Express Center. ASAE Paper No. 001084, Milwaukee, Wisconsin, USA, pp. 5­9. SOLTANPOUR P.N., SCHWAB A.P., 1977 ­ A new soil test for javascript:void(0) javascript:void(0) javascript:void(0) javascript:void(0) https://www.sciencedirect.com/journal/journal-of-applied-research-on-medicinal-and-aromatic-plants https://www.sciencedirect.com/journal/journal-of-applied-research-on-medicinal-and-aromatic-plants https://www.sciencedirect.com/journal/journal-of-applied-research-on-medicinal-and-aromatic-plants https://www.sciencedirect.com/journal/journal-of-applied-research-on-medicinal-and-aromatic-plants/vol/9/suppl/C https://ideas.repec.org/a/eee/agisys/v85y2005i1p59-81.html https://ideas.repec.org/a/eee/agisys/v85y2005i1p59-81.html https://ideas.repec.org/a/eee/agisys/v85y2005i1p59-81.html https://ideas.repec.org/s/eee/agisys.html javascript:void(0) javascript:void(0) javascript:void(0) javascript:void(0) javascript:void(0) javascript:void(0) javascript:void(0) https://link.springer.com/article/10.1007/s40003-020-00455-6#auth-2 https://link.springer.com/article/10.1007/s40003-020-00455-6#auth-3 https://link.springer.com/article/10.1007/s40003-020-00455-6#auth-3 https://link.springer.com/article/10.1007/s40003-020-00455-6#auth-3 https://link.springer.com/article/10.1007/s40003-020-00455-6#auth-4 https://www.sciencedirect.com/science/article/abs/pii/S092666902030683X#! https://www.sciencedirect.com/science/article/abs/pii/S092666902030683X" /l "! https://www.sciencedirect.com/science/article/abs/pii/S092666902030683X" /l "! https://www.sciencedirect.com/science/article/abs/pii/S092666902030683X" /l "! https://www.sciencedirect.com/science/article/abs/pii/S092666902030683X" /l "! https://www.sciencedirect.com/science/article/abs/pii/S092666902030683X" /l "! https://www.sciencedirect.com/science/journal/09266690 https://www.sciencedirect.com/science/journal/09266690/155/supp/C Khakipour et al. ‐ Forecasting chamomile essential oil yield through soil physicochemical 219 simultaneous extraction of macro‐ and micronutrients in alkaline soils. ­ Commun. Soil Sci. Plant Analysis, 8: 195­207. TASHAKORI F., MOHAMMADI TORKASHVAND A., AHMADI A., ESFANDIARI M., 2020 ­ Comparison of different methods of estimating saffron yield based on soil prop‐ erties in Golestan province. ­ Commu. Soil Sci. Plant Anal., 51(13): 1767­1779. TOFIGHI ALIKHANI T., TABATABAEI S.J., MOHAMMADI TORKASHVAND A., KHALIGHI A., TALEI D., 2021 ­ Effects of silica nanoparticles and calcium chelate on the mor‐ phological, physiological and biochemical characteris‐ tics of gerbera (Gerbera jamesonii L.) under hydroponic condition. ­ J. Plant Nutr., 44(7): 1039­1053. UPADHYAY R.K., PATRA D.D., 2011 ­ Influence of secondary plant nutrient (Ca and Mg) on growth and yield of chamomile (Matricaria recutita L.). ­ Asian J. Crop Sci., 3: 151­157. VEREECEN H., DIELST J., VAN ORSHOVEN J., FEYEN J., BOUMA J., 1992 ­ Functional evaluation of pedotrans‐ fer functions for the estimation of soil hydraulic proper‐ ties. ­ Soil Sci. Soc. Amer. J., 56: 1371­1378. WALKLEY A.J., BLACK I.A., 1934 ­ Estimation of soil organic carbon by the chromic acid titration method. ­ Soil Sci., 37: 29­38. ZHOU T., SHI P.J., LUO J.Y., SHAO Z.Y., 2008 ­ Estimation of soil organic carbon based on remote sensing and process model. ­ Frontiers Forestry China, 3: 139­147. https://scholar.google.com/scholar?oi=bibs&cluster=859759651701355766&btnI=1&hl=en https://scholar.google.com/scholar?oi=bibs&cluster=859759651701355766&btnI=1&hl=en https://scholar.google.com/scholar?oi=bibs&cluster=859759651701355766&btnI=1&hl=en https://scholar.google.com/scholar?oi=bibs&cluster=859759651701355766&btnI=1&hl=en https://scholar.google.com/scholar?oi=bibs&cluster=859759651701355766&btnI=1&hl=en https://scholar.google.com/scholar?oi=bibs&cluster=859759651701355766&btnI=1&hl=en https://scholar.google.com/scholar?oi=bibs&cluster=859759651701355766&btnI=1&hl=en