Turmeric samples were collected from 119 sites covering 8 agroclimatic regions at different altitudes (2.8 m -872 m) from different districts of Odisha, from May to July, 2013. From each site, representative plant samples were collected in replicates of three. The interval between two consecutive replicates was 2 m -5m. The collected fresh rhizomes were washed with running tap water to remove the soil particles, followed by washing with double distilled water. The washed rhizomes were then air dried and used for curcumin estimation. Soil samples from each sampling site were collected in replicates and brought to the laboratory for analysis of soil nutrients. From each sampling site, the data on environmental factors such as temperature, humidity and rainfall were taken as monthly averages from May, 2012 to February, 2013 and properly documented.

Rhizomes of turmeric were collected from field before onset of dormancy (i.e., November) for extraction of curcumin. The rhizomes were cleaned thoroughly with water, cut into small pieces and air dried. The air dried rhizomes were powdered in mortar with a pestle, 0.1 g of powdered rhizome was taken in a flat bottom flask and 75 ml of acetone was added and refluxed for 4 h. The refluxed residue was cooled, filtered and washed with 100 ml of acetone. The absorbance of the diluted sample and that of the standard curcumin (95% HPLC Purified, Charak) solution were measured at 420 nm by spectrophotometer (Thermo Scientific, Evolution 220 UV Visible) and curcumin percentage of the sample was estimated according to the ASTA method (American Spice Trade Association, 1997).

The soil samples were collected in replicates of three from each site. Soil samples were collected from interior inter-row area of turmeric growing region from 0 to 15 cm depth of each agroclimatic zone. Soil samples were thoroughly mixed and stored in zip-lock polythene bags to maintain the moisture level similar to the field conditions. Samples were then stored in the laboratory for 2–3 weeks. About 200 g of soil were collected and then sieved through a 2 mm mesh. The fine soil was used for nutrient analysis.

pH of the soil samples was determined in 1:2 soil:water suspension after equilibration for half an hour with intermittent stirring using the Systronics pH meter (Model MKVI).

Total nitrogen was determined by using alkaline KMnO₄ method (Subbiah and Asija, 1956). Hundred milliliter of 0.32% KMnO₄ solution was added to 20 g of soil sample in an 800 ml Kjeldahl flask and then 2.5% NaOH solution was added with some distilled water. Distillation was continued and was collected at receiver tube in the 250 ml conical flask containing 20 ml boric acid (2%) with mixed indicator. The distillate was titrated against 0.02N H₂SO₄ taken in burette to a pink color end point and available nitrogen was calculated.

Total phosphorus in the surface soil samples were determined using Brays No-1 method. 2 g soil was extracted with 40 ml of Bray’s-1 solution (0.03NH₄F and 0.025N HCl) and was shaken for 5 min by mechanical shaker and filtered through Whatman filter paper. 0.5 ml aliquot was transferred into a 25 ml flask 0.5 ml of ammonium molybdate solution was added and distilled water was added to make the volume up to 25 ml. Diluted SnCl₂ (0.5 ml was diluted to 66 ml) was added and the volume was made up to the mark. Phosphorous concentration was analyzed by the help of spectrophotometer (Model: Systronics 166) at 660 nm. The concentration was calculated from the standard graph prepared by taking different phosphorous concentration. Available phosphorous in the profile soil samples were determined by extracting the soil with Olsen’s reagent (0.5 M NaHCO₃, pH 8.5). The phosporous concentration was determined colorimetrically following the procedure of ascorbic acid reduced sulphomolybdic acid blue color in sulphuric acid system at 882 nm by using spectrophotometer (Model: Systronics 166) (Olsen and Sommers, 1982). Soil potassium content was determined by taking 5 g of soil sample in 100 ml conical flask and 25 ml of 1N NH₄OAc solution was added to it. Then it was shaken with the help of a mechanical shaker for 5 min and the potassium concentration in the filtrate was analyzed with the help of a flame photometer (Model:Systronics 128).

Organic carbon (OC) content of the soil was determined by Wet digestion procedure of Walkley and Black as outlined in soil chemical analysis (Jackson, 1973).

The relationship between the curcumin content with the individual parameter of soil and climatic factors was investigated using simple linear regression model. Minitab statistical package was used for the regression analysis. The degree and nature of relationship was estimated applying pearson correlation coefficient (r-value). Further the statistical significance of the relationship was measured at different probability level such as p = 0.05, p = 0.01 and p = 0.001.

For predictions and classifications of data neural networks are the useful tools. The main structure of the ANN is made up of the input layer and the output layer (Yang et al., 2014). For a complicated system a neural network model can determine the input–output relationship based on the strength of their interconnections presented in a set of sample data (Mark et al., 2000). Data approximation and signal filtering functions can be provided by such a model (Clifford and Lau, 1992). In this case, a back propagation (BP) neural-network model was created using STATISTICA software (statsoft) and trained using the environmental parameters and soil parameters as the inputs and curcumin content as the output.

The structure of the neural network model consisted of 11 input neurons in the input layer and one output neuron in the output layer to match the 11:1 input – output pattern of the training data set. For the neural network determined by a trial and error method, one hidden layer with eight neurons was the best topology (Figure 1). The neural- network model was trained in an iterative training process using the obtained training set as follows:

The input variables in this model were normalized based on their possible ranges to avoid data saturation using the following equation.

Where a, a_min, a_max and a_norm are the real valued input variable, the minimum and maximum values of the input variable and its normalized value respectively. The output from this model is an indexed value that corresponds to the input variable. To get the real-valued output, the indexed output value needs to be denormalized according to the following equation:

Where b, b_min, b_max and b_norm are the real-valued output variable, the minimum and maximum possible values of the real-valued output and the indexed output value from the ANN model respectively.

For predictions and optimization neural networks are useful tools. In this case, a back-propagation (BP) neural-network model was created using STATSTICA software and trained using the environmental and soil parameters as the inputs and curcumin content as the output. The structure of the neural network model consisted of 11 input neurons in the input layer and one output neuron. One hidden layer with eight neurons was the optimal topology for the neural network determined by a trial and error method (Figure 1). Heinzow and Tol (2003) and Khazaei et al. (2008) have chosen four-layer backpropagation networks with two hidden layers for prediction of crop yield. These results again confirm that given sufficient hidden layer and neurons multi-layer feed-forward network architectures can approximate virtually any function of interest to any desired degree of accuracy. Mohammadi et al. (2010) devised ANN models to estimate yield level of kiwi fruit production in Mazandaran province of Iran. They used annual energy consumption per hectare of fruit production by different inputs as input variables and the yield level of fruit as output parameter. From this study they concluded that the ANN model with 6-4-1 structure was the best model for predicting the kiwi fruit yield in the surveyed region. Rahman and Bala (2010) reported that a model with 6-9-5-1, i.e., a network having an input layer with six neurons, two hidden layers with 9 and 5 neurons and one neuron in the output layer, was the best model for predicting jute production in Bangladesh. Pahlavan et al. (2012) developed various ANN models to estimate the production yield of greenhouse basil in Iran.

Results showed that feed-forward neural networks trained by BP algorithm had a good ability for creating of non-linear mapping between input and output parameters. Among the various structures model of good performance was produced by the MLFN-8 (Figure 1). Before arriving at this optimum, several tests were carried out with different configurations of the neural network. The configuration that had a minimal dimension and which gave satisfying results was retained with trial and error method. The number of hidden layers and neurons had a large effect on model performance. It was found that increasing the number of hidden layers increased the modeling capability. For the final network, 11-8-1 structure, the training, testing and validation RMSE values were 0.54, 0.17 and 0.9 respectively (Figure 2). Ideally, the RMSE values should be close to zero, indicating that on an average there were appropriate training and prediction performances.

Training results of MLFN-8: Training results of MLFN-8 model were obtained from the experiments. In training process, the comparison result between predicted values and actual values is presented in Table 1. A plot between target (actual) values and output (predicted) values is provided in Figure 3A. The relationship between predicted values and actual values implies that the training process is precise.

Testing results of MLFN-8: In testing process, as shown in Table 2 and Figure 3B, the comparison between predicted values and actual values are also very close, which means that the MLFN-8 model is precise while predicting.

Validation results of MLFN-8: the comparison result between predicted values and actual values is depicted in Table 3. A plot between target and output values is provided in Figure 3C. The relationship between predicted values and actual values implies that the validation process is precise.

The data obtained from soil analysis and climatic factors of 119 places were statistically analyzed using ANOVA test. ANOVA is a method of portioning variability into identifiable sources of variation and the associated degree of freedom in the model. Eleven control parameters were considered in the present study. Each factor affected the response to a varying degree.

The response surface curves were plotted to understand the interaction of the variables and to determine the optimum level of each variable for maximum response. The response surface plot and contour plot for minimum relative humidity, altitude and curcumin content are shown in Figures 4A,B respectively. The figures show the combined effect of minimum relative humidity and altitude on curcumin content. The curcumin content increases with gradual increase in altitude value nearly up to 550 m and minimum relative humidity. The response surface plot and contour plot for altitude, nitrogen content of soil and curcumin content are shown in Figures 5A,B and they show the effect of altitude, and nitrogen content of the soil up to a certain range. The nitrogen content range lies between 200 and 600 kg ha^-1 and the altitude range 0–100 m then from 540 m onward. The response surface plot and contour plot for minimum relative humidity, nitrogen content of soil and curcumin content are shown in Figures 6A,B and it shows the effect of minimum relative humidity and nitrogen content of soil on curcumin content. Curcumin content is increased with gradual increase in minimum relative humidity and nitrogen content of the soil.

The response surface plot and contour plot for soil pH, altitude and curcumin content are shown in Figures 7A,B and it shows the effect of pH of soil and altitude on curcumin content. Slightly acidic soil (pH value ~6.2) with low altitude favored curcumin content.

The response surface plot and contour plot for minimum relative humidity, pH and curcumin content are shown in Figures 8A,B and it shows the effect of minimum relative humidity and pH on curcumin content. Curcumin content is favored by acidic pH (~6.2) and an increasing value of minimum relative humidity. The response surface plot and counter plot for pH, nitrogen content and curcumin content are shown in Figures 9A,B and they show the effect of pH and nitrogen content on curcumin content. The curcumin content is favored by the increase in both pH (in slightly alkaline range) and nitrogen content of the soil.

Several studies on the optimized conditions for the osmotic dehydration process using response surface plots have been published for papaya, potato, diced pepper and banana (El-Aouar et al., 2006; Ozdemir et al., 2008; Mercali et al., 2011; Maran et al., 2013).

Results of the sensitivity analysis are shown in Figure 10. The calculated curcumin content showed the greatest sensitivity to minimum relative humidity with regression value of 0.26. The next greatest sensitivity was to altitude, followed by soil nitrogen content and pH with regression value of 0.21, 0.18, and 0.16 respectively. Maximum relative humidity came fifth in influencing the curcumin production. The calculated curcumin content showed considerably less sensitivity to the remaining 6 input factors (Relative Humidity, Avg Rainfall, Maximum avg temperature, OC, phosphorous content and Potasium content). A sensitivity analysis is the method of studying the behavior of a model, and assessing the significance of each input variable on the values of the output variable of the model. Sensitivity analysis provides insight into the usefulness of individual variables. By the help of this kind of analysis, it is possible to judge which inputs for curcumin content parameters should be considered as the most significant and least significant. To evaluate the predictive ability and validity of the developed models, a sensitivity analysis was performed with the best network for yield of curcumin. Dai et al. (2011) examined effect of soil moisture and salinity stress on sunflower yield by using ANN. Results of the sensitivity analysis indicated that at each level of salinity stress, at different soil moisture, the yield would vary. Pahlavan et al. (2012) carried out sensitivity analysis of input parameters on basil production. Sensitivity analysis revealed that chemical fertilizer; farm yard manure (FYM), diesel fuel and other chemical energies had the highest sensitivity on output; while the sensitivity of electricity, human labor and transportation energies was relatively low.

Curcumin content was analyzed in triplicate from 119 accessions of turmeric. It was found that the curcumin content from the rhizomes obtained from Rayagada was comparatively more (7.2%) than that in the rhizome samples collected from other districts with a minimum from 0.4%, (Kandhamal). The variation in curcumin content was significant among the districts. From the developed model it was found that the curcumin content in the rhizome sampled from these sites decreased progressively from low altitude. The respective correlation coefficient (r) was 0.04. The variation in curcumin content of turmeric is highly dependent on climatic factors. The variation in curcumin content was related positively with humidity; r = 0.02 (afternoon) and r = 0.08 (forenoon). The correlation coefficient between curcumin content was 0.01 with average rainfall. The linear correlation coefficient (r) was 0.002 for maximum temperature and 0.1 for minimum temperature and are positively correlated with curcumin content. The correlation coefficient (r) was 0.02 between soil OC and curcumin content. The results indicated that OC favored curcumin production in the turmeric rhizome.

The result of the regression between soil pH and curcumin was 0.03. The correlation coefficients (r) was 0.05 between soil nitrogen content (N) and curcumin content. The statistical analysis results of the linear regression between phosphorous (P) and curcumin content was 0.06. The correlation content (r) was 0.003 between soil potassium (K) and curcumin content. According to the findings, curcumin content of turmeric can be improved through the soil management to create the soil conditions similar to the original ones. An optimization process using the earlier developed ANN model is shown in Figure 11.

Medicinal productivity of many medicinal herbs is more in their original habitat than in cultivated lands. Soil nutrient and environmental factors similar to original habitats must be most suitable for the active compound production. According to the results of this research, curcumin production of turmeric can be improved through the soil management to create the soil conditions similar to the original ones.

Further, plant secondary metabolism is a complex physiological process. The secondary metabolite production is influenced by the plant’s own physiological age, status and other environmental factors. So, the effects of soil on curcumin content of turmeric are more complicated. However, it is assumed that the variation in curcumin content is dependent on these factors, which needs further research.

The prediction model (ANN) (Figure 12) developed in this study will be helpful to get maximum yield of curcumin in turmeric. This model could predict the curcumin content for a new site which is very close to the experimental value. The results demonstrated that using a combination of soil and environmental data, we were able to successfully predict curcumin content with the developed ANN model. The ANN model was 80% accurate when tested using the reserved data set. The same ANN model, without retraining was tested at Pottangi. This model predicted the curcumin content for Pottangi is 0.8 which is similar to the experimental value (Figure 12).