عنوان مقاله [English]
Introduction: Phosphorus is important as an essential element in the production of agricultural products. On the other hand, its ability to induce essential micronutrient deficiency and its negative effects on the environment, have attracted more attention to this element. The knowledge of phosphorus availability conditions in the soil and consequently the accurate management of fertilizer consumption has a key role in the environmental protection. The degree of phosphorus absorption in the soil depends on the environmental factors, soil characteristics and compositions, and phosphorus fertilizer which have been used. The amount of available phosphorus in the soil has relationship with some of the physical and chemical properties of the soil. Since, the soil characteristics are important factors in the reaction of phosphorus in the soil, the present study aimed to investigate and determine the most important soil characteristics affecting the availability of phosphorus using regression and artificial neural network techniques, in the soils of Sistan plain.
Materials and Methods: Soil sampling was done in 1.5×1.5 km intervals, from 0-30 cm depth, and 200 soil samples were taken. The amounts of available phosphorus and the soil properties including the percentages of clay , organic matter, calcium carbonate and the amount of pH were measured. Then, stepwise multivariate linear regression analysis was performed to determine linear relation between available phosphorus and the soil properties. In order to model and validate the regression model, respectively 80 and 20% of data were selected and entered into SPSS software. To train the neural network, multilayer perceptron (MLP) network was used by MATLAB 7.6 package. In this type of network, 70% of data is selected for training, 15% for validation and 15% for testing the model. Levenberg-Marquardt algorithm and hyperbolic tangent (as a transfer function) were used to train the network. The numbers of neurons in the hidden layer were calculated based on the trial and error method and finally the best structure was selected according to the highest R2 and the lowest RMSE value. Moreover, quantifying the importance of variables in the neural network was done through employing connection weight approach. In this method, the connection weights of input-hidden and hidden-output neurons were used to indicate the significance of variables.
Results and Discussion: The values of the coefficient of variation for the soil properties were in the range of 5.66 for pH (the lowest) and 69.90 for available phosphorus (the highest). The high variation of the available phosphorus could be due to the different amounts of phosphorus fertilizers consumption and their diverse rate of conversion to less soluble forms. The validation results of regression and neural network methods showed that the latter technique was more accurate compared with the multivariate linear regression method, in the estimation of available phosphorus, as multi-layer perceptron neural network with 4-6-1 layout predicts nearly 90% of available phosphorus variability using soil properties (percentage of clay, organic matter, calcium carbonate and the amount of pH); however, the obtained regression equation could explain only 43% of phosphorus variances. The reasons for this could be: 1) considering nonlinear relations between the variables in the artificial neural network method, and 2) less sensitivity of this method to the existence of error in input data, comparing with the regression method. The values of R2 and RMSE were 0.43 and 11.23, respectively for training the multivariate linear regression method and 0.91 and 4.28, respectively for training the artificial neural network method. From the investigated soil properties in the current study, the percentage of organic matter and clay were entered in the regression model, and the values of standardized regression coefficient (beta) showed that the first variable is more important to explain the available phosphorus variability. The results of quantifying the importance of variables by the connection weight method showed that pH had the greatest contribution in the variability of phosphorus in the study area. In the other words, the high values of pH were the most important limiting factor for the availability of phosphorus in Sistan soils.
Conclusion: Considering nonlinear and complicated relations between variables, the artificial neural network model is an effective tool to assess the effect of soil properties on the availability of phosphorus in the study area. The results of quantifying the importance of variables by using the connection weight method showed that pH had the greatest contribution in the variability of phosphorus in the study area. In fact, the existence of lime in the soils of the study area, arid climate and lack of precipitation have resulted in the accumulation of basic cations in the soil and consequently increased pH values. Furthermore, the observed average values of pH that are more than 8.5 demonstrated the risk of soil sodicity in the study area. Thus, the management of this area by cultivating tolerant plants could be resulted in increasing organic matter content, which along with using chemical amendments such as sulfur will decrease pH values and increase the availability of phosphorus in Sistan plain. Applying such practices and through it modifying soil characteristics, decreasing the consumption of phosphate fertilizers and preventing their hazardous environmental effects would be expected in long run.