كليدواژه :
گيلان , ويژگي هاي ديريافت , رگرسيون , شبكه عصبي مصنوعي , توابع انتقالي خاك
چكيده لاتين :
Background and Objective: Direct measurement of some soil properties may be difficult, costly and time consuming. So, these properties can be predicted using readily available data. Soil cation exchange capacity (CEC) is an important indicator of soil fertility and pollutant sequestration capacity. Soil hydraulic properties drive the water flow in the soil-plant-atmosphere system and hence control processes such as aquifer recharge or nutrient fluxes between soil and vegetation. Knowledge of soil hydraulic is important for modeling the physical processes related to soil water content. Despite great advances in measurement methods, it is still difficult to determine soil hydraulic properties accurately, especially for undisturbed soils and in the dry range. However, the measurement of the soil hydraulic properties and CEC is time-consuming, labor and expensive. Therefore, the present study aimed to derive the pedotransfer functions (PTFs) for the estimation of field capacity (FC), permanent wilting point (PWP) and CEC in some soils of Guilan province. Materials and Methods: Study area is located in south of Guilan Province, Gilevan region, northern Iran. The climate is aridic. The annual precipitation is 245 mm and the average temperature is 18 °C. A total of 240 soil samples from 0-30 cm layer of this region were collected. Then, both difficult and readily available soil properties such as clay, sand and silt percent, CaCO3, organic matter, bulk density and gypsum were measured. The first step for using statistical methods is to study the normality of data. For this purpose, Kolmogorov-Smirnov test was used. Data were divided into two groups of test (%25) and train (%75). This division carried out in such a way that statistical characteristics of two groups such as minimum, maximum, standard deviation, etc. were similar. Then regression and artifitial neural network (ANN) models set on training data. For prevention of error in ANN process, data converted in standard scale from 0.1 to 0.9. Multi-layer percepteron, feed forward backpropagation and Levenberg-Marquardt functions were used for extension of ANN. Relative root mean square error (RMSE), determination coefficient (R2) and model efficiency factor (MEF) criteria were used for evaluation of models. Results: In regression analysis, for CEC, clay and organic matter percent, in FC moisture content, clay, silt as well as bulk density and for PWP, clay percent showed significant effects in created models. Coefficients of determination in created linear models for CEC, FC and PWP were 0.72, 0.84 and 0.73, respectively.While these coefficients for non-linear models were 0.78, 0.87 and 0.74 for CEC, FC and PWP, respectively. The best PTFs for prediction of difficult available properties in ANNs obtained by multi-layer perceptron model with 2 hidden layers, 8 neurons for FC and PWP, 6 neurons for CEC and considering all inputs. Coefficients of determination for CEC, FC and PWP were 0.98, 0.99 and 0.98, respectively. ANNs designed for prediction of difficulty available properties with inputs include soil readily available properties that have the most sensitivity coefficient with difficulty available properties. Test results of these models were similarity non-linear regression models. The results of models compared with test data showed that the models obtained from ANNs were more accurate than the regression model. Conclusion: In regression method, non-linear models for prediction of soil difficulty available properties were more accurate than linear models. In ANNs, models with inputs including all of the soil readily available properties were more accurate than models with inputs include soil easily available properties that have the most sensitivity coefficient with difficulty available property. Besides, according to results, if the number of readily available data weren’t adequite, regression models could be used with acceptable accuracy. While, these data were enough, then the ANN provides much more accurate results. These means that ANN accuracy decreases by reducing input variables.
