aboalhasan fathabadi; hamed rouhani
Abstract
Introduction: In order to implement watershed practices to decrease soil erosion effects it needs to estimate output sediment of watershed. Sediment rating curve is used as the most conventional tool to estimate sediment. Regarding to sampling errors and short data, there are some uncertainties in estimating ...
Read More
Introduction: In order to implement watershed practices to decrease soil erosion effects it needs to estimate output sediment of watershed. Sediment rating curve is used as the most conventional tool to estimate sediment. Regarding to sampling errors and short data, there are some uncertainties in estimating sediment using sediment curve. In this research, bootstrap and the Generalized Likelihood Uncertainty Estimation (GLUE) resampling techniques were used to calculate suspended sediment loads by using sediment rating curves.
Materials and Methods: The total drainage area of the Sefidrood watershed is about 560000 km2. In this study uncertainty in suspended sediment rating curves was estimated in four stations including Motorkhane, Miyane Tonel Shomare 7, Stor and Glinak constructed on Ayghdamosh, Ghrangho, GHezelOzan and Shahrod rivers, respectively. Data were randomly divided into a training data set (80 percent) and a test set (20 percent) by Latin hypercube random sampling.Different suspended sediment rating curves equations were fitted to log-transformed values of sediment concentration and discharge and the best fit models were selected based on the lowest root mean square error (RMSE) and the highest correlation of coefficient (R2). In the GLUE methodology, different parameter sets were sampled randomly from priori probability distribution. For each station using sampled parameter sets and selected suspended sediment rating curves equation suspended sediment concentration values were estimated several times (100000 to 400000 times). With respect to likelihood function and certain subjective threshold, parameter sets were divided into behavioral and non-behavioral parameter sets. Finally using behavioral parameter sets the 95% confidence intervals for suspended sediment concentration due to parameter uncertainty were estimated. In bootstrap methodology observed suspended sediment and discharge vectors were resampled with replacement B (set to 3000) times. Sediment rating curves equation was fitted to each sampled suspended sediment and discharge data sets. Using these sediment rating curve and their residual suspended sediment concentration were calculate for test data. Finally using the 2.5 and 97.5 percentile of the B bootstrap realizations, 95% bootstrap prediction intervals were predicted.
Results and Discussion: Results showed that Motorkhane and MiyaneTonelShomare 7 stations were best fitted by a sigmoid function and Stor and Glinak stations were best fitted by second order polynomial and liner function, respectively The first 50 of the B bootstrapped curves were plotted for all stations.with respect to these plots implied that bootstrapped curves more scattered whereas observed data were less. The suspended sediment curve parameters estimated more accurately where, the suspended sediments were sampled more, as a result of reduced uncertainty in estimated suspended sediment concentration due to parameter uncertainty. In addition to sampling density bootstrapped curves, uncertainty depends on the curve shape. For GLUE methodology to assess the impact of threshold values on the uncertainty results, threshold values systematically changed from 0.1 to 0.45. Study results showed that 95% confidence intervals are sensitive to the selected threshold values and higher threshold values will result in an increasing 95% confidence interval. However, the highest 95% confidence intervals obtained by GLUE method (when threshold value was set to 0.1) was little than those values obtained by Bootstrap.
Conclusions: The uncertainty of sediment rating curves was addressed in this study by considering two different procedures based on the GLUE and bootstrap methods for four stations in Sefidrod watershed.Results showed that nonlinear equation fitted log-transformed values of sediment concentration and discharge better than linear equation. Uncertainty result using GLUE depend on chosen threshold values. As threshold values increased, 95% confidence intervals decreased. Uncertainty results showed that 95% confidence intervals estimated by bootstrap were higher than the biggest 95% confidence intervals (when threshold value set to 0.1) estimated by GLUE method. Overall, in all stations, 95% confidence intervals arising from suspended sediment curve shapes (e.g, linear, second order polynomial and sigmoid function), data sampling density and uncertainty estimation methods (here were GLUE and Bootstrap).
N. Zabet Pishkhani; S.M. Seyedian; A. Heshmat Pour; H. Rouhani
Abstract
Introduction: In recent years, according to the intelligent models increased as new techniques and tools in hydrological processes such as precipitation forecasting. ANFIS model has good ability in train, construction and classification, and also has the advantage that allows the extraction of fuzzy ...
Read More
Introduction: In recent years, according to the intelligent models increased as new techniques and tools in hydrological processes such as precipitation forecasting. ANFIS model has good ability in train, construction and classification, and also has the advantage that allows the extraction of fuzzy rules from numerical information or knowledge. Another intelligent technique in recent years has been used in various areas is support vector machine (SVM). In this paper the ability of artificial intelligence methods including support vector machine (SVM) and adaptive neuro fuzzy inference system (ANFIS) were analyzed in monthly precipitation prediction.
Materials and Methods: The study area was the city of Gonbad in Golestan Province. The city has a temperate climate in the southern highlands and southern plains, mountains and temperate humid, semi-arid and semi-arid in the north of Gorganroud river. In total, the city's climate is temperate and humid. In the present study, monthly precipitation was modeled in Gonbad using ANFIS and SVM and two different database structures were designed. The first structure: input layer consisted of mean temperature, relative humidity, pressure and wind speed at Gonbad station. The second structure: According to Pearson coefficient, the monthly precipitation data were used from four stations: Arazkoose, Bahalke, Tamar and Aqqala which had a higher correlation with Gonbad station precipitation. In this study precipitation data was used from 1995 to 2012. 80% data were used for model training and the remaining 20% of data for validation. SVM was developed from support vector machines in the 1990s by Vapnik. SVM has been widely recognized as a powerful tool to deal with function fitting problems. An Adaptive Neuro-Fuzzy Inference System (ANFIS) refers, in general, to an adaptive network which performs the function of a fuzzy inference system. The most commonly used fuzzy system in ANFIS architectures is the Sugeno model since it is less computationally exhaustive and more transparent than other models. A consequent membership function (MF) of the Sugeno model could be any arbitrary parameterized function of the crisp inputs, most like lya polynomial. Zero and first order polynomials were used as consequent MF in constant and linear Sugeno models, respectively. In addition, the defuzzification process in Sugeno fuzzy models is a simple weighted average calculation. The fuzzy space was divided via grid partitioning according to the number of antecedent MF, and each fuzzy region was covered with a fuzzy rule.
Results Discussion: The statistical results showed that in first structure determination coefficient values for both the training and test was not good performance in precipitation prediction so that ANFIS and SVM had determination coefficient of 0.67 and 0.33 in training phase and 0.45 and 0.40 in test phase. Also the error RMSE values showed that both models had failed to predict precipitation in first structure. The results of second structure in precipitation prediction showed that determination coefficient of ANFIS at training and testing was 0.93 and 0.87 respectively and RMSE was 7.06 and 9.28 respectively. MBE values showed that the ANFIS underestimated at training phase and overestimated at test phase. Determination coefficient of SVM at training and testing was 0.89 and 0.91 respectively and RMSE was 9.28 and 5.59 respectively. SVM underestimated precipitation at train phase and overestimated it at test phase. ANFIS and SVM modeled precipitation using precipitation gauging stations with reasonable accuracy. Determining coefficient in the test phase was almost the same for ANFIS and SVM but the RMSE error of SVM model was about 20% lower than the ANFIS. The coefficient of determination and error values indicated SVM had greater accuracy than ANFIS. ANFIS overestimated precipitation for less than 20 mm but for higher values of uniformly distributed around the 1:1. SVM underestimated precipitation for more than 90 mm precipitation due to the low number of data in the training phase, which made this model, did not train well. When meteorological parameters were introduced as input, minimum determination coefficient and maximum error in the test phase occurred while humidity parameters were removed. By removing any of the parameters of temperature, pressure and wind speed the error values and coefficient of determination in test phase was approximately equal.
Conclusion: The potential of the support vector machine (SVM) and neuoro fuzzy inference system (ANFIS) in monthly precipitation pattern were analyzed. In order to model, two data sets were used containing meteorological parameters (temperature, humidity, pressure and wind speed) and the stations precipitation. The results showed that the simulated precipitation using meteorological parameters by ANFIS and SVM had low accuracy. Precipitation forecasting using stations precipitation in the region had good accuracy by ANFIS and SVM. Comparing the results of this study showed the high efficiency of SVM in simulating precipitation. This method can be successfully used in modeling precipitation to increase efficiency of precipitation modelling.