Modeling of estimation of suspended sediment in Gharasu watershed of Ardabil province using artificial neural network models RBF and MLP
Document Type : Original Article
Authors
1
Department of Physical Geography, Faculty of Social Sciences, University of Mohaghegh Ardabili, Iran
2
Ph.D. Student in Physical Geography, University of Mohaghegh Ardabili
10.22034/gmpj.2023.406147.1445
Abstract
IIntroduction
Sediment that consists of solid particles with organic matter transported by water is called suspended sediment. In other words, the sediment load is the flow of the total sediment output from the catchment or drainage basin that can be measured in the desired cross-section and in a certain period of time. In most natural rivers, most of the sediments are transported as suspended load. The sediments collected by the rivers cause many problems, including sedimentation in the reservoirs of dams and reducing their useful volume, changing the course of the river due to sedimentation in their bed, reducing the capacity of canals and water transfer facilities, and changing the quality of water in terms of drinking and agriculture. . However, no direct or indirect experimental model developed to evaluate this process has been universally accepted. Therefore, sediment transport has been considered by engineers from different aspects and different methods have been used to estimate it. One of the methods of estimating suspended sediment is artificial neural network. Artificial neural network is a computing mechanism that is able to provide a series of new information by taking information and calculating it. Considering that the structure of the human brain has a very high ability to process complex, non-linear and parallel information,
Methodology
As one of the sub-basins of the Aras catchment basin, Gharasu catchment is located in the geographical coordinates of 47°31' to 48°47' east longitude and 37°47' to 38°52' north latitude.
In this study, the statistics and information of 17 variables in 13 sub-basins of the Gharasu River, which were extracted by the regional water organization of Ardabil province, were obtained from this organization. In order to model the artificial neural network from the data of 3834 daily sediments recorded in 13 sediment measuring stations in the studied sub-basins during a 50-year statistical period corresponding to the statistical period of 1350 to 1399 and also from the digital topographic maps of the 1:25000 scale of the Geographical Organization of the Armed Forces to Validation of basin demarcation was used. In choosing this common time base, criteria of completeness, sufficient length of data and use of the latest available data were taken into consideration. Then the normality and correlation between the obtained data were evaluated and two methods of Radius Axis Function (RBF) and Multilayer Perceptron (MLP) were used in SPSS software to model the artificial neural network.
Results and Discussion
, recorded suspended sediment (3834 cases) in the relevant statistical period was considered as a dependent variable and flow rate as an independent variable separately for each sub-basin, and Pearson's correlation method was used to check the correlation between the independent variable and the dependent variable.
According to the correlation matrix of the variables, it can be seen that Barouk sub-basin has the highest correlation and Arrab Kendi and Pol Almas sub-basins have the lowest correlation. After modeling the data by two artificial neural network models (RBF and MLP), the amount of sediment for each year was predicted by these models and R2 and RMSE values were also calculated for them. In order to determine the number of neurons in the hidden layer, the values of the neurons in this layer were evaluated by trial and error, and according to the results, choosing the number of 4 neurons for the RBF model and 3 neurons for the MLP model has the highest prediction accuracy in the evaluation data and It also shows in the test data. The accuracy of prediction in RBF model with correlation coefficient R2=0.941 and RMSE=65.002 is compared to MLP model with R2=0.917 and RMSE=88.244.
Based on the scatter diagram between the real data and the estimated data, it was determined that the average of the real values is 4.636, which is 4.367 for the RBF model and 3.534 for the LMP model, which indicates better accuracy in Modeling and the closeness of the RBF model value to the real value. Regarding the median index and the mode index, which represent the most repeated data in the statistical collection, for the real values, the numbers are 4.117 and 3.246, respectively, and for the RBF model, the numbers are 4.425 and 4.213, respectively, which are the closest values. It is considered as a real amount.
Conclusion
So far, various forecasting models have been used to estimate river sedimentation. Some of these models have estimated the amount of sediment by combining different physical parameters of the basin, climate and even the output of satellite images. Artificial neural network models are widely used in forecasting geographic models today.
In this research, two artificial neural network models, radial axis function (RBF) and multi-layer perceptron (MLP) in SPSS software have been used to estimate the sediment of Gharasu River in Ardabil province. In this study, recorded suspended sediment (3834 cases) in a 50-year statistical period was considered as a dependent variable and flow rate as an independent variable separately for each sub-basin, and Pearson's correlation method was used to check the correlation between the independent variable and the dependent variable. It was found that Barouk sub-basin had the highest correlation and Arbab Kendi and Pol Almas sub-basins had the lowest correlation. After modeling the data by artificial neural network model, the amount of sediment for each year was predicted by these models and R2 and RMSE values were also calculated for them. The prediction accuracy of RBF model with correlation coefficient R2=0.941 and RMSE=65.002 is higher than MLP model with R2=0.917 and RMSE=88.244, and it has a better performance in estimating suspended sediment in the study basin.
Also, the average value of the real values is equal to 4.636, which is equal to 4.367 for the RBF model. This research showed that in all studied stations, the RBF method provides more accurate estimates of suspended sediment than the MLP model. Of course, due to the existence of complex relationships between flow rate and suspended sediment, the appropriate model should be determined in each hydrometric station to estimate this variable more accurately,
Sediment that consists of solid particles with organic matter transported by water is called suspended sediment. In other words, the sediment load is the flow of the total sediment output from the catchment or drainage basin that can be measured in the desired cross-section and in a certain period of time. In most natural rivers, most of the sediments are transported as suspended load. The sediments collected by the rivers cause many problems, including sedimentation in the reservoirs of dams and reducing their useful volume, changing the course of the river due to sedimentation in their bed, reducing the capacity of canals and water transfer facilities, and changing the quality of water in terms of drinking and agriculture. . However, no direct or indirect experimental model developed to evaluate this process has been universally accepted. Therefore, sediment transport has been considered by engineers from different aspects and different methods have been used to estimate it. One of the methods of estimating suspended sediment is artificial neural network. Artificial neural network is a computing mechanism that is able to provide a series of new information by taking information and calculating it. Considering that the structure of the human brain has a very high ability to process complex, non-linear and parallel information,
Methodology
As one of the sub-basins of the Aras catchment basin, Gharasu catchment is located in the geographical coordinates of 47°31' to 48°47' east longitude and 37°47' to 38°52' north latitude.
In this study, the statistics and information of 17 variables in 13 sub-basins of the Gharasu River, which were extracted by the regional water organization of Ardabil province, were obtained from this organization. In order to model the artificial neural network from the data of 3834 daily sediments recorded in 13 sediment measuring stations in the studied sub-basins during a 50-year statistical period corresponding to the statistical period of 1350 to 1399 and also from the digital topographic maps of the 1:25000 scale of the Geographical Organization of the Armed Forces to Validation of basin demarcation was used. In choosing this common time base, criteria of completeness, sufficient length of data and use of the latest available data were taken into consideration. Then the normality and correlation between the obtained data were evaluated and two methods of Radius Axis Function (RBF) and Multilayer Perceptron (MLP) were used in SPSS software to model the artificial neural network.
Results and Discussion
, recorded suspended sediment (3834 cases) in the relevant statistical period was considered as a dependent variable and flow rate as an independent variable separately for each sub-basin, and Pearson's correlation method was used to check the correlation between the independent variable and the dependent variable.
According to the correlation matrix of the variables, it can be seen that Barouk sub-basin has the highest correlation and Arrab Kendi and Pol Almas sub-basins have the lowest correlation. After modeling the data by two artificial neural network models (RBF and MLP), the amount of sediment for each year was predicted by these models and R2 and RMSE values were also calculated for them. In order to determine the number of neurons in the hidden layer, the values of the neurons in this layer were evaluated by trial and error, and according to the results, choosing the number of 4 neurons for the RBF model and 3 neurons for the MLP model has the highest prediction accuracy in the evaluation data and It also shows in the test data. The accuracy of prediction in RBF model with correlation coefficient R2=0.941 and RMSE=65.002 is compared to MLP model with R2=0.917 and RMSE=88.244.
Based on the scatter diagram between the real data and the estimated data, it was determined that the average of the real values is 4.636, which is 4.367 for the RBF model and 3.534 for the LMP model, which indicates better accuracy in Modeling and the closeness of the RBF model value to the real value. Regarding the median index and the mode index, which represent the most repeated data in the statistical collection, for the real values, the numbers are 4.117 and 3.246, respectively, and for the RBF model, the numbers are 4.425 and 4.213, respectively, which are the closest values. It is considered as a real amount.
Conclusion
So far, various forecasting models have been used to estimate river sedimentation. Some of these models have estimated the amount of sediment by combining different physical parameters of the basin, climate and even the output of satellite images. Artificial neural network models are widely used in forecasting geographic models today.
In this research, two artificial neural network models, radial axis function (RBF) and multi-layer perceptron (MLP) in SPSS software have been used to estimate the sediment of Gharasu River in Ardabil province. In this study, recorded suspended sediment (3834 cases) in a 50-year statistical period was considered as a dependent variable and flow rate as an independent variable separately for each sub-basin, and Pearson's correlation method was used to check the correlation between the independent variable and the dependent variable. It was found that Barouk sub-basin had the highest correlation and Arbab Kendi and Pol Almas sub-basins had the lowest correlation. After modeling the data by artificial neural network model, the amount of sediment for each year was predicted by these models and R2 and RMSE values were also calculated for them. The prediction accuracy of RBF model with correlation coefficient R2=0.941 and RMSE=65.002 is higher than MLP model with R2=0.917 and RMSE=88.244, and it has a better performance in estimating suspended sediment in the study basin.
Also, the average value of the real values is equal to 4.636, which is equal to 4.367 for the RBF model. This research showed that in all studied stations, the RBF method provides more accurate estimates of suspended sediment than the MLP model. Of course, due to the existence of complex relationships between flow rate and suspended sediment, the appropriate model should be determined in each hydrometric station to estimate this variable more accurately,
Keywords
ابوالفتحی، د.، مددی، ع، و اصغری، ص.، 1397. مدل سازی تخمین میزان رسوب رودخانه به کمک روش شبکه عصبی مصنوعی (نمونه موردی: رودخانه گلرود)، پژوهشهای ژئومورفولوژی کمی، 7(2)، صص 208-196.
ابوالفتحی، د.، مددی، ع، و اصغری، ص.، 1401. مدلسازی تخمین میزان رسوب رودخانه به کمک روش شبکه عصبی مصنوعی (رودخانه ونایی)، تحقیقات کاربردی علوم جغرافیایی، ۲۲(۶۶)، صص 56-41.
اصغری سراسکانرود، ص، مددی، ع، باشکوه، ز، قلعه، ا.، 1400. بررسی الگو و دینامیک رودخانه قرهسو اردبیل محدوده پل سامیان تا سد سبلان. هیدروژئومورفولوژی 8(27)، 35-19.
آوریده، ب.، و شمسی، م.، ۱۳۸۰. کاربرد شبکههای عصبی مصنوعی جهت تخمین دبی رسوب رودخانهها، سومین کنفرانس هیدرولیک، تهران، صص۳۶7 - ۳66.
حجازی، ا، و برومند، ر.، 1393. تحلیلی بر ارزیابی علل فرسایش و برآورد میزان رسوب در حوضههای کوهستانی مناطق نیمه خشک (مورد: حوضه آبریز فاروب رومان نیشابور)، پژوهشهای ژئومورفولوژی کمی، 3(1)، صص 146-128.
دهقانی، ا.، زنگانه، م.، مساعدی، ا، و کوهستانی، ن.، ۱۳۸۸. مقایسه تخمین بار معلق به دو روش منحنی سنجه رسوب و شبکه عصبی مصنوعی، علوم کشاورزی و منابع طبیعی، سال 16، شماره 1، صص 276-266.
دولت کردستانی، م.، نوحه گر، ا، و جانی زاده، س.، 1397. ارزیابی کارایی مدلهای شبکه عصبی-فازی تطبیقی و شبکه عصبی مصنوعی در پیشبینی رسوب معلق روزانه (مطالعه موردی: حوضه آبخیز گرو)، پژوهشهای ژئومورفولوژی کمی، 6(4)، صص 130-120.
طهمورث، م.، احمدی، ح، و عسکری، م.، ۱۳۸8. مقایسه دقت مدلهای شبکه عصبی مصنوعی ژئومورفولوژی(GANNs) و رگرسیونی(RM) در برآورد رسوب طالقان رود، پژوهشهای آبخیزداری، سال 22، شماره 84، صص 27-19.
غلامی، ل.، صادقی، ح.، خالدی درویشان، ع، و تلوری، ع.، 1387. مدلسازی رسوب ناشی از رگبارها با استفاده از متغیرهای باران و رواناب، علوم و صنایع کشاورزی، 22(2)، صص 271-263.
محمدی، ص.، 1398. شبیهسازی بار رسوب معلق با استفاده از روشهای شبکه عصبی مصنوعی، عصبی-فازی و منحنی سنجه رسوب در حوزه آبخیز هلیلرود، مهندسی و مدیریت آبخیز، 11(2)، صص 466-452.
مصفایی، ج.، صالح پورجم، ا، و طباطبائی، م.، 1396. مقایسه کارآیی مدل سنجه رسوب و شبکه عصبی مصنوعی در برآورد بار کف رودخانهها. جغرافیا و پایداری محیط، 7(3)، صص 44-33.
نورانی، ب، سلماسی، ف، و قربانی، م، 1399. مدلسازی و تخمین نیروی بالابرنده وارد بر سدهای وزنی با استفاده از روشهای المتن محدود و شبکه عصبی مصنوعی بهینه شده با الگوریتم وال، مهندسی عمران امیرکبیر، 52(7)، صص 1608-1595.
Abrahart, R.J., See, L.M., Heppenstall, A.J., & White, S.M. (2008). Neural network estimation of suspended sediment: potential pitfalls and future directions, Practical hydroinformatics, 4(16), 139-161.
Boukhrissa, Z.A., Khanchoul, K., Le Bissonnais, Y., & Tourki, M. (2013). Prediction of sediment load by sediment rating curve and neural network (ANN) in El Kebir catchment, Algeria, Journal of Earth System Science, 122(5), 1303-1312.
Cigizoglu, H.K. (2004). Estimation and forecasting of daily suspended sediment data by multilayer perceptrons, Advances in Water Resources, 27(2), 185-195.
Cigizoglu, H.K., & Kisi, Ö. (2006). Methods to improve the neural network performance in suspended sediment estimation, Journal of Hydrology, 317(3-4), 221-238.
Dawson, C.W., Harpham, C., Wilby, R.L., & Chen, Y. (2002). Evaluation of artificial neural network techniques for flow forecasting in the River Yangtze, China, Hydrology and Earth System Sciences Discussions, 6(4), 619-626.
Jongho, K., & Ivanov, Y.V. (2014). On the Non uniqueness of Sediment Yield at the Catchment Scale: The Effects of Soil Antecedent Conditions and Surface Shield, Water Resources Research, 50(2), 1025-1045.
Lee S., Ryu J.H., Lee M.J., & Won J.S. 2006. The Application of artificial neural networks to landslide susceptibility mapping at Janghung, Korea, Mathematical Geology, 38(2), 199-220.
Shadkani, S., Abbaspour, A., Samadianfard, S., Hashemi, S., Mosavi, A., & Band, S. (2021). Comparative study of multilayer perceptron-stochastic gradient descent and gradient boosted trees for predicting daily suspended sediment load: The case study of the Mississippi River, U.S., International Journal of Sediment Research, 36(4), 512-523.
Waga, H., Eicken, H., Light, B., & Fukamachi, Y. (2022). A neural network-based method for satellite-based mapping of sediment-laden sea ice in the Arctic, Remote Sensing of Environment, 270(112861), 1-19.
