Modeling the Erosion and Sedimentation Rate of Sojasrood River before and after Construction of Golaber Dam with Machine Learning Algorithms

Dams are one of the most important human structures along rivers that are constructed with the aim of generating electricity, flood control, and providing water for agriculture and urban centers. Today, very few large and small rivers remain uncontrolled; what is important geomorphologically is the changes that occur in the performance of erosion processes downstream of the river after dam construction. These changes are not limited to after dam construction, but these morphological changes are the result of changes in the performance of erosion and deposition processes in drainage basins, completely transforming the face of the basin and canals in the area close to the constructed dam. Analyzing the geomorphological changes of rivers due to the creation of human facilities such as dams is one of the most important tasks of geomorphologists. This study is done with different models and methods. Among the common methods of the last decade are the GCD model and machine learning. The GCD method is the result of subtracting two digital elevation models at different times, which are produced by different methods of these dams. To analyze the geomorphological changes caused by the construction of the dams in downstream of the river, in addition to using historical dams, machine learning methods can be used for more accurate modeling by involving a variety of effective maps in detecting changes. The main purpose of this study is to apply the machine learning method using the data obtained from the GCD model to generate regression maps due to the impact of dam operation downstream of the river.

This study was carried out in the Sojasrood River and downstream of Golaber dam. Software and tools used in this research included the following: Arcgis, envi, sagagis software, R software, GCD software and extension, Google Earth Engine system, AutoCAD software, Excel, topographic maps of 1/50000, Elevation digital model and Garmin GPS. Machine learning methods were used to evaluate the effects of Golaber Dam in the period before and after the construction of the dam. In order to access the data required for this research, digital elevation models of stereo pair images of L1A series and L1B satellite ester were used as time series. First, through the GCD model, the volume changes of erosion and sediment downstream of the dam were calculated. Then, the data obtained from this model were used as a target variable along with nine layers of geomorphometry and precipitation and runoff as predictive data to implement machine learning algorithms in three methods of multiple linear regression, decision tree, and random forest. 70% of the data were used for modeling and 30% of the data were used for evaluation in R programming software.

Given that continuous data on erosion and sedimentation rates from the GCD model have been used for the machine learning method, naturally, a regression method (prediction) should be used for the output of the machine learning models. Three steps were taken to achieve the result of machine learning. First, the models were run one by one in R software and evaluated with 30% of the experimental data, and finally, the model maps and their correlation coefficient and RMSE error were calculated. Comparison of the output results of multiple linear regression models and decision tree and random forest showed differences in statistical data and time series maps before and after dam construction. Therefore, the output maps of the models before and after the construction of the dam were also different from each other. The main reason for this is a significant reduction in the runoff, land-use changes, increased vegetation of the bed and riverside, which has led to changes in the independent variable research data in the period after the construction of the dam. Although statistically in multiple linear regression, the p-value was less than 0.05, the output of maps of this model were associated with a large error. And the model did not predict the rate of erosion and sediment well. In the multiple linear regression model, the correlation coefficient of the map before the construction of the dam was higher than the period after the construction of the dam. CART method was used for decision tree modeling. The map produced by this method with a correlation coefficient above 0.6 showed better performance compared to the multiple linear regression model. The best method for modeling erosion and sedimentation rates in both periods was the random forest method. This model with a correlation coefficient above 0.7 provided the most accurate prediction in this study.

Various methods and models have been proposed to estimate the rate of erosion and sediment in rivers. In recent years, new and more accurate models have been formed, especially in the analysis of the time series of river developments. New methods include the GCD model and machine learning (ML). In this study, in order to observe the changes in erosion and sedimentation rate due to the construction of Golaber Dam in the period before and after its construction, first, the GCD model of volumetric erosion and sedimentation changes was estimated through the model of published errors and multiple time series maps were produced. Then, from the data obtained from this model, along with maps and geomorphometric layers and maximum rainfall and runoff data, in order to more accurately predict the impact of the dam on the riverbed in terms of erosion and sedimentation rates and geomorphological changes of the river, machine learning method was used. The results of modeling showed that dam utilization was strongly effective in erosion and sedimentation of river bed and Random Forest algorithm with a correlation coefficient above 70% and RMSE less than the other two models showed the best prediction for both periods before and after dam construction. The maps produced by the decision tree method also modeled the erosion and sedimentation process in the riverbed in both time series analyses well, but the output of the linear regression model was not accurate enough. For an overview of machine learning algorithms, in addition to evaluating the experimental data of the models themselves, the overall average results of some morphometric indices of the river such as number of meanders, center angle, channel length, and Sinuosity index were also evaluated. This comparison showed the accuracy of modeling decision tree and random forest algorithms applied in the present study.