The use of quantitative drainage network in the placement of check dam based on machine learning

Background and Objectives: The loss of life and property damage caused by floods and the overflowing of water storage dams with sediment are causes for concern, especially in recent years due to the unsustainable use of natural resources. In order to address these issues, soil and water conservation measures have been implemented in the form of various watershed management programs, including structural (biomechanical and mechanical), management, and biological measures, at the national level. The implementation of watershed management plans, as an infrastructure priority, plays an important role in reducing the damage caused by runoff and floods. One of the most important goals of implementing check dam along rivers, especially upstream of residential areas and watersheds leading to cities and villages, is flood and sediment control. The most important step in implementing mechanical plans as check dam measures is to correctly identify the locations needed to implement these plans. The correct location of check dam has a great impact on reducing the cost of watershed management activities and increasing their effectiveness. In this study, a number of machine learning-based approaches, including artificial neural network (ANN) and maximum entropy (ME) models, were used with the aim of determining the best machine learning model for locating check dam in the Quein watershed.
Materials and Methods: The Quein watershed is located in Alborz Province, north of Taleghan, and lies between the geographical coordinates (50°46' to 50°56' E and 36°10' to 36°18' N). In the present research, 11 effective indicators in the selection of check dam locations, including topographic, hydrological, geological, land use, and economic factors, were used to determine the appropriate areas for the construction of watershed structures. The points of the existing check dam were divided into two groups modeling data (training) and validation data (test) using a random method. The digital elevation model layer was prepared using elevation points and contour lines with a resolution of 10x10 meters. The slope layer was prepared using the digital elevation model and the Slope function. The distance from the watercourse and watercourse density layers were prepared based on the watercourse map (extracted from SAGA GIS software) using the Euclidean distance and Line Density functions in ArcGIS software, respectively. The lithology layer was considered an important factor in the spatial and temporal changes in drainage, permeability, hydrology, and sediment production of the watershed and was extracted from the geological map. The land use layer was obtained from the General Directorate of Natural Resources and Watershed Management of Alborz Province. The rainfall map was prepared using data from 15 rain gauge and meteorological stations over a statistical period of 18 years in the form of a rainfall gradient. The SPI factors, watercourse rank, and flow accumulation were prepared using the digital elevation model map in SAGA GIS software. The road density layer was prepared based on the road map using the Line Density function in ArcGIS software. In this study, two models maximum entropy and artificial neural network were used. Preprocessing of input factors was performed to check for the absence of multicollinearity using the variance inflation factor (VIF) and the tolerance index. The importance of each factor in explaining the model was determined using the maximum entropy model and the jackknife diagram, performed in MaxEnt software. The efficiency of the models in the training and validation stages was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC).
Results: The results indicated that there was no collinearity between the factors; therefore, all factors were used in the modeling process. The results showed that the factors of distance from the watercourse, watercourse rank, flow accumulation, elevation, and average rainfall were the most important factors affecting the placement of check dam and were effective in predicting areas with potential for structure construction. The prediction accuracy of the maximum entropy model was excellent in both the training (0.994) and validation (0.993) phases. The prediction accuracy of the artificial neural network model was also excellent in both the training (1) and validation (1) phases. Considering the field reality in the Quein watershed, it seems that the artificial neural network model was over fitted, making the final potential map for check dam subject to errors, whereas the results of the maximum entropy model provided more reasonable outputs. Field visits to verify the model results showed that in all the streams studied, the maximum entropy model correctly and with great accuracy identified critical streams in terms of flood and sediment load. A total of 47 km of critical streams were identified in terms of flood and sediment load. Therefore, during the subsequent visits, 22 check dam were installed in the studied streams.
Conclusion: The results of field reconnaissance showed that in all the streams studied, the maximum entropy model correctly and accurately identified critical streams in terms of flood and sediment load. This indicates the high capability of machine learning-based methods to integrate and analyze complex spatial data, thus increasing accuracy in locating check dam measures. The results of this study emphasize that using machine learning-based methods can significantly improve the accuracy of locating check dam. Therefore, it is suggested that future studies combine machine learning models with optimization and uncertainty analysis methods to further improve location accuracy. This research represents an important step toward using data-driven models to optimize check dam measures and can serve as a model for similar studies in other watersheds. This method not only increases the accuracy of decision-making but also reduces implementation costs and improves the efficiency of watershed management projects.