Zoning of Susceptible Areas with Debris Flows Using Machine Learning Algorithms (Case study of Tangarah Basin -Golestan Province)

Since the beginning of human civilization, human life has always been threatened by natural disasters. Floods are one of the most severe and dangerous natural events that cause a lot of human and financial losses. Although humans today have realized how they are originated, they are still not able to fully control such natural events (Qanavati et al., 2009). Different definitions of flood and their types have been used in different parts of the world (Chuboklu, 2016; 1). One of these cases is sedimentary currents, which are local hydrological phenomena that occur in small catchments with an area of ​​several square kilometers to several hundred square kilometers with a response time of several hours or less. Therefore, these phenomena are one of the most important natural hazards that have serious casualties and financial losses around the world, especially in mountainous basins (Vilanova et al., 2010) in addition to natural environmental conditions, human activities, and lack of planning. also causes the increase and increase of frequency and volume, as well as financial and human losses caused by deposit flows (Qanavati, 2013). One of the ways that can reduce the damage caused by floods is to determine the areas that produce floods and apply appropriate methods to control them (Ahmadabadi, 1399; 318). These currents are related to torrential rains and are a destructive and common hazard in the study areas. Phenomena such as slipping on a steep slope along a small sloping canal and sediment deposition are of this category (Jacob, Hanger, 2005: 9). Floods of deposit streams can lead to catastrophes that pose a serious threat to the lives and property of individuals and economic development. This doubles the need to assess the vulnerability of areas prone to these currents in the area.

To perform sensitivity analysis and prepare a zoning map of floods with debris flow in the basin, the importance and relationship of each of the factors affecting the debris flow should be evaluated. The most important factors for creating debris flows are the height, slope, and sudden precipitation with a very high flow rate. In this study, according to extensive field studies,10 effective factors in creating debris flow floods were identified and used, including elevation, slope percentage, distance from the stream, drainage density, geology, rainfall, land use, slope direction, soil layer, waterway strength index, and topographic moisture index. To prepare these layers from basic data, 1: 50000 topographic maps and to prepare geological maps and fault maps from 1: 250000 and 1: 100000 scale geological maps published by the Geological Survey of Iran were used. The digital elevation model with a ground separation of 30 meters has been used to prepare factors such as height, slope, slope direction, and slope curvature. Digitization of base maps and functions has been done using ARC GIS 10.5 software and Euclidean distance has been used to prepare lithological factors, distance from the road, and distance from the waterway. The flow rate index indicates the strength of water flow in terms of erosion, which was prepared in the GIS software environment. To prepare the map of total annual rainfall, the hydrometric stations of Strait, Galikesh, Tamar, and Dome of the Nightmare have been used, so that it has been interpolated in ARC GIS 10.5 software using mathematical functions. Land use invoice based on Landsat 7 (2014) satellite images and Google Earth archive, as well as the land use information layer of Golestan province, which had been prepared by the General Department of Natural Resources, had been used. Topographic Humidity Index (TWI) is one of the indicators affecting the occurrence and potential of floods in watersheds (Wooger 2019). To prepare this index, ARC GIS software was used and the TWI index was obtained from Equation One.

For training and experimentation in the artificial neural network model, the data of 605 flood points and 300 non-flood points were used. 0.070 has been used for model training and 0.030 for validation. Usually, the error propagation algorithm with a low learning ratio gives the best answer (Kia, 2010; 229). The number of neurons in the hidden layer was changed between 10 and 20, and by selecting 15 neurons in the hidden layer, the least amount of error was obtained. The final network structure was identified with 10 neurons in the input layer, 15 neurons in the latent layer, and 1 neuron in the output layer, And based on this structure, the final zoning was done. Using the final stage weights related to network training, the whole area, which included 63495 pixels, and each of the pixels had 10 characteristics related to 10 factors influencing the debris flow floods, was included in the network structure. After analyzing this data by the neural network, values ​​between zero and one were obtained based on each of the pixels. Finally, the hazard zoning map was prepared by transferring the weight of the final pixels in the ARC GIS environment into 5 hazard classes: low, very low, medium, high, and very high. In the implementation of the stochastic forest model, the identification of trees is very important in OOB error training and validation. To determine the appropriate number of trees, using the mean square error (MSE) criterion, first, some initial values ​​for the number of trees were determined, and then the model was implemented. By examining the mean square error and the coefficient of determination, the optimal model with the lowest error was designed. The resulting model was obtained with a coefficient of determination of 0.94 and a mean squared error of 0.24 for the training stage. Based on the results of the rocking curve, the area of ​​the curve in the area using the random forest algorithm is estimated to be 0.94.

The results showed that the highest weight with a criterion rank of 2.06 in the random Forest approach and a value of 2.04 in the artificial neural network approach is related to rainfall, which shows the accuracy of the evaluation of the two models in estimating debris flow flood zones. The second factor in creating these movements is the height variable with a weight of 1.68. Since the northern and southeastern parts have the highest points of the region, this factor, along with the rainfall factor, which was in the first category, has been the factor often causing the city to activate the force of gravity and to move the sediment and runoff masses on the slope surfaces. After identifying the core structure in the RF and ANN models and providing the information needed to teach the approaches as well as achieving the promised error, the network prepared to analyze the areas it did not encounter. For this purpose, holding the weights of the final stage related to artificial neural network training with a coefficient of 0.2 along with 15 neurons in the hidden layer of the whole region, which included pixels, was provided to ANN and RF models in order to determine the degree of risk from 0 to 1 for each pixel. By classifying the obtained values ​​and transferring these values ​​to ARC GIS software, the final map obtained was divided into 5 vulnerability classes with intervals of 0.2. Among the proposed solutions, the residential areas of Tangarah villages are formed on parts of cones that are exposed to the flow of materials. These dwellings are formed at the outlet of the basin and should be developed away from the outlet of the basin waterways and should be located in higher areas and very old barracks of the river. Due to the non-resistance capacity of bridges and structures built in this area, most areas are at risk of possible damage.