Landslide susceptibility predicting using the maximum entropy machine learning algorithm (Bar catchment of Nishapur)

Introduction
Landslides are one of the most destructive natural disasters. Landslides cause damage to a variety of engineering structures, residential areas, vital arteries such as roads, water and gas pipelines, power lines, forests and pastures, agricultural lands, sediments and muddy floods. When landslides occur not much work can be done, but damage can be reduced with pre-determined planning. Therefore, it is necessary to prepare a comprehensive landslide susceptibility map to reduce possible damage to people and infrastructure. In recent years, the use of machine learning methods and geographic information systems (GIS) in landslide sensitivity zoning has expanded and landslide sensitivity maps are prepared with acceptable accuracy. The purpose of this study is to prioritize the factors affecting the occurrence of landslides and zoning the sensitivity of its occurrence using the maximum entropy method in the Neishabour watershed, located in Khorasan Razavi province.
Materials and methods
Bar watershed is one of the sub-basins of Bar River, which is located in the north of Neishabour city. This basin is limited from the northeast, east and south to the heights of Lar limestone ridge, from the west to Jurassic shales and from the northwest to the marls of Delichai Formation. In this study, landslide zoning was performed using the maximum entropy model. First, the landslides in the basin were extracted through the landslide map in the General Department of Natural Resources of Khorasan Razavi Province. Then, by conducting field visits in the basin, using local information and GPS devices, as well as using Google Earth satellite images, this map was corrected, and finally, the landslide inventory map was prepared as a point. The most important factors affecting landslides in the region include slope, slope direction, elevation classes, geology, drainage network (distance from river), road (distance from road), fault (distance from fault), topographic indexes (Steam Power Index (SPI)), Topographic Wetness Index (TWI) and Slope Length index (LS)), geomorphological indexes (Topographic Position Index (TPI)), Topographic Roughness Index (TRI) and Curvature Index, land use, Normalized Vegetation Difference Index (NDVI) Co-precipitation lines were identified and analyzed as effective factors in landslide occurrence in the study area. In order to model the sensitivity of landslides, 70% of landslides were randomly selected for training (calibration) of the model and 30% for validation of model results. The maximum entropy model was modeled as a dependent model of the presence of potential regions in MaxEnt software. The final landslide susceptibility map was zoned in five talent classes (very low, low, medium, high and very high) based on the turning points of the cumulative frequency of the pixels. In order to evaluate the results of the model, ROC curve was used using two sets of training and validation data.
Results and discussion
Using information provided by the General Department of Natural Resources of Khorasan Razavi Province, extensive field studies, and the use of Google Earth images, a total of 74 landslides were identified in the Bar Watershed. The sensitivity of landslides has decreased with increasing altitude up to 1650 meters and has increased from 1650 to 3200 meters. The highest susceptibility to landslides is in the northern and eastern slopes of the study area. The results of the slope study showed that with increasing the slope to 40 degrees, the sensitivity to landslides has increased and after 40 degrees the sensitivity has decreased. Landslide susceptibility has decreased with increasing distance from the fault. Regarding the effect of distance from the road on the occurrence of landslides, with increasing distance from the road to a distance of 4000 meters from the road, there is a decreasing sensitivity trend, but after this distance, this trend has increased, which is probably due to other factors. Among the different uses of the study area, weak pasture uses and rocky outcrops have the highest susceptibility to landslides, respectively. Q2 and PLQm rock units have the greatest impact on landslide susceptibility. With increasing TPI index, the sensitivity of landslide is also increasing. With increasing TRI index, the sensitivity of landslide occurrence has decreased. Slip sensitivity has also increased with increasing slope length. With increasing Stream Power Index (SPI), landslide sensitivity has increased and then decreased. According to the results of Jackknife diagram, among the selected parameters in the modeling process, the slope length (LS) layers, respectively, slope direction and slope have the most participation and impact in the occurrence of landslide landslides. According to the results, 13% of the area has high and very high sensitivity, 73% has low and very low sensitivity and 14% of the area has medium sensitivity. The area under the curve (AUC) based on the relative performance detection curve indicates excellent accuracy (AUC = 0.92) in the training phase and very good (AUC = 0.87) in the validation phase. Based on the results of the maximum entropy model, about 13% of the Neishabour load field is located in the high and very high sensitivity zones.
Conclusion
Distribution factor analysis showed that the slope length (LS) factors for slope and slope direction have the highest participation and impact in the occurrence of landslide landslides and distance from the fault, NDVI index and TWI index have the least impact on landslide occurrence. Interpretation of the results of ROC curve mapping showed that the accuracy of the model in estimating sensitive areas in both training and validation stages was excellent and very good, respectively, which according to the results of Phillips et al. (2006) means excellent performance of the model. is. Based on the obtained results, it can be said that the maximum entropy model has high speed and accuracy in determining landslide-sensitive areas due to the unlimited ability of the maximum entropy algorithm to measure complex linear and other linear relationships. Identifying sensitive areas, using this model can be considered as a solution. The results of this study also showed that the maximum entropy model is a promising approach for modeling landslide sensitivity. Because the plan has a high accuracy in identifying and separating landslide sensitive areas, decision makers and engineers to introduce areas with different landslide sensitivities in order to build a suitable place to prevent the destruction of sediment structures, slope management, drainage and transfer Water from sensitive areas close to the implementation of structures, road network development and land use planning programs will help. The resultant landslide susceptibility maps can be useful in appropriate watershed management practices and for sustainable development in the region.