Document Type : Research Article
Authors
1
Department of Remote Sensing and GIS, Faculty of Earth Sciences, University of Shahid Beheshti, Tehran, Iran
2
Department of Remote Sensing and GIS, Faculty of Geography, University of Tehran, Tehran, Iran
3
Department of Physical Geography, Faculty of Earth Sciences, University of Shahid Beheshti, Tehran, Iran
4
Department of Remote Sensing, Faculty of Environmental Sciences, University of Tabriz, Tabriz, Iran
Abstract
Introduction:
Landslides are among the most frequent and significant natural hazards worldwide. During this geomorphic process, soil and surface materials become unstable and move downslope under the influence of gravity. This instability is often triggered by factors such as heavy or prolonged rainfall, earthquakes, water saturation of the soil, and human activities including vegetation removal and improper engineering practices.This issue becomes particularly critical along transportation and communication corridors, such as roads, railways, power transmission lines, and energy pipelines. The occurrence of landslides along these routes can lead to traffic disruptions lasting days or even weeks, human casualties, damage to vehicles and equipment, and severe destruction of infrastructure, including asphalt surfaces, retaining walls, bridges, and technical structures. Additionally, landslides can delay emergency response efforts and significantly increase maintenance and repair costs.
Material and Methods:
In this study, aimed at preparing a landslide susceptibility map for the Saqqez–Marivan corridor, a comprehensive set of spatial data—including raster layers and landslide point data—was first collected and integrated. The raster layers represented factors commonly used in landslide susceptibility analyses, such as elevation, slope, aspect, curvature, profile and plan curvature, distance to watercourses, fault distance and density, drainage density, precipitation, and other morphometric variables. Before processing, all raster layers were standardized in terms of coordinate system, resolution, dimensions, and spatial extent. The elevation layer was selected as the reference raster to ensure that all sampling and analytical operations were performed consistently based on its coordinate grid and cell structure. In addition to raster data, a vector layer containing landslide and non-landslide points was used, with each point assigned a target class value (1 for landslide, 0 for non-landslide). To ensure accurate sampling from the raster layers, the coordinate system of this vector dataset was transformed to match that of the rasters, after which the corresponding values of all raster layers were extracted for each point. This sampling process was performed using Rasterio functions, resulting in a complete feature vector for every point. Following feature extraction, the dataset was examined for invalid values, including NaN, Inf, and out-of-range values. Rows containing missing or invalid data were identified and removed during the initial cleaning stage. For remaining missing values that did not require row deletion, the mean of the corresponding training-set column was used for imputation. This approach prevented information leakage from the test set into the model and ensured the integrity of the modeling process.After confirming that the feature table was complete, all variables were standardized to have a mean of zero and a variance of one, due to the substantial differences in measurement scales among the predictors. Standardization facilitated neural network convergence and prevented disproportionate influence of variables with larger numeric ranges.The processed dataset was then divided into three subsets to enable proper model training and reliable performance evaluation. Seventy percent of the data was allocated to the training set, while the remaining 30% was split equally between the validation and test sets, each comprising 15% of the total data. The training set was used to learn model parameters, the validation set to tune model architecture and control overfitting, and the test set to provide an independent and unbiased assessment of model performance.
Results and Discussion:
The results of this study demonstrate that the Artificial Neural Network (ANN) model possesses a strong capability for identifying and mapping landslide susceptibility along the Marivan–Saqqez corridor. The model successfully predicts the spatial distribution pattern of landslides with considerable accuracy. Analysis of the contributing variables indicates that landslide occurrence in the region is governed by a combination of environmental and geomorphological factors, and its spatial distribution is highly structured and dependent on natural conditions.
According to the findings, the highest concentration of landslides occurs in the southern parts of the study area zones that exhibit greater susceptibility due to higher rainfall, proximity to faults and rivers, and specific topographic and morphological characteristics. In contrast, the northern parts of the region show the lowest landslide frequency, reflecting more stable environmental conditions.The assessment of variable importance further revealed that annual rainfall is the most influential factor, playing a decisive role in increasing landslide susceptibility. This is followed by distance to faults, elevation, slope, and other topographic parameters. These results highlight the dominant role of hydrological and structural processes in driving slope instability within the region.Morphological analysis also showed that slopes with intermediate elevations, gentle to moderate gradients, and southern to southwestern aspects exhibit the greatest overlap with landslide occurrences. Slopes with near‑zero surface curvature or positive cross‑sectional curvature were also more susceptible—conditions that promote surface flow concentration and increased soil moisture, ultimately reducing slope stability. In terms of model performance, the statistical indicators obtained from the training, validation, and testing phases confirm the high efficiency of the ANN model in landslide prediction. Elevated values of metrics such as AUC, Accuracy, F1‑score, and Cohen’s Kappa indicate not only strong predictive accuracy but also model stability and generalizability. The model effectively reconstructed the actual landslide pattern without significant overfitting.
Conclusion:
The test phase modeling results indicate that the developed ANN model achieved an overall accuracy of 92.5%, a Kappa coefficient of 0.85, and an AUC value of 0.9905. These metrics collectively confirm the model’s exceptional capability in discriminating landslide from non landslide areas and demonstrate the absence of overfitting.The feature importance analysis, conducted using the weighted neural network approach, revealed that mean annual rainfall (importance coefficient = 0.55) constitutes the most influential factor governing landslide occurrence in the study area. This variable is followed, in descending order of importance, by distance to faults, elevation, fault density, and slope. These findings underscore the dominant role of hydrological and structural controls in shaping the spatial distribution of landslides.The final landslide susceptibility map was classified into five hazard categories very low, low, moderate, high, and very high. Spatial analysis of this map demonstrated that the high and very high susceptibility zones are predominantly concentrated in the southern portion of the Marivan–Saqqez corridor. These areas correspond closely with regions characterized by steep slopes, elevated drainage density, proximity to major fault structures, and higher precipitation levels. Conversely, the northern parts of the region fall within the low to very low hazard classes, primarily due to their greater distance from fault lines, lower drainage density, and reduced rainfall.A comparison of the present findings with previous research (including studies by Rezaei Moghadam et al., Rakei et al., and Pradhan et al.) reveals a high degree of consistency between the ANN based results and field observations, as well as a relative advantage of the ANN model over traditional statistical techniques such as logistic regression and frequency ratio.
Keywords
Subjects
