عنوان مقاله [English]
Geomorphological hazards such as mass movements are one of the potentially harmful phenomena. Landslides as one of the most traditional earth movements involved all slope failures in which a mass of materials (soil and rock) moves down the slope due to decreasing of safety factor and overcoming of the destructive forces on the resistance forces on a slope.
Kurdistan province has been frequently exposed to mass movement hazards owing to its characteristics of topography, geomorphology, pedology, geology and climatology. Therefore, the objective of this study is 1) to prepare the spatial prediction map of mass movements for the Kurdistan province to manage the areas prone to the hazards and to use the map in the land use planning projects and development of the rural regions 2) to introduce the Bayesian logistic regression (BLR) ensemble and to compare its efficiency with the logistic regression (LR), and 3) to identify the most significant conditioning factors on mass movements occurrence in the Kurdistan province.
Based on the literature review, 18 factors affecting landslide susceptibility were selected for modeling including slope degree, slope aspect, elevation above sea, curvature, profile curvature, plan curvature, stream power index, topographic wetness index, length-angle of slope, lithology, rainfall, land use, distance to fault, fault density, distance to stream, stream density, distance to road, and road density. A total of 895 mass movements in the Kurdistan province were divided into 70 % (626 locations) for modelling using training dataset and 30% (269 locations) to evaluate based on the validation dataset. Additionally, 626 locations were randomly selected as areas where mass movements have not occurred and they were classified into 70% and 30% for training and validation in modelling process. Based on the Information Gain Ratio index to modelling process, 7 factors were then applied to the LR method and 11 factors were used in the BLR method. The evaluation of models were performed using Specificity, Sensitivity, Accuracy, the area under the Receiver Operating Characteristic Curve (AUROC), and the Root Mean Square of Errors (RMSE) indices.
Result and discussion
According to the results of spatial prediction of mass movements by quantile approach, about 8.06% of the Kurdistan province area is very sensitive to mass movement occurrence using the LR model; while, the BLR model shows a sensitive area of about 12.46%.
The majority of very sensitive and sensitive areas are located to the west of the province. These regions include Oramanat mountainous areas, Kosalan and Chelchama highlands, the Salvatabad Saddle (East of Sanandaj), the Morvarid Saddle (between Sanandaj and Kamyaran), The Khan Saddle (between Saghez and Baneh), and the Arez Saddle (between Sanandaj and Marivan).
In order to assess the spatial map accuracy of mass movement sensitive areas prediction, both training and validating datasets were used. The results showed that the area under the curve of SRC is 0.714 in LR based on training data, indicating that this approach has a potential of 71.4 percent to predict sensitive areas to mass movement. BLR showed a value of 0.672 using the same data which implies a potential of 67.2% for prediction of sensitive areas.
Based on the validating dataset, the area under the curve of PRC was 0.732 and 0.729 for LR and BLR, respectively. These results confirmed the accuracy of the maps prepared by both models; however, it should be noted that LR was slightly providing better results.
To analyze the probability of mass movement occurrence, the potential of the LR and BLR models were evaluated using Friedman test at 5% significant level. The test revealed that there is no significant difference between these two models and both are reasonably able to evaluate spatial prediction of mass movements.
The results showed that there is no significant difference (95% level of significance) between the efficiency and the susceptibility maps prepared by the two models for spatial predicting of landslide in the study area; therefore, the BLR model can be applied as an index model for the study area and other similar areas.
Therefore the hybrid model of Bayesian Logistic Regression has a high capability of identifying sensitive areas to mass movement such that it can be compared with previously successfully testes methods such as artificial neural network, fuzzy logic, decision tree algorithms (Naïve, Random Forest, Baes Forest, …).