Evaluation of the role of input data type on the accuracy of snow avalanche hazard map with Shannon's entropy data-driven approach

Today, despite all the activities that have been done for the residents of mountainous areas, facing a phenomenon called avalanche in mountainous areas is still a normal thing. For years, the people of mountainous areas have to deal with this natural threat. The manner and place of residence and the type of construction, forestry and land use plans in recent decades have minimized the risk of avalanches in mountainous areas. During the occurrence of natural crises that are associated with social anomalies, an efficient management plan must be prepared and adjusted in advance to firstly identify high-risk areas and secondly, the shortest and most reliable ways to access the centers and the best decision can be made in setting up relief centers. In Iran, studies on the phenomenon of snow and snow avalanches are more limited than other studies. In this study, the Absafid waterfall area of Aligudarz has been investigated in order to prepare an avalanche zoning map and the factors affecting the occurrence of this hazard.

In this research, the input data to the model was prepared first. The input data included sampling points that were from avalanche prone areas (dependent variable of the model) which were identified in order to identify avalanche prone spots through field visits, local inquiries and using the information of natural resources experts in the region. Of the 4 sampling methods used, the first method is to select points that were in the area of rocky outcrops (model 1), the second method is to select points that have slopes of more than 45 degrees (model 2), the third method is to select random points in the areas of snow accumulation (model 3), The fourth method of selecting points in a grid with a regular distance of 150 meters from each other in the area of snow accumulation (model 4). And the independent variables of the model, which include indices (Slope, Aspect, LS , TWI , Wind effect, VRM , MBI , profile and general curvature , NDVI ) were used for modeling.

AUC index values for each of the 4 models, respectively, in training and test mode for: the first model) 0.808 and 0.75, the second model) 0.885 and 0.868, the third model) 0.875 and 0.881, the fourth model) 0.947 and 0.94 and the results show that all models are in the category of excellent models. The results of the Jackknife statistics analysis show that in model 1, NDVI and Slope indices have the greatest impact, model number 2 relates to Slope and TWI indices, model 3 relates to NDVI and Slope indices, and in model number 4, NDVI and Slope indices have had the greatest impact on the output of the model. The results of matching the zoning maps with each other and the percentage difference between the maps obtained from the 4 models show that the lowest matching rate among the above maps is the result of the comparison between model 1 and 2, which is 53% and also the highest The compliance rate is related to model 3 and 4, and their lack of compliance is 9%. Also, comparing the results with field observations shows that model number 3 has predicted the best result in identifying avalanche spots in the evaluated area.

In recent decades, the use of different modeling algorithms has been developed. On the other hand, in all data-driven models, in order to initially train the model, it is necessary to introduce the points related to the place of occurrence of the desired phenomenon. Therefore, the type of distribution of input data can influence the output maps. The most important goal of this research is to evaluate the type of input data introduced in a data-oriented method (based on Shannon entropy and logistic regression, Max.Ent) in preparing a snow avalanche risk zoning map in the limits of the White Water waterfall in Aliguderz city. In order to prepare an avalanche risk zoning map, 10 geomorphometric factors were used, and based on the field survey, the area of avalanche occurrence was identified in the region, and snow avalanche prone points were introduced to the model in four different ways. The results of prediction models' accuracy (AUC), determining the importance and degree of sensitivity of each of the criteria used showed that the accuracy of the predicted maps varied between 0.81 and 0.95. Also, from the results of the most effective environmental indicators in the output of each of the models, it was evident that, in general, the criteria related to vegetation and slope were given the most weight. The results of the spatial evaluation of the prepared maps indicate a difference of 53%-9% between the avalanche prone areas among the four introduced methods. In general, the northwest to southwest regions are the most sensitive to the onset of avalanches in the study area, and this issue is evident in almost all models.