Comparing the performance of different machine learning models in avalanche risk zoning on Khalkhal-Shahroud road
Document Type : Original Article
Authors
1
Masters graduate, Geomorphology and Environmental Management, Faculty of Social Sciences University of Mohaghegh Ardabili. Iran
2
Professor of Geomorphology, Faculty of Social Sciences, University of Mohaghegh Ardabili, Ardabil, Iran
3
Assistant professor of Geomorphology, Department of Geography, Faculty of Letters and Humanities, Ferdowsi University of Mashhad, Mashhad, Iran
10.22034/gmpj.2024.448966.1493
Abstract
Avalanche is the fast and downward movement of large masses of snow and avalanches can endanger human lives and cause huge financial losses. Therefore, zoning and evaluating areas prone to avalanches is one of the necessities of environmental planning to prevent crises and The reduction of human and financial losses in different regions.
The connection road between Khalkhal and Shahrood sector has attracted special attention due to the development of the communication network in the region and its economic, tourism, and transit importance. This route is located in a mountainous area with special geomorphic and geological conditions, and the occurrence of avalanches in this area has caused many human and financial losses every year. For this reason, the evaluation and zoning of avalanche-prone areas on the Khalkhal Road to the Shahrood section is very important.
One of the most basic stages of conducting any research is collecting data and information. In this research, various data and information have been used, including a geological map of Ardabil province with a scale of 1.250000, from which the information about faults was extracted, topographic map of Khalkhal. With a scale of 1/20000, the road map is extracted from this map, the field data of the avalanche points has been collected through the field survey and GPS device, and the important remote sensing data and information include vegetation, land use, elevation map, slope map, slope direction map It was calculated from the images of Sentinel 2 and ALOS-PALSAR satellites. This research is based on field, analytical, and statistical works. Arc GIS, SPSS Modeler, and ENVI software were used to prepare the layers and implement the research model.
According to field studies and analysis of satellite images to identify factors affecting avalanches on the road from Khalkhal to Shahrood, 8 layers in order: 1- Height 2- Vegetation 3- Slope direction 4- Distance from the fault 5- Distance from the road 6- Snow area 7- Land use 8- Slope was used with the standard raster format and all the layers were collected together after production and a single raster image was created. After finishing the classification process, modeling was done in SPSS Modeler software with 8 input neurons, 8 intermediate neurons and 1 output, and 70% of the data were allocated for model training and 30% of the data for model testing. The type of algorithm of SVM and MLP model is after error propagation. At the same time, this algorithm is very efficient and makes the model learn as well as possible, and the model process is completed when the minimum amount of error is reached.
The results show that in the support vector machine model, the highest weight value for the snow layer is 0.26 and for the slope layer and the distance from the road is 0.18 and 0.15, respectively, which indicates that avalanches and hazards It is more dependent on these variables. Also, in the multi-layer perceptron model, the highest weight value was assigned to the snow area factor with a value of 0.20, followed by the layers of the distance from the road, both values of 0.17 and 0.13.
Also the area of risk classes shows that in the support vector machine model, the most class in terms of risk is 36.34 square kilometers and in the multi-layer perceptron model it is 45.83 square kilometers, which shows that the multi-layer perceptron model is better than the volume vector machine model. A large part of the area is classified as high-risk class. Also, in the support vector machine model, the value of 61.57 square kilometers for the high-risk class and 49.32 square kilometers for the multi-layer perceptron model is classified, which shows that there is a big difference between the area values of the models that the support vector machine model had a realistic performance and was able to correctly identify the areas where there was an avalanche process and the risks arising from it, while the multilayer perceptron model did not have a good performance in the high-risk section. In the medium risk section, the results of the gear mark area values are that the support vector machine model with an area of 23.93 and the multilayer perceptron model with a value of 25.83 square kilometers are classified and there is not much difference between the values of the areas, but in the low risk category, the multilayer perceptron model with The area of 15.73 square kilometers is slightly different from the area of the support vector machine model with an area of 14.87 square kilometers.
According to the results of two models regarding avalanche risk zoning in Khalkhal axis to Shahroud, also according to the mechanism of avalanche and the risks arising from it, it has been found out by field studies that the morphology of the slopes overlooking the road, the slopes of the central part in the area of Eskistan village due to the angle A slope of 20 to 35% and too close to the sacred road with snow and its accumulation during the winter seasons with the least mobility causes avalanches. The occurrence of avalanches in the region is in the form of slides and masses and does not have a powder state, because the length of the range and also the road factor have a great impact on the occurrence of this avalanche model. The cause of this provocation is known to be the avalanche. The results of the support vector machine model with relatively few deviations from the multi-layer perceptron model have high reliability and have been able to identify dangerous areas well. . Also, the overlap of the identified areas with real points is high in both models, but the results related to the identification of avalanche-prone areas in the support vector machine model have better performance than the multilayer perceptron and can identify the areas realistically and carefully. Therefore, both of these models can be used to identify areas prone to avalanches.
The connection road between Khalkhal and Shahrood sector has attracted special attention due to the development of the communication network in the region and its economic, tourism, and transit importance. This route is located in a mountainous area with special geomorphic and geological conditions, and the occurrence of avalanches in this area has caused many human and financial losses every year. For this reason, the evaluation and zoning of avalanche-prone areas on the Khalkhal Road to the Shahrood section is very important.
One of the most basic stages of conducting any research is collecting data and information. In this research, various data and information have been used, including a geological map of Ardabil province with a scale of 1.250000, from which the information about faults was extracted, topographic map of Khalkhal. With a scale of 1/20000, the road map is extracted from this map, the field data of the avalanche points has been collected through the field survey and GPS device, and the important remote sensing data and information include vegetation, land use, elevation map, slope map, slope direction map It was calculated from the images of Sentinel 2 and ALOS-PALSAR satellites. This research is based on field, analytical, and statistical works. Arc GIS, SPSS Modeler, and ENVI software were used to prepare the layers and implement the research model.
According to field studies and analysis of satellite images to identify factors affecting avalanches on the road from Khalkhal to Shahrood, 8 layers in order: 1- Height 2- Vegetation 3- Slope direction 4- Distance from the fault 5- Distance from the road 6- Snow area 7- Land use 8- Slope was used with the standard raster format and all the layers were collected together after production and a single raster image was created. After finishing the classification process, modeling was done in SPSS Modeler software with 8 input neurons, 8 intermediate neurons and 1 output, and 70% of the data were allocated for model training and 30% of the data for model testing. The type of algorithm of SVM and MLP model is after error propagation. At the same time, this algorithm is very efficient and makes the model learn as well as possible, and the model process is completed when the minimum amount of error is reached.
The results show that in the support vector machine model, the highest weight value for the snow layer is 0.26 and for the slope layer and the distance from the road is 0.18 and 0.15, respectively, which indicates that avalanches and hazards It is more dependent on these variables. Also, in the multi-layer perceptron model, the highest weight value was assigned to the snow area factor with a value of 0.20, followed by the layers of the distance from the road, both values of 0.17 and 0.13.
Also the area of risk classes shows that in the support vector machine model, the most class in terms of risk is 36.34 square kilometers and in the multi-layer perceptron model it is 45.83 square kilometers, which shows that the multi-layer perceptron model is better than the volume vector machine model. A large part of the area is classified as high-risk class. Also, in the support vector machine model, the value of 61.57 square kilometers for the high-risk class and 49.32 square kilometers for the multi-layer perceptron model is classified, which shows that there is a big difference between the area values of the models that the support vector machine model had a realistic performance and was able to correctly identify the areas where there was an avalanche process and the risks arising from it, while the multilayer perceptron model did not have a good performance in the high-risk section. In the medium risk section, the results of the gear mark area values are that the support vector machine model with an area of 23.93 and the multilayer perceptron model with a value of 25.83 square kilometers are classified and there is not much difference between the values of the areas, but in the low risk category, the multilayer perceptron model with The area of 15.73 square kilometers is slightly different from the area of the support vector machine model with an area of 14.87 square kilometers.
According to the results of two models regarding avalanche risk zoning in Khalkhal axis to Shahroud, also according to the mechanism of avalanche and the risks arising from it, it has been found out by field studies that the morphology of the slopes overlooking the road, the slopes of the central part in the area of Eskistan village due to the angle A slope of 20 to 35% and too close to the sacred road with snow and its accumulation during the winter seasons with the least mobility causes avalanches. The occurrence of avalanches in the region is in the form of slides and masses and does not have a powder state, because the length of the range and also the road factor have a great impact on the occurrence of this avalanche model. The cause of this provocation is known to be the avalanche. The results of the support vector machine model with relatively few deviations from the multi-layer perceptron model have high reliability and have been able to identify dangerous areas well. . Also, the overlap of the identified areas with real points is high in both models, but the results related to the identification of avalanche-prone areas in the support vector machine model have better performance than the multilayer perceptron and can identify the areas realistically and carefully. Therefore, both of these models can be used to identify areas prone to avalanches.
Keywords
احمدی، و.، مرادی، م.،. 1398. پهنهبندی وقوع بهمن با استفاده از مدل شبکه عصبی مصنوعی (MLP) (مطالعه موردی: کوهستانهای شمالی استان البرز)، نشریه علمی مهندسی نقشهبرداری و اطلاعات مکانی، دوره 10، شماره 3، صص 1-14.
سلیمانی مطلق، م.، طالبی، ع.، اکرامی، م،. 1399. امکان پیشبینی احتمال وقوع بهمن با استفاده از روش نزدیکترین همسایه در محیط نرمافزار GIS (مطالعه موردی: محورهای بهمن خیز استان تهران)، مجله علوم و مهندسی آبخیزداری ایران، دوره 5، شماره 16، صص33-38.
اسدامرجی، م.، نهاوندی، ن.، 1395. الگوی اولویتبندی قطعات جادههای کوهستانی ازلحاظ ریسک ریزش بهمن با بهکارگیری احتمال برخورد وسیله نقلیه. پژوهشنامه حملونقل، دوره 13، شماره 4، صص180-192.
اسدامرجی، م.، صفارزاده، م.، حسنپور، ش.، رسولی، ن.، 1392. ارائهی مدل ارزیابی خطر ریزش بهمن در محورهای کوهستانی (مطالعه موردی: محور کرج-چالوس)، مهندسی عمران مدرس، دوره 13، شماره 5، صص 15-28.
تاج بخش، س.، نیکفال، ا م، 1397. پیش بینی پتانسیل سقوط بهمن با استفاده از یک مدل پیش بینی عددی (مطالعه موردی: منطقه شهرستانک, 26-28 دسامبر 2016)، مجله فیزیک زمین و فضا، دوره 44، شماره 3، صص641-658.
جهاندار، ش.، آقاگل زاده، ع.، کاظمی تبار، ج.، 1398. بازشناسی کور کدهای بلوکی در حضور نویز بالا با استفاده از روش های آماری. علوم و فناوری پدافند نوین (علوم و فناوری های پدافند غیر عامل، دوره 10، شماره 4، صص 373-381.
نصرتی، ک.، کیاشمشکی، س.، حسین زاده، م م.، 1398. ارزیابی و مدل سازی پارامتر های محیطی مؤثر بر وقوع بهمن در محور میگون-شمشک با استفاده از تکنیک های آماری چندمتغیره، اکوهیدرولوژی، دوره 6، شماره 1، صص 89-98.
نیری، ه.، کرمی، م ر.، چاره خواه، ب.، 1397. پهنه بندی گذرگاه های بهمن خیز حوضه سیروان با تلفیق مدل تحلیل سلسله مراتبی و شبکه های عصبی مصنوعی، فضای جغرافیایی، دوره 18، شماره 61، صص، 203-219.
Bartelt, P., Bühler, Y., Christen, M., & Schweizer, J. (2012). Snow avalanch.
Beato Bergua, S., Poblete Piedrabuena, M. Á., & Marino Alfonso, J. L. (2018). Snow avalanche susceptibility in the eastern hillside of the aramo range (Asturian central massif, cantabrian mountains, nw spain). Journal of Maps, 14(2), 373-381.
Bian, Rui, et al. "Snow avalanche susceptibility assessment based on ensemble machine learning model in the central Shaluli Mountain." Frontiers in Earth Science 10 (2022): 880711.
Blagovechshenskiy, V., Medeu, A., Gulyayeva, T., Zhdanov, V., Ranova, S., Kamalbekova, A., & Aldabergen, U. (2023). Application of Artificial Intelligence in the Assessment and Forecast.
Caldera, H. J., & Wirasinghe, S. C. (2022). A universal severity classification for natural disasters. Natural hazards, 111(2), 1533-1573.
Chiroiu, P., Onaca, A. L., Favillier, A., Voiculescu, M., Corona, C., Urdea, P., & Stoffel, M. (2024). Snow avalanche synchronicity derived from a multi-path tree-ring reconstruction in the Făgăraș Mountains (Southern Carpathians, Romania). Quaternary Geochronology, 79, 101474.
Chiroiu, P., Onaca, A. L., Favillier, A., Voiculescu, M., Corona, C., Urdea, P., & Stoffel, M. (2024). Snow avalanche synchronicity derived from a multi-path tree-ring reconstruction in the Făgăraș Mountains (Southern Carpathians, Romania). Quaternary Geochronology, 79, 101474.
Choubin, B., Borji, M., Mosavi, A., Sajedi-Hosseini, F., Singh, V. P., & Shamshirband, S. (2019). Snow avalanche hazard prediction using machine learning methods. Journal of Hydrology, 577, 123929.
Choubin, B.; Borji, M.; Hosseini, F.S.; Mosavi, A.; Dineva, A.A. Mass wasting susceptibility assessment of snow avalanches using machine learning models. Sci. Rep. 2020, 10, 18363.
Christen, M., Bühler, Y., Bartelt, P., Leine, R., Glover, J., Schweizer, A., ... & Volkwein, A. (2012, April). Integral hazard management using a unified software environment. In 12th Congress Interpraevent (pp. 77-86).
Devishri, K., Thombson, C., Rajkumar, K. S., Maisnam, M., & Pallipad, J. (2024). Investigating snow avalanche in the Sikkim Himalaya through coherence and displacement mapping of Sentinel-1A SAR data.
Föhn, P. M. B. (1988, October). Snow cover stability tests and the areal variability of snow strength. In Proceedings International Snow Science Workshop, Whistler, British Columbia, Canada pp. 12-15.
Haegeli, P., McClung, D. M., & Schaerer, P. (2020). The avalanche handbook. The Mountaineers Books.
Hagan, M. T., Demuth, H. B., Beale, M. H., and Jesús, O. D., 2014- Neural Network Design, Martin Hagan (2 edition), 1012 pp. ISBN-13:A, Volume 113, Number 6.
Joachims, T. (1998). Making large scale SVM learning practical. Technical reports.
Kumar, S., Srivastava, P. K., & Bhatiya, S. (2019). Geospatial probabilistic modelling for release area mapping of snow avalanches. Cold Regions Science and Technology, 165, 102813.
Lee S, Ryu J-H, Lee M-J, Won J-S. 2006. The application of artificial neural networks to landslide susceptibility mapping at Janghung, Korea. Mathematical Geology, 38(2): 199,220.
Mao, K., Zhu, J., Su, L., Cai, G., Li, Y., & Dong, Z. (2023, June). FinalMLP: an enhanced two-stream MLP model for CTR prediction. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 37, No. 4, pp. 4552-4560).
McClung, D. M., & Schaerer, P. (2006). The avalanche handbook. The Mountaineers Books.
Padeli, P., Martono, A., & Sudaryono, S. (2023). Application of Naive Bayes Model, SVM and Deep Learning Predicting. CICES.
Phoon, K.K.; Zhang, W. (2023). Future of machine learning in geotechnics. Georisk-Assess. Manag. Risk Eng. Syst. Geohazards 2023, 17, 7–22.
Tyrrell, E. (2024). Avalanche Forecasting with Geospatial Analysis in Southwestern Colorado.
Varol, N. (2022). Avalanche susceptibility mapping with the use of frequency ratio, fuzzy and classical analytical hierarchy process for Uzungol area, Turkey. Cold Regions Science and Technology, 194, 103439.
Yom, E. (2004). DR AF T A parallel training algorithm for large scale support vector machines.
