مدل‌سازی مکانی و پهنه‌بندی خطر فرونشست زمین در استان چهارمحال‌ و بختیاری با استفاده از الگوریتم SOM و تحلیل عملکرد مدل در شناسایی دشت‌های بحرانی

Introduction and Goal
Land subsidence, as one of the silent environmental hazards, poses a serious threat to infrastructure, water and soil resources, and food security, and its occurrence is exacerbated in arid and semi-arid regions due to the over-extraction of groundwater resources. Chaharmahal and Bakhtiari Province, with its strategic location as the roof of Iran and the source of the Karun and Zayandeh-Rud rivers, plays a vital role in supplying the country's water and food. However, in recent years, the emergence of numerous fissures in important plains such as Shahrekord, Borujen, Faradonbeh, and Saman has sounded a serious alarm for this region. Despite scattered reports of subsidence occurrence, the lack of a comprehensive and systematic study that investigates the dimensions of this crisis using modern technologies is quite evident. Therefore, the present study was conducted with the aim of spatial zoning of land subsidence hazard in Chaharmahal and Bakhtiari Province using the capabilities of the Self-Organizing Map (SOM) machine learning model and satellite imagery. This research seeks to identify critical areas and the factors influencing this phenomenon, ultimately providing accurate and practical maps for risk management and the adoption of preventive measures by planners and executive managers.
Materials and Methods
SOM Model: SOM is an artificial neural network that uses unsupervised learning. It takes complex, high-dimensional data and projects it onto a two-dimensional map, placing similar data points close together and enabling the discovery of hidden patterns. Data: For modeling, 30 primary variables affecting subsidence were initially considered. After analyzing the correlation matrix and removing variables with high correlation (greater than 0.7), 23 final variables were selected for model implementation. These variables included elevation, slope, aspect, distance to fault, distance to river, distance to road, geology, land use, topographic indices (such as TWI, TRI, TPI), climatic data (temperature, precipitation), and vegetation. Data on subsidence points were collected using GPS data from the Sanat Maadan va Tejarat organization and field surveys around irrigation canals, power poles, and concrete structures. Model Training and Evaluation: The data were randomly divided into training and testing sets. The model was trained with 2,352 points (including 1,859 subsidence points and 493 non-subsidence points) and evaluated with 772 points (including 536 subsidence points and 236 non-subsidence points). Model performance was assessed using the Area Under the Curve (AUC), Precision, and Recall indices.
Results and Discussion
Model Performance: The evaluation of the SOM model demonstrated its good capability. The AUC value was 0.7685, indicating the model's suitable ability to distinguish areas susceptible to subsidence. The Precision index was calculated as 0.70 (acceptable), and the Recall index was very high at 0.99, showing that the model successfully identified the subsidence points. Hazard Zonation: Based on the model's output, a subsidence hazard zonation map was produced with five classes (Very Low, Low, Moderate, High, and Very High). According to this map, approximately 34% of the province's area falls into the Low hazard class, 26% into Moderate, 3% into High, and 37% into the Very High hazard class. In total, 40% of the province is located in high and very high hazard zones. Critical Areas: Based on the final map, the Shahrekord, Borujen, Ben, and Saman plains were identified as the most critical areas with the highest degree of subsidence risk. Parts of the Lordegan and Ardal plains are also classified as areas with very high risk. In contrast, the Kuhrang plain is assessed to be at low and very low risk. Comparison with Other Studies: The accuracy of the SOM model in this study (AUC=0.76) is comparable to the results of a study in the Tabriz plain using a regression model (AUC=0.8) and a study in the Salmas plain using fuzzy logic (AUC=0.8), confirming its performance.
Conclusion and Suggestions
The present study confirms the effectiveness of the Self-Organizing Map (SOM) machine learning model in zonation of land subsidence hazard in Chaharmahal and Bakhtiari Province. The findings reveal an alarming situation, with more than one-third of the province's area (37 percent) exposed to a very high risk of subsidence, and the plains of Shahrekord, Boroujen, Ben, and Saman identified as the most critical zones. The innovation of this study lies in the simultaneous use of field data and artificial intelligence modeling to provide a clear and scientific picture of this crisis. The results of this research can serve as an effective tool for managers and planners to adopt preventive measures and ensure sustainable management of water and soil resources in the province. Accordingly, it is recommended that management actions be urgently concentrated in the critical plains, implementing strict control over aquifer extraction and executing aquifer rehabilitation projects, while continuous monitoring of subsidence rates using advanced radar interferometry technology is pursued. Furthermore, conducting additional research using more advanced machine learning models, more precise spatial and temporal data (such as time series of radar images and piezometric data), and a quantitative investigation of the impacts of human factors and climate change scenarios on the intensification of subsidence is recommended to achieve higher accuracy and a deeper understanding of this phenomenon.