Document Type : Research Paper

Authors

• Ahmadreza Karimipour
• Saleh Yousefi
• Sara Mardanian

Soil Conservation and Watershed Management Research Department, Chaharmahal and Bakhtiari Agricultural and Natural Resources Research and Education Center, AREEO, Shahrekord, Iran

Abstract

Introduction
The phenomenon of land subsidence, as a global challenge, has affected many plains and metropolitan areas, including various regions of Iran. This phenomenon is primarily caused by the uncontrolled exploitation of groundwater resources and geological factors, which can lead to serious consequences such as damage to infrastructure and the destruction of agricultural lands. Recent advances in remote sensing have enabled more accurate monitoring of this phenomenon, utilizing various techniques such as PS-InSAR and CPT. In addition to remote sensing, machine learning algorithms have also been used in various studies to predict subsidence. Accordingly, employing a suitable model with high accuracy in this field is of great importance. The AdaBoost model, due to its high capability in addressing the complex nonlinear relationships governing the subsidence phenomenon, can play an effective role in zoning subsidence risk levels, thereby contributing to risk management and land-use planning in a region.
 
Materials and methods
This study was conducted in Chaharmahal and Bakhtiari province, located in the heart of the Zagros Mountains. Initially, thirty factors related to land subsidence were considered, encompassing topographic (elevation, slope, aspect, curvature, TWI, TPI, TRI), hydrological (distance from rivers, flow accumulation), geological (distance from faults, lithology), environmental (vegetation cover, land moisture index, land use), and climatic parameters (temperature, precipitation, snow depth). To avoid multicollinearity, a correlation matrix analysis was performed, leading to the removal of seven variables with a correlation coefficient greater than 0.7. Subsequently, 23 variables were retained for modeling. The AdaBoost algorithm was trained on 2,352 samples (1,859 subsidence and 493 non-subsidence) and validated on an independent test set of 772 samples (536 subsidence and 236 non-subsidence). The model's performance was assessed using the Area Under the Curve (AUC), Precision, Recall, and Kappa coefficient metrics.
 
Results and discussion
The data of the obtained statistical indices for the Area Under the Curve (AUC), model accuracy (Precision), Recall index (Recall), and Kappa coefficient (0.974, 0.936, 0.981, and 0.855, respectively) in evaluating the AdaBoost model indicate the model's highly desirable performance in predicting subsidence risk. Based on the classification obtained in the final subsidence risk zoning map, the study area was categorized into five classes: very low risk, low risk, moderate, high, and very high. The results showed that the plains of Boroujen and Shahrekord are at the highest risk, respectively, while limited parts of the Lordegan plain fall into the very high-risk category. Other plains in the province are mainly faced with low to moderate risk levels. Furthermore, the results indicated that 36% of the province is exposed to moderate subsidence risk.
 
Conclusions
The findings of this study demonstrate that the AdaBoost model is an effective tool for zoning land subsidence risk in Chaharmahal and Bakhtiari Province. According to the SHAP analysis, the three main parameters influencing subsidence, in order of importance, are: land slope angle, surface sand percentage, and groundwater level fluctuations. It was further established that an inverse relationship exists between land slope and subsidence intensity, meaning areas with gentler slopes exhibit greater vulnerability to this phenomenon. Conversely, a decline in the groundwater level and its fluctuations have a direct correlation with increased subsidence, while a higher surface sand percentage exerts a mitigating effect on subsidence occurrence. The results underscore the necessity for proper management of groundwater resources. The subsidence risk zoning map and the evaluation charts presented in this study can serve as practical tools for preventing and mitigating damages caused by land subsidence.
 

Keywords

• Groundwater
• Groundwater level fluctuations
• Machine learning algorithms
• Risk
• Shahrekord