Risk assessment of geologic hazards in county-slope units based on RF-AER-AHP combined models Authors: Yang Wen, Ying Huang, Jian Ou, Can Wang, Kaichun Zhou, Yuqi Huang Published: 2026 ABSTRACT As among the most destructive types of natural hazards, geological disasters pose serious threats to human life, property safety, and socioeconomic development. Accurate identification and assessment of geological disaster risk are crucial for revealing the spatial distribution and underlying mechanisms of disasters and for providing scientific support for regional disaster prevention and territorial spatial planning. However, traditional risk assessment methods often suffer from high subjectivity in weight determination, limited model precision, and insufficient coupling among influencing factors, making it difficult to capture the complexity of disaster risk. To address these limitations, in this study, a coupled random forest–assumption of extreme rainfall–analytic hierarchy process (RF–ARE–AHP) model based on slope units for county-level geological disaster risk assessment was developed, using Ningyuan County in Hunan Province, China, as a case study. Model performance was validated using the area under the ROC curve (AUC), confusion matrix parameters, and field investigation results, and a high-resolution risk zoning map was generated at the slope-unit scale. The results show that the susceptibility model achieved an AUC of 0.92, with true positive, true negative, and overall accuracy rates exceeding 85% and false rates below 15%, indicating high model reliability and discriminative capacity. The hazard of geological disasters increased significantly with increasing rainfall intensity, expanding both the spatial extent and the severity of high-hazard areas. In the exposure assessment, the introduction of a damping coefficient effectively differentiated the exposure levels of the threatened elements. The medium–high-risk zones were concentrated mainly in Jiuyi Mountain, Mianhuaping, and Tongmulou Yao ethnic townships, which is consistent with field observations. Overall, the proposed RF–ARE–AHP model reduces subjectivity and enhances accuracy, providing a robust and scalable approach for county-level geological disaster risk evaluation and disaster prevention planning. Introduction Geological disasters usually refer to catastrophic events caused by a series of geological processes that pose serious threats to the safety of human life and property and natural ecosystems1. As a typical mountainous country, approximately 70% of China’s land area consists of mountains and plateaus. This complex terrain pattern results in geological disasters such as landslides, collapses and debris flows, which are characterized by high-frequency development and strong destruction. For example, in July 2024, under the influence of Typhoon Gemei, in Zixing city, Hunan Province, extreme rainstorms occurred, and large-scale geological disasters occurred in many towns and villages, including Bamianshan Yao township, resulting in the collapse of 1,641 houses and damage to 76,300 mu of cultivated land2. This event highlights the suddenness and complexity of geological disasters against the background of climate change and reflects the significant impact of small- and medium-scale rainstorm processes on the stability of the geological environment in mountainous areas. The occurrence of geological disasters is essentially the result of the interaction of geological environment evolution and external disturbance, and its formation mechanism is driven mainly by two kinds of factors: background conditions, such as regional geological structure, lithology and landforms, and external triggering factors, such as rainfall, earthquakes and anthropogenic activities3,4. Academic researchers have long carried out investigations of the underlying mechanisms and coupling effect of these two factors5–8, among which the issue of how to quantitatively describe the spatial distribution and dynamic evolution of geological disaster risk has gradually become an important direction of geological disaster research9. The core goal of geological disaster risk assessment is to quantitatively describe the occurrence probability and potential loss of a specific region under the action of geological disasters with different intensities. This process not only is an important extension of the study of geological disaster mechanisms but also provides a scientific basis for regional disaster prevention and mitigation, project site selection and government risk management and control10. Risk assessment usually includes susceptibility, hazard, exposure and vulnerability, which are interrelated and constitute a comprehensive risk system for geological disasters11. In terms of susceptibility assessment, the research methods can be divided into knowledge-driven methods, data-driven methods and comprehensive integrative methods12. The knowledge-driven method relies on expert experience and theoretical judgment. Although it is simple and efficient, it is subjective13. With the rapid development of computer technology and artificial intelligence algorithms, data-driven methods based on machine learning have gradually become mainstream14. Among them, algorithms such as support vector machine (SVM), convolutional neural network (CNN) and random forest (RF), which have strong feature extraction and nonlinear fitting capabilities, have been widely applied and achieved good performance in landslide susceptibility assessment15–17. In recent years, landslide susceptibility assessment has further advanced beyond conventional machine learning approaches toward more sophisticated frameworks, such as UAV-based landslide detection using generative AI, hybrid models combining physics-based approaches with extreme value analysis, uncertainty-aware modeling of hydrological parameters, and physics-informed deep learning models18–21.Hazard assessment focuses on the dynamic triggering process of geological disasters. A rainfall-induced model based on the assumption of extreme rainfall (AER) can reflect the nonlinear influence of rainfall intensity change on the probability of geological disasters, especially in areas with severe rainfall change and detailed disaster records22. This method reveals the dynamic law of geological disaster hazard changes with fluctuations in climatic factors and provides theoretical support for the process simulation and early warning of rainstorm-related geological disaster hazard.The focus of exposure assessment lies in the identification and characterization of the spatial exposure level and damage characteristics of disaster-affected entities. This process typically relies on existing survey data to analyze the distribution, quantity, and spatial coupling relationships between different types of disaster-affected entities23. However, on a large regional scale, the types of disaster-bearing bodies are diverse, and their structures are complex; thus, fully reflecting their risk characteristics by relying solely on data-driven quantitative methods is often difficult24. Therefore, the combination of knowledge-driven and data-driven comprehensive methods, such as the spatial multicriteria evaluation method, TOPSIS method and AHP method, is more scientific and applicable in the exposure assessment of geological disasters24,25. In terms of susceptibility assessment, the research methods can be divided into knowledge-driven methods, data-driven methods and comprehensive integrative methods12. The knowledge-driven method relies on expert experience and theoretical judgment. Although it is simple and efficient, it is subjective13. With the rapid development of computer technology and artificial intelligence algorithms, data-driven methods based on machine learning have gradually become mainstream14. Among them, algorithms such as support vector machine (SVM), convolutional neural network (CNN) and random forest (RF), which have strong feature extraction and nonlinear fitting capabilities, have been widely applied and achieved good performance in landslide susceptibility assessment15–17. In recent years, landslide susceptibility assessment has further advanced beyond conventional machine learning approaches toward more sophisticated frameworks, such as UAV-based landslide detection using generative AI, hybrid models combining physics-based approaches with extreme value analysis, uncertainty-aware modeling of hydrological parameters, and physics-informed deep learning models18–21. In terms of susceptibility assessment, the research methods can be divided into knowledge-driven methods, data-driven methods and comprehensive integrative methods12. The knowledge-driven method relies on expert experience and theoretical judgment. Although it is simple and efficient, it is subjective13. With the rapid development of computer technology and artificial intelligence algorithms, data-driven methods based on machine learning have gradually become mainstream14. Among them, algorithms such as support vector machine (SVM), convolutional neural network (CNN) and random forest (RF), which have strong feature extraction and nonlinear fitting capabilities, have been widely applied and achieved good performance in landslide susceptibility assessment15–17. In recent years, landslide susceptibility assessment has further advanced beyond conventional machine learning approaches toward more sophisticated frameworks, such as UAV-based landslide detection using generative AI, hybrid models combining physics-based approaches with extreme value analysis, uncertainty-aware modeling of hydrological parameters, and physics-informed deep learning models18–21. Hazard assessment focuses on the dynamic triggering process of geological disasters. A rainfall-induced model based on the assumption of extreme rainfall (AER) can reflect the nonlinear influence of rainfall intensity change on the probability of geological disasters, especially in areas with severe rainfall change and detailed disaster records22. This method reveals the dynamic law of geological disaster hazard changes with fluctuations in climatic factors and provides theoretical support for the process simulation and early warning of rainstorm-related geological disaster hazard. Hazard assessment focuses on the dynamic triggering process of geological disasters. A rainfall-induced model based on the assumption of extreme rainfall (AER) can reflect the nonlinear influence of rainfall intensity change on the probability of geological disasters, especially in areas with severe rainfall change and detailed disaster records22. This method reveals the dynamic law of geological disaster hazard changes with fluctuations in climatic factors and provides theoretical support for the process simulation and early warning of rainstorm-related geological disaster hazard. The focus of exposure assessment lies in the identification and characterization of the spatial exposure level and damage characteristics of disaster-affected entities. This process typically relies on existing survey data to analyze the distribution, quantity, and spatial coupling relationships between different types of disaster-affected entities23. However, on a large regional scale, the types of disaster-bearing bodies are diverse, and their structures are complex; thus, fully reflecting their risk characteristics by relying solely on data-driven quantitative methods is often difficult24. Therefore, the combination of knowledge-driven and data-driven comprehensive methods, such as the spatial multicriteria evaluation method, TOPSIS method and AHP method, is more scientific and applicable in the exposure assessment of geological disasters24,25. The focus of exposure assessment lies in the identification and characterization of the spatial exposure level and damage characteristics of disaster-affected entities. This process typically relies on existing survey data to analyze the distribution, quantity, and spatial coupling relationships between different types of disaster-affected entities23. However, on a large regional scale, the types of disaster-bearing bodies are diverse, and their structures are complex; thus, fully reflecting their risk characteristics by relying solely on data-driven quantitative methods is often difficult24. Therefore, the combination of knowledge-driven and data-driven comprehensive methods, such as the spatial multicriteria evaluation method, TOPSIS method and AHP method, is more scientific and applicable in the exposure assessment of geological disasters24,25. In addition, recent studies have emphasized the importance of vulnerability and risk assessment, including semi-quantitative estimation of risk to people in buildings and quantitative modeling of debris-flow impact forces26,27. These studies indicate that hazard assessment alone is insufficient for supporting decision-making, and that a comprehensive framework integrating susceptibility, hazard, exposure, and vulnerability is required. However, such integrated frameworks are still limited at the county scale, particularly in terms of combining high-resolution susceptibility mapping, rainfall-triggered hazard scenarios, and terrain-corrected exposure assessment. Notably, the accuracy of risk assessment depends not only on the model method but also on the scale of the evaluation unit. At present, the commonly used risk assessment units include grid units, regional units, homogeneous condition units, landform units and slope units (SUs)28. Among them, the SUs is defined by the ridge line and the valley line, which can better reflect the spatial consistency of landform evolution, hydrological processes and geological disaster breeding mechanisms29. As the basic terrain unit of geological disasters, evaluation based on SUs can significantly improve the spatial resolution and rationality of risk identification and subsequently increase the scientific and application value of model results30. Although scholars worldwide have made rich achievements in the study of the susceptibility to, hazard of and exposure of geological disasters, some problems still exist with the existing methods, such as strong subjectivity in weight determination and insufficient model accuracy, and comprehensive methods need to be improved31. To solve this problem, this paper relied on the basic data and samples provided by the 1:10,000 geological disaster investigation and evaluation project in Ningyuan County, Hunan Province; the Ningyuan County region was selected as the research area, the susceptibility model was established by using the RF method, the risk model was constructed by using the AER, and exposure evaluation was carried out by combining the analytic hierarchy process (AHP). On this basis, the incorporation of Pearson’s correlation coefficient, Bayesian algorithms, and slope loss coefficient served as optimization techniques. These were then integrated into a coupled model, namely, the random forest–assumption of extreme rainfall–analytic hierarchy process (RF-AER-AHP), for the systematic, quantitative, and spatial evaluation of geological hazard risks in Ningyuan County. Therefore, this study aims to develop a county-scale geohazard risk assessment framework by integrating susceptibility, hazard, and exposure. The specific objectives are to: (1) construct a slope-unit-based susceptibility model using Bayesian-optimized RF; (2) quantify hazard under multiple rainfall scenarios; and (3) improve exposure estimation using slope-loss correction. The contribution of this study does not lie in introducing a completely new standalone method, but in developing an integrated county-scale geohazard risk assessment framework. Specifically, this study contributes in three aspects: (1) implementing susceptibility assessment on slope units to better reflect geomorphological controls; (2) improving model robustness through Bayesian optimization of the RF model; and (3) coupling susceptibility, multi-scenario rainfall-triggered hazard, and exposure with a slope-loss correction to enhance the spatial realism of risk assessment. The rest of this paper is organized as follows. Section "Overview of the study area and data sources" describes the study area and data sources. Section “Research method” presents the methodology, including the Bayesian-optimized RF model, rainfall scenario design, and risk assessment framework. Section “Results” reports the results. Section 5 discusses the findings and limitations. Finally, Sect. "Conclusion and Outlook" concludes the study. Overview of the study area and data sources Overview of the study area Ningyuan County is located in the southern part of Hunan Province on the northern slope of Mengzhuling, which is one of the Five Ridges Mountains. It lies between the Qiyi Mountains and Yangming Mountains. Its geographical coordinates are 111°43′25″-112°15′10″E and 25°11′39″-26°06′20″N. It is elongated from north to south and narrow from east to west, resembling a boat. Under its jurisdiction are 12 towns, 4 subdistricts, and 4 ethnic minority townships. It belongs to the Central Subtropical Humid Monsoon Climate Zone, characterized by long summers and winters and short springs and autumns, with a distinct rainy season. The southern and northern regions have higher terrains, while the central area features alternating hills and plains with relatively lower elevations, making it one of the few counties in Hunan Province with significant topographic variations. The exposed strata within the region are predominantly composed of shallowly metamorphosed rocks of Ordovician age and carbonate rock formations of Carboniferous and Devonian systems, which are widely distributed. According to statistics, there are 92 geological disasters in Ningyuan County, mainly landslides, including 83 landslides, 8 collapses and 1 debris flow. Geological disasters in Ningyuan County are mostly developed on both sides of tributaries of various water systems, on both sides of artificial roads, and behind houses built by cutting slopes. Affected by the unfavorable combination of tectonic action, lithologic structure, river landform evolution, human activities, vegetation coverage and other disaster-causing backgrounds, it is concentrated or banded. Meteorological precipitation conditions are the most important factor that can induce geological disasters in Ningsanyuan County, and the disaster occurrence time is highly consistent with that during the main flood season, especially when it is affected by extreme meteorology caused by a typhoon, which has obvious characteristics of regional mass, time concentration and local suddenness. Data sources The data were mainly from the 1:10,000 geological disaster investigation and evaluation project in Ningyuan County, and related data were collected (Table 1), including the 1:10,000–1:2,000 topographic map of Ningyuan County; the 1:250,000 geological map of Ningyuan County; Sentinel-2 data; Ningyuan County information; and the field survey data of the project team. The data were true and detailed and provide solid data support for the risk assessment of geological disasters in Ningyuan County. Table 1Data categories and features.data typeSpatial resolution and scaleData usageSentinel-2Multispectral 10 mNDVI calculationtopographic map1:2000, 1:10000Generation of a digital terrain elevation model (DEM) and calculation of the slope, aspect, etc.geologic map1:2500000Extraction of the lithology of the strata, engineering geological rock groups, and fault distancesFengdu Integrated Land and Housing Data of Ningyuan CountyExtraction of exposure factors Table 1Data categories and features.data typeSpatial resolution and scaleData usageSentinel-2Multispectral 10 mNDVI calculationtopographic map1:2000, 1:10000Generation of a digital terrain elevation model (DEM) and calculation of the slope, aspect, etc.geologic map1:2500000Extraction of the lithology of the strata, engineering geological rock groups, and fault distancesFengdu Integrated Land and Housing Data of Ningyuan CountyExtraction of exposure factors Data categories and features. Slope units division Slope units(SUs) were adopted as units in this study because they take ridge and valley lines as natural boundaries, which are consistent with the topographic and hydrological laws of geohazard gestation and occurrence. This unit avoids the evaluation defects of grid and administrative units, accurately characterizes the geohazard characteristics of a single slope, and matches the refined risk assessment requirements of this study, thus improving the scientificity and spatial accuracy of county-scale geohazard risk assessment32.On the basis of the principles of hydrological methods33, in this study, forward and backward digital elevation model (DEM) data were used, and the boundary between the watershed and catchment area was extracted through the steps of filling the depression, flow direction and flow calculation to delimit the slope unit of the study area34. After processing, a total of 18,416 SUs were obtained for Ningyuan County, providing basic spatial unit support for the subsequent risk assessment. On this basis, further integration with high-resolution hazard-bearing body image data was conducted to carry out a refined investigation in key areas characterized by dense geological disasters development, complex geological environments, frequent engineering construction activities, and high population aggregation. On the basis of the spatial distribution characteristics and representativeness, 4760 SUs were selected for field investigation. The investigation encompasses multiple dimensions, including hazardogenic geological conditions, the intensity of human engineering activities, vegetation and land use status, abnormal phenomena such as surface seepage and rockfalls, and onsite and historical deformation indicators, to comprehensively assess the susceptibility and risk levels of geological hazards across various SUs.The susceptibility grades evaluated in field investigations were taken as supplementary model samples for machine learning modeling., Research method The geological disaster risk assessment workflow of this study takes SUs as the core evaluation units and is carried out in three key stages: susceptibility, hazard and exposure assessment, ultimately generating a geological disaster risk map. For susceptibility assessment, positive and negative samples are first delineated from SUs with defined susceptibility grades by synthesizing field investigations and historical geological disaster data. A total of 16 evaluation factors across five categories, namely topographic, geological, hydrological, human engineering activity and other factors, are selected and filtered via Pearson correlation analysis. The filtered factors are split into a training set and a test set at a ratio of 7:3 to construct a Bayesian-optimized RF model, whose accuracy is verified using ROC curves and confusion matrix parameters. For hazard assessment, the Assumption of Extreme Rainfall (AER) method is adopted to calculate and compare the geological disaster hazard levels under four rainfall scenarios: 50 mm/24 h, 100 mm/24 h, 150 mm/24 h and 166.6 mm/24 h. For exposure assessment, population, building density, road density and other infrastructure are selected as evaluation factors; the analytic hierarchy process (AHP) is used to determine the weight of each factor, and a slope loss coefficient is introduced to correct the assessment results (with the vulnerability of exposed elements simplified and set to 1). Finally, the results of the three aforementioned assessments are integrated to generate a high-resolution geological disaster risk distribution map of the study area, as shown in Fig. 1. Fig. 1Flowchart of risk assessment. Flowchart of risk assessment. Susceptibility assessment Selection of susceptibility evaluation factors The occurrence process of geological disasters is complex and diverse, and it is influenced by a variety of terrains, geological and hydrological characteristics and anthropogenic activities. On the basis of previous research results and the geological environmental characteristics of the study area15,16,24,35,36, 16 evaluation factors were selected in this study to construct a susceptibility evaluation system. Among them, the landform factors included the elevation, slope, aspect, curvature, ground roughness and topographic relief index(TRI); the geological factors included the formation lithology, geological rock group engineering, slope structure and fault distance; and the hydrological condition factors included the topographic wetness index(TWI), stream power index(SPI) and water system distance. The factors of human engineering activities included the cut-slope building density of medium- and high-risk slope cutting and road distance. The other factor was the vegetation normalization index (NDVI). The source and spatial resolution of each factor are shown in Table 2. To realize the unified treatment of factors, all the susceptibility evaluation factors were converted into a grid unit of 10 m×10 m, and each factor was graded and quantified by using the natural breakpoint method. Afterward, the “Zonal Statistics as Table” tool in ArcGIS 10.3 was employed. Using the slope unit polygon layer as the zonal feature, the median value of each susceptibility factor raster layer within each unit was calculated, thus generating a classification map of susceptibility assessment factors based on SUs. Table 2Data sources of the susceptibility assessment factors.Susceptibility evaluation factorData sourceResolution ratioElevation, slope, aspect, curvature, topographic relief index, ground roughness, topographic wetness index and stream power index.Third Institute of Surveying and Mapping, Hunan Province, 1: 10,000 ~ 1: 2,000 topographic map.10 mWater system distance1:2000Highway distance1:2000Stratigraphic lithologyHunan Provincial Natural Resources Affairs Center, 1:250,000 geological map1:250,000Engineering geological rock group1:250,000Discontinuity distance1:250,000Slope structureField surveys were conducted to collect dataNormalized Vegetation Index https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR_HARMONIZED 30 m Table 2Data sources of the susceptibility assessment factors.Susceptibility evaluation factorData sourceResolution ratioElevation, slope, aspect, curvature, topographic relief index, ground roughness, topographic wetness index and stream power index.Third Institute of Surveying and Mapping, Hunan Province, 1: 10,000 ~ 1: 2,000 topographic map.10 mWater system distance1:2000Highway distance1:2000Stratigraphic lithologyHunan Provincial Natural Resources Affairs Center, 1:250,000 geological map1:250,000Engineering geological rock group1:250,000Discontinuity distance1:250,000Slope structureField surveys were conducted to collect dataNormalized Vegetation Index https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR_HARMONIZED 30 m Data sources of the susceptibility assessment factors. Based on Pearson correlation coefficient analysis, factors with absolute correlation coefficients exceeding 0.5 (specifically, ground roughness, topographic relief index, and topographic wetness index) were eliminated37, ultimately retaining 12 factors for subsequent model calculation (Figs. 2 and 3). Fig. 2Pearson correlation coefficient results. Pearson correlation coefficient results. Fig. 3Susceptibility assessment factor graphs(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview). Susceptibility assessment factor graphs(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview). Bayesian-optimized random forest model Random forest (RF) is an idea based on bagging ensemble learning. By constructing multiple decision trees and integrating the voting results, the optimal classification result can be obtained38, as shown in Formula 1:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat {C}_{{rf}}^{B}(x){\text{ }}={\text{ }}majority{\text{ }}vote{\text{ \{ }}{\hat {C}_b}{\mathrm{(}}x{\mathrm{)\} }}_{{\mathrm{1}}}^{B}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat {C}_b}(x)$$\end{document} represents the classification result of the b-th tree, B is the total number of decision trees, and the final classification result \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat {C}_b}(x)$$\end{document} is obtained by integrating the majority votes of the b-th decision tree. Each decision tree extracts features in an independent and random way, which makes the model robust and reduces the risk of overfitting39. In addition, compared with traditional linear models, the RF classifier has a strong ability to fit datasets and can effectively address complex nonlinear relationships, and its calculation efficiency is better; thus, it has been widely used in the analysis of geological disaster susceptibility35. Grid search and the Bayesian algorithm are commonly used hyperparametric optimization methods for machine learning. Unlike the grid search method36, the Bayesian algorithm defines the objective function and iteratively estimates the posterior distribution of the objective function and the hyperparametric subspace, which may optimize the objective function. Specifically, for each iteration, Bayesian optimization first samples the hyperparameters in a given range using a uniform distribution and then estimates the posterior distribution of the objective function of each sampled superparameter combination using a Gaussian process. An acquisition function is subsequently used to identify the hyperparametric region that may optimize the objective function. The acquisition function uses tuning parameters to balance the posterior mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu (x)$$\end{document} and variance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma (x)$$\end{document} of the objective function, which can be expressed as Eq. (1)40:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$AC(x){\text{ = }}\upmu {\mathrm{(}}x{\text{) + k}}\sigma (x)$$\end{document} By using the acquisition function to identify the optimal subspace of the hyperparameter, Bayesian optimization can avoid the region that reaches the optimal value of the objective function but has a large variance. By iteratively estimating the posterior distribution of the objective function and adaptively updating the hyperparametric subspace, Bayesian optimization can use the information of prior evaluation and focus only on those subspaces that are most likely to result in improvements, which simplifies the hyperparametric search process and saves computational resources41. Model samples In view of the relatively limited number of known geological hazards in the study area, in this paper, the training of the RF model in this study was based on a binary classification. The positive samples were defined by integrating 92 SUs containing historically inventoried geological disaster locations (providing objective evidence) with 256 SUs identified as medium- or high-susceptibility zones through detailed field investigations (incorporating expert prior knowledge).Negative samples were randomly selected from the pool of SUs determined as low susceptibility through the same field survey, maintaining a positive-to-negative sample ratio of 1:5, resulting in 1740 negative samples, and the training set and test set were divided according to a ratio of 7∶3 to support the training and verification of the geological disaster susceptibility model.At present, the mainstream sampling strategies for non-geohazard/negative sample units are mainly divided into four categories: random sampling42, buffer zone sampling43, environmental feature space sampling44 and prior model sampling45. In contrast to random sampling, buffer zone sampling that tends to excessively exclude potentially relevant areas, single environmental feature space sampling that is prone to ignoring the complex disaster-causing mechanisms of geological hazards, and prior model sampling that may introduce prior errors, the strategy adopted in this study fully integrates field investigation data. By delimiting low-prone areas based on field surveys as the negative sample pool, it excludes regions with unclear geological conditions and uncertain susceptibility assessment, which is more consistent with the actual development characteristics of geological hazards in the study area and can effectively reflect the real differences in the susceptibility of SUs within the study area. Evaluation of the susceptibility accuracy To comprehensively evaluate the performance of the model, a ROC curve and confusion matrix parameters were used. ROC characteristic curve. ROC characteristic curve. ROC characteristic curve. The receiver operating characteristic (ROC) curve is used to visually evaluate the discrimination ability of a classification model by depicting the relationship between the true positive rate (TPR) and false-positive rate (FPR) under different classification thresholds46. The area under the curve (AUC) is a quantitative index to measure the overall performance of the model. The closer the AUC value is to 1, the better the classification performance of the model is47. Because the ROC curve is insensitive to the change in the proportion of positive and negative samples, it can still reflect the stability of the model under the condition of unbalanced samples. (2)Confusion matrix parameters. (2)Confusion matrix parameters. Confusion matrix parameters. To further verify the accuracy of the model, five confusion matrix indicators were used for evaluation48. ① The true-positive rate (TPR), that is, the proportion of correctly classified positive samples; ② the false-positive rate (FPR), that is, the proportion of wrongly classified positive samples; ③ the true-negative rate (TNR), that is, the proportion of correctly classified negative samples; (4) the false-negative rate (FNR), that is, the proportion of wrongly classified negative samples; and ⑤ the overall accuracy rate (OA), that is, the proportion of correctly predicted samples to all samples. A schematic diagram of the confusion matrix is shown in Fig. 4, where P and N represent the numbers of positive samples and the negative samples SUs, respectively, obtained from the field investigation, TP and TN are the numbers of correctly classified samples, and FP and FN are the numbers of incorrectly classified samples. Fig. 4Confusion matrix parameters. Confusion matrix parameters. Hazard assessment In accordance with the internationally accepted risk paradigm49,50, a comprehensive hazard assessment requires a probabilistic analysis encompassing three dimensions: space, time and intensity, namely addressing the questions of where a disaster may occur, when it may occur, and how large its magnitude may be. Susceptibility assessment primarily characterizes the spatial dimension of disaster occurrence, reflecting the inherent and static disaster-forming potential of a region. To evaluate the temporal dimension, however, the effects of external triggering factors must be taken into account. As the primary triggering factor, the dynamic incorporation of rainfall is precisely intended to quantify the probability of disaster occurrence under specific rainfall conditions. However, in large-scale regional assessments, it is often challenging to directly obtain precise data on the timing of slope instability (e.g., recurrence intervals) and movement intensity (e.g., velocity, impact force) for each individual slope. To address this, this study employs the AER as an indirect modeling strategy to quantify the temporal dimension51. This method is designed to estimate the normalized hazard index of geological hazards, which comprehensively characterizes the hazard degree of geological hazards in each SUs of the study area under different extreme rainfall conditions. It fully considers the historical geological disasters caused by extreme rainfall events in the study area and combines the measured extreme rainfall records with the dynamic rainfall intensity changes, emphasizing the functional relationship between geological disaster hazard and rainfall-induced probability, which can truly reflect the dynamic evolution process of hazard22. By calculating the hazard index of each slope unit, the hazard zoning under different rainfall conditions is obtained, and the formula is as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}=\frac{{{Y_i}}}{{{Y_{max}}}}{P_i}$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_i}=\frac{{{L_i}}}{{{L_{max/day}}}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document} is the hazard index of the i-th slope unit in Ningyuan County under a certain rainfall condition; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y_i}$$\end{document} is the susceptibility index of the i-th slope unit; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y_{\hbox{max} }}$$\end{document} is the maximum susceptibility index of the Ningyuan County slope unit; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_i}$$\end{document} is the instability probability of the i-th slope unit in Ningyuan County during a given period of time under a certain rainfall condition; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L_{max/day}}$$\end{document} is the maximum daily rainfall since the monitoring record was recorded in the study area; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L_i}$$\end{document} is the annual average maximum daily rainfall under different working conditions (50 mm/24 h, 100 mm/24 h, 150 mm/24 h and 166.6 mm/24 h). Exposure assessment Exposure factor selection On the basis of comprehensive consideration of the characteristics of geological disasters in Ningyuan County, such as the population distribution, building density and infrastructure, five types of exposure assessment factors, namely, population, number of buildings, traffic roads, other living facilities (oil and gas, water transmission, communication lines) and the density of important facilities (schools, hospitals, nursing homes, power stations, etc.), were selected (Table 3). These factors can comprehensively reflect the degree of regional population concentration, the intensity of social and economic activities and the spatial distribution characteristics of key infrastructure and are important for depicting the potential losses of geological disasters. See Table 5 for the data sources of each evaluation factor. The data of exposure factors based on grid units were quantified and graded by the natural breakpoint method on the ArcMap platform, and the median value of factors under the scale of SUs was extracted by using the “Zonal Statistics As Table” tool, thus generating a exposure factor classification map with SUs as evaluation units (Fig. 5). Table 3Data sources of the exposure assessment factors.Exposure evaluation factor(s)Data sourceResolution ratiopopulationNingyuan county Natural Resources Bureau, integrated data of premisesbuildingThird Institute of Surveying and Mapping, Hunan Province, 1:2000 topographic map1:2000Traffic roadOther living facilities (oil and gas, water supply, power transmission and communication lines)Density of important facilities (schools, hospitals, nursing homes, power stations, etc.) Table 3Data sources of the exposure assessment factors.Exposure evaluation factor(s)Data sourceResolution ratiopopulationNingyuan county Natural Resources Bureau, integrated data of premisesbuildingThird Institute of Surveying and Mapping, Hunan Province, 1:2000 topographic map1:2000Traffic roadOther living facilities (oil and gas, water supply, power transmission and communication lines)Density of important facilities (schools, hospitals, nursing homes, power stations, etc.) Data sources of the exposure assessment factors. Fig. 5Exposure assessment factor graphs(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Exposure assessment factor graphs(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Analytic hierarchy process The analytic hierarchy process (AHP) was proposed by the American logistics scientist Saaty, and it is a multilevel decision-making method that combines qualitative analysis and quantitative calculation52. By constructing a multilevel evaluation structure system, this method judges the relative importance of each factor and calculates its weight coefficient; thus, it is widely used to determine the exposure weight of geological disasters53. The specific steps are as follows. Build a hierarchical structure. The exposure assessment system of geological disasters is decomposed into three levels: target level, criterion level and scheme level. Among them, the target layer is comprehensive exposure. The criterion layer includes population exposure, material exposure and the exposure of important facilities. The scheme layer consists of five specific factors: population, buildings, traffic roads, other living facilities and the density of important facilities.Construct a judgment matrix. In accordance with expert knowledge and experience, the factors of each criterion layer and scheme layer are compared, a judgment matrix is constructed, and the weight relationship of the relative importance of each factor is determined.By calculating the consistency ratio (CR), the rationality of the judgment matrix is verified to ensure the consistency and credibility of the weight distribution. Build a hierarchical structure. The exposure assessment system of geological disasters is decomposed into three levels: target level, criterion level and scheme level. Among them, the target layer is comprehensive exposure. The criterion layer includes population exposure, material exposure and the exposure of important facilities. The scheme layer consists of five specific factors: population, buildings, traffic roads, other living facilities and the density of important facilities. Build a hierarchical structure. The exposure assessment system of geological disasters is decomposed into three levels: target level, criterion level and scheme level. Among them, the target layer is comprehensive exposure. The criterion layer includes population exposure, material exposure and the exposure of important facilities. The scheme layer consists of five specific factors: population, buildings, traffic roads, other living facilities and the density of important facilities. Construct a judgment matrix. In accordance with expert knowledge and experience, the factors of each criterion layer and scheme layer are compared, a judgment matrix is constructed, and the weight relationship of the relative importance of each factor is determined. Construct a judgment matrix. In accordance with expert knowledge and experience, the factors of each criterion layer and scheme layer are compared, a judgment matrix is constructed, and the weight relationship of the relative importance of each factor is determined. By calculating the consistency ratio (CR), the rationality of the judgment matrix is verified to ensure the consistency and credibility of the weight distribution. By calculating the consistency ratio (CR), the rationality of the judgment matrix is verified to ensure the consistency and credibility of the weight distribution. Slope loss coefficient The topography of Ningyuan County is rather undulating, with the elevation of Benji Wo Mountain, the highest point of Jiuyi Mountain in the south, reaching 1,959.2 m, and the lowest point of the county, the sliding stone slab, reaching 165 m, with a vertical elevation difference of 1,794.2 m and a gradient of 36.6‰. Significant topographical relief leads to the concentration of population and buildings in flat areas such as county seats and townships. In traditional exposure assessments, these flat zones are often categorized as having high or extremely high exposure because of the dense concentration of exposure targets. However, in reality, these areas do not present significant geological hazard risks. If such results are directly used in the calculation of “Risk (R)= Hazard (H)× Exposure (H)× Vulnerability (V)”, excessive expansion of the risk area will occur, thus misleading the judgment of local governments in terms of disaster prevention and mitigation and project site selection. To overcome this problem, in this study, the slope loss coefficient was introduced to correct the assessment results. A large number of studies have demonstrated a significant positive correlation between slope and the movement distance of geohazard bodies: the steeper the slope, the farther the movement distance and the wider the influence range of geohazard bodies, and the corresponding broader the scope of disaster-bearing bodies at the slope foot exposed to risks; conversely, gentle slopes lead to a significant narrowing of the geohazard influence range54–56. Based on this, differential loss ratios were set in this study according to slope intervals (Table 4), so as to reflect the restrictive effect of site topography on the actual disaster-affected range.When the slope gradient of a slope element is less than 5°, there is virtually no risk of instability; when the gradient is between 5° and 15°, buildings located near the toe of the slope are at risk, and when the gradient exceeds 35°, all components of the slope element are at high risk of experiencing disaster impacts. The loss coefficient was determined through an expert judgment approach, aiming to better align the exposure assessment with the actual threatened area. Table 4Exposure slope loss coefficient.Slope classification< 5°5–15°15–25°25–35°> 35°Exposure loss coefficient00.10.40.71 Table 4Exposure slope loss coefficient.Slope classification< 5°5–15°15–25°25–35°> 35°Exposure loss coefficient00.10.40.71 Exposure slope loss coefficient. Risk assessment Based on the risk assessment formula R = H×E×V, the hazard (H) and exposure (E) were both quantitatively obtained by the RF-AER coupling method and the AHP process described above. Aiming at the demand of risk assessment at the county-scale large regional level, the vulnerability (V) was uniformly set to 1 in this study. This treatment is a common simplified strategy based on specific research constraints in large-scale geohazard risk assessment, with the core assumption that all types of structures and disaster-bearing bodies within the scope of the geohazard impact path will suffer irreversible damage when a geohazard occurs57.it enables the research to achieve an overall grasp and spatial comparison of the regional risk pattern at the macro scale, which is more in line with the actual needs of macro prevention, control and planning of geohazard risk at the county scale.Based on this, the risk index for each SUs was calculated. Afterward, the natural breakpoint method was used to partition the risk results under the four rainfall conditions, and the study area was divided into low-risk areas, medium-risk areas, high-risk areas and very high-risk areas. Results Susceptibility assessment results On the basis of the optimal hyperparameter combination obtained by Bayesian optimization, the RF susceptibility model was constructed, and the ROC curve was analyzed (Fig. 6). The results show that the AUC value of the model was 0.92 and that the optimal classification threshold was 0.41. The results of the confusion matrix calculated according to this threshold (Fig. 7) show that the true positive rate, true negative rate and overall accuracy rate were all greater than 85%, while the false-positive rate and false-negative rate were all less than 15%, which shows that the model has excellent classification performance and strong generalization ability. Following the utilization of an optimal model for global prediction, the natural breaks method was employed on the ArcGIS platform to delineate the susceptibility zones of geological hazards in Ningyuan County (Fig. 8). The results reveal that low susceptibility areas account for 58.7%, medium susceptibility areas account for 20.1%, high susceptibility areas account for 16.1%, and very high susceptibility areas account for 5.1%. The high susceptibility and very high susceptibility areas are concentrated mainly in Jiuyishan Yao Township, Wulong Mountain Scenic Resort Township and Mianhuaping Yao Township, and their spatial distribution characteristics are highly consistent with the topographic and geological conditions. To further verify the reliability of the model, the geological disaster points in Ningyuan County are superimposed on the prone zoning map for analysis. The results show that the proportion of disaster points in SUs with moderate or above susceptibility is 88.0%. The overall results show that the RF susceptibility model based on Bayesian optimization has high accuracy and stability and can provide scientific support for the risk prevention and control of geological disasters and spatial planning at the county level 7. Fig. 6ROC curve for the RF model with Bayesian optimal hyperparameters. ROC curve for the RF model with Bayesian optimal hyperparameters. Fig. 7Confusion matrix parameters with optimal thresholds. Confusion matrix parameters with optimal thresholds. Fig. 8Geological hazard susceptibility partitioning of Ningyuan County(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Geological hazard susceptibility partitioning of Ningyuan County(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Hazard assessment results In accordance with the meteorological records of Ningyuan County, the annual maximum daily precipitation since the monitoring data were available was 166.6 mm (June 5, 2004), which was defined as the extreme rainfall condition in this study. In accordance with the precipitation grade standard formulated by the National Standardization Administration Committee58, A 24 h cumulative precipitation of 25–49.9 mm is defined as “moderate rain”, 50–99.9 mm as “heavy rain”, and 100–249.9 mm as “torrential rain”. In combination with the multi-year rainfall characteristics of Ningyuan County, four typical rainfall scenarios were divided: 50 mm/24 h, 100 mm/24 h, 150 mm/24 h and 166.6 mm/24 h (extreme rainfall). On the basis of the evaluation results of geological disaster susceptibility, the hazard index of each SUs under different rainfall conditions was calculated via the ArcGIS platform combined with Formulas (3) and (4) and the risk division standard51 in Table 5, and the geological disaster hazard zoning map of Ningyuan County was drawn accordingly (Fig. 9). The results show that under the condition of 50 mm/24 h, the whole study area was mainly a low-hazard area; with increasing rainfall intensity, the range and grade of dangerous areas increased gradually. When the rainfall intensity reached 50 mm/24 h, 100 mm/24 h, and 150 mm/24 h and the extreme value was 166.6 mm/24 h, the proportions of medium- and high-hazard areas within the study area were 0%, 4.9%, 6.8% and 8.4%, respectively. There was a significant positive correlation between rainfall and geological disaster hazard. Fig. 9Geological hazard zones in Ningyuan County under four rainfall conditions(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Geological hazard zones in Ningyuan County under four rainfall conditions(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Table 5Standards for Geologic Hazard Level Classification48.Risk degree of geological disastersVery High-HazardHigh-HazardMedium-HazardLow-HazardRisk index\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}> 0.80.6<\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}≦0.80.4<\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}≦0.60<\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}≦0.4 Table 5Standards for Geologic Hazard Level Classification48.Risk degree of geological disastersVery High-HazardHigh-HazardMedium-HazardLow-HazardRisk index\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}> 0.80.6<\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}≦0.80.4<\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}≦0.60<\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${H_i}$$\end{document}≦0.4 Standards for Geologic Hazard Level Classification48. Exposure assessment results The weights of each exposure factor were calculated using the AHP, with the results presented in Table 6.The maximum eigenvalue λ_max was 5.1448, and the consistency 10 ratio Cr was 0.032 < 0.1, which shows that the judgment matrix met the consistency test requirements. Table 6Exposure factor matrix and weighting values according to the AHP.Evaluation factorpopulationbuildingtransportation facilitiesKey facilitiesOther living facilitiesweightPopulation136460.50Buildings1/312220.18transportation facilities1/61/211/310.08Key facilities1/41/23130.16Other living facilities1/61/211/310.07 Table 6Exposure factor matrix and weighting values according to the AHP.Evaluation factorpopulationbuildingtransportation facilitiesKey facilitiesOther living facilitiesweightPopulation136460.50Buildings1/312220.18transportation facilities1/61/211/310.08Key facilities1/41/23130.16Other living facilities1/61/211/310.07 Exposure factor matrix and weighting values according to the AHP. The weight value of each factor and its subordinate relationship were obtained by the AHP, and the exposure index of geological disasters was obtained by combining the classification standard of exposure evaluation factors with Formula (10).5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E_i}=\sum\limits_{{j=1}}^{n} {{\omega _j}} {y_j}{S_i}{\text{ }}(j=1,2,3, \ldots )$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E_i}$$\end{document} is the comprehensive exposure index of geological disasters as the evaluation unit; i is the slope unit; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\omega _j}$$\end{document} is the weight value of the j evaluation index of the i-th slope unit; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y_j}$$\end{document} is the normalized value of the j evaluation index of the i-th slope unit. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S_i}$$\end{document} is the slope loss coefficient of the i-th slope unit. The natural break method was employed to categorize the disaster exposure of Ningyuan County into four levels: low exposure, medium exposure, high exposure, and very high exposure (Fig. 10). After the slope loss coefficient was introduced, the areas with flat terrain and no obvious threat of geological disasters, such as county towns and township centers, were effectively eliminated from the middle and high exposure grade areas, thus significantly reducing the risk misjudgment. After revision, the proportion of highly exposure areas decreased by 7%, and the proportion of very high exposure areas decreased by 6.9%. The high and very high exposure areas are distributed mainly in hilly areas such as Jiuyi Mountain, Yao Township, Shuishi town, Wanjing town and Wulong Mountain Scenic Resort, Yao Township. This area has a large relief, high population and high building density and is threatened by geological disasters over a relatively wide range. Fig. 10Exposure zoning for geological hazards in Ningyuan County (the left chart omits the exposure slope loss factor)(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Exposure zoning for geological hazards in Ningyuan County (the left chart omits the exposure slope loss factor)(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Risk assessment results Based on the “R = H×E×V” model, the risk results under the four rainfall conditions were calculated (Figs. 11 and 12), and the results showed that with increasing rainfall intensity, the risk level of geological disasters in Ningyuan County increased significantly, and the proportions of medium- and high-risk areas were 3.2%, 5.6%, 7.1% and 7.4%, respectively, indicating a continuous growth trend. The very high- and high-risk areas are concentrated mainly in low mountain areas such as the Jiuyishan–Shuishi section, Mianhuaping–Baijiaping section and Wulong Mountain Scenic Resort–Tongmuluo section. The roads in these areas are densely populated, and the terrain fluctuates greatly, which is highly consistent with the distribution areas of known geological hazards. The medium-risk areas are not only widely distributed in the above sections but also scattered in hilly areas such as Wanjing town, Zhonghe town and Qingshuiqiao town. The low-risk areas are located mainly in mountainous forestland, uninhabited areas and the center of cities and towns with relatively flat terrain. Overall, the risk spatial pattern in the study area is characterized as “high in the southeast and low in the northwest”, with significant spatial overlap between high-risk areas and population aggregation and construction areas. On the basis of “Technical Requirements for 1:10000 Geohazard Investigation and Risk Assessment in Hunan Province (Trial)”59, the field investigation for the geohazard investigation and risk assessment project in Ningyuan County was grounded in a 50-year return period rainfall scenario, which is comparable to the extreme rainfall conditions set in this study. To assess the reliability of the model, a comparative analysis was conducted between the empirical risk outcomes obtained from field measurements and the risk evaluation results of extreme rainfall scenarios derived from the coupled RF–AER–AHP model. The comparison results reveal that the classification accuracy of medium- and high-risk SUs was 93.4%, the classification accuracy of low-risk SUs was 87.2%, and the overall comprehensive accuracy was 87.6%. The high degree of consistency observed in both the spatial distribution patterns and the risk level categorization between the two models indicates that the coupled model is capable of accurately representing the spatial distribution rules and the differentiation characteristics of geological hazards in Ningyuan County. The excellent alignment of the model outcomes with the results from field investigations and observations serves as a further validation of its scientific validity and reliability. Fig. 11Geological hazard risk zones in Ningyuan County under four rainfall conditions(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Geological hazard risk zones in Ningyuan County under four rainfall conditions(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Fig. 12Geological hazard risk zones in Mianhuapingyaozuxiang under four rainfall conditions(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Geological hazard risk zones in Mianhuapingyaozuxiang under four rainfall conditions(Maps are made using ArcGIS (version: 10.3; URL link: https://www.esri.com/en-us/arcgis/about-arcgis/overview ). Conclusion and outlook This study develops an integrated geohazard risk assessment framework at the county scale by coupling Bayesian-optimized RF susceptibility modeling, multi-scenario rainfall hazard analysis, and AHP-based exposure with slope-loss correction. On the basis of the slope unit, a coupled RF-AER-AHP model was established for Ningyuan County, Hunan Province, and a risk assessment of county geological disasters was systematically carried out from three aspects: susceptibility, hazard and exposure. The research results show the following: In the susceptibility evaluation, a total of 16 evaluation indicators were selected from five aspects: topography, geological conditions, hydrological conditions, human engineering activities and other environmental factors. After high correlation factors were removed through Pearson’s correlation coefficient, the Bayesian optimized random forest algorithm was used to construct the model. The AUC value was 0.92, the true positive rate was 89.8%, and the overall accuracy rate was 85.8%. The model is highly accurate and stable for spatial identification and classification prediction of geological disasters;In the hazard evaluation, four rainfall conditions of 50 mm/24 h, 100 mm/24 h, 150 mm/24 h and 166.6 mm/24 h were established on the basis of the extreme rainfall assumption method. The results revealed that the hazard area increased significantly with increasing rainfall intensity. The proportions of the hazard area were 0%, 4.9%, 6.8%, and 8.4%, providing a quantitative basis for local governments to implement hierarchical risk management and control under different rainfall warning levels;In the exposure assessment, the weights of the five factors of population, buildings, transportation roads, other living facilities and important facility density were determined by the analytic hierarchy process, and a correction method for the slope loss coefficient was proposed to assess the exposure of the whole region. The distribution of disaster-bearing bodies was focused on areas threatened by geological disasters, effectively avoiding the phenomenon in which flat terrain areas are misjudged as highly vulnerable areas;In the comprehensive risk evaluation, the RF-AER-AHP coupled model was used to perform quantitative coupling of multidimensional factors and risk simulation under multiple working conditions. The results show that extremely high- and high-risk areas are concentrated mainly in Jiuyi Mountain-Shuishi, Mianhuaping-Baijiaping and Wulongshan-Tongmuluo and other low mountainous and hilly areas, which are highly consistent with the spatial distribution of geological disaster points. The accuracy rate of the medium- and high-risk SUs reached 93.4%, and the comprehensive accuracy rate was 87.6%. In the susceptibility evaluation, a total of 16 evaluation indicators were selected from five aspects: topography, geological conditions, hydrological conditions, human engineering activities and other environmental factors. After high correlation factors were removed through Pearson’s correlation coefficient, the Bayesian optimized random forest algorithm was used to construct the model. The AUC value was 0.92, the true positive rate was 89.8%, and the overall accuracy rate was 85.8%. The model is highly accurate and stable for spatial identification and classification prediction of geological disasters; In the susceptibility evaluation, a total of 16 evaluation indicators were selected from five aspects: topography, geological conditions, hydrological conditions, human engineering activities and other environmental factors. After high correlation factors were removed through Pearson’s correlation coefficient, the Bayesian optimized random forest algorithm was used to construct the model. The AUC value was 0.92, the true positive rate was 89.8%, and the overall accuracy rate was 85.8%. The model is highly accurate and stable for spatial identification and classification prediction of geological disasters; In the hazard evaluation, four rainfall conditions of 50 mm/24 h, 100 mm/24 h, 150 mm/24 h and 166.6 mm/24 h were established on the basis of the extreme rainfall assumption method. The results revealed that the hazard area increased significantly with increasing rainfall intensity. The proportions of the hazard area were 0%, 4.9%, 6.8%, and 8.4%, providing a quantitative basis for local governments to implement hierarchical risk management and control under different rainfall warning levels; In the hazard evaluation, four rainfall conditions of 50 mm/24 h, 100 mm/24 h, 150 mm/24 h and 166.6 mm/24 h were established on the basis of the extreme rainfall assumption method. The results revealed that the hazard area increased significantly with increasing rainfall intensity. The proportions of the hazard area were 0%, 4.9%, 6.8%, and 8.4%, providing a quantitative basis for local governments to implement hierarchical risk management and control under different rainfall warning levels; In the exposure assessment, the weights of the five factors of population, buildings, transportation roads, other living facilities and important facility density were determined by the analytic hierarchy process, and a correction method for the slope loss coefficient was proposed to assess the exposure of the whole region. The distribution of disaster-bearing bodies was focused on areas threatened by geological disasters, effectively avoiding the phenomenon in which flat terrain areas are misjudged as highly vulnerable areas; In the exposure assessment, the weights of the five factors of population, buildings, transportation roads, other living facilities and important facility density were determined by the analytic hierarchy process, and a correction method for the slope loss coefficient was proposed to assess the exposure of the whole region. The distribution of disaster-bearing bodies was focused on areas threatened by geological disasters, effectively avoiding the phenomenon in which flat terrain areas are misjudged as highly vulnerable areas; In the comprehensive risk evaluation, the RF-AER-AHP coupled model was used to perform quantitative coupling of multidimensional factors and risk simulation under multiple working conditions. The results show that extremely high- and high-risk areas are concentrated mainly in Jiuyi Mountain-Shuishi, Mianhuaping-Baijiaping and Wulongshan-Tongmuluo and other low mountainous and hilly areas, which are highly consistent with the spatial distribution of geological disaster points. The accuracy rate of the medium- and high-risk SUs reached 93.4%, and the comprehensive accuracy rate was 87.6%. In the comprehensive risk evaluation, the RF-AER-AHP coupled model was used to perform quantitative coupling of multidimensional factors and risk simulation under multiple working conditions. The results show that extremely high- and high-risk areas are concentrated mainly in Jiuyi Mountain-Shuishi, Mianhuaping-Baijiaping and Wulongshan-Tongmuluo and other low mountainous and hilly areas, which are highly consistent with the spatial distribution of geological disaster points. The accuracy rate of the medium- and high-risk SUs reached 93.4%, and the comprehensive accuracy rate was 87.6%. It should be noted that the novelty of this study does not lie in the independent use of RF, rainfall scenarios, or AHP, which have been widely reported in previous studies. Instead, the main contribution lies in the integration and refinement of the full risk chain, particularly in combining slope-unit-based susceptibility, rainfall-scenario hazard modeling, and terrain-corrected exposure within a unified county-scale framework. Overall, the coupled model effectively integrates the adaptive advantages of machine learning, the dynamic characteristics of rainfall-induced mechanisms, and the structured weight system of the AHP. It not only addresses the limitations of traditional methods, such as strong subjectivity in weight determination and insufficient dynamic response but also provides a scientific and efficient technical approach for identifying and spatially delineating geological hazard risks at the county scale. Despite the effectiveness of the proposed framework, several limitations remain. First, the rainfall scenarios used in this study are predefined and do not fully account for spatial variability in rainfall thresholds and antecedent conditions. Second, vulnerability is simplified as a constant, which may not reflect differences in building structure, population characteristics, and coping capacity. Third, the slope-loss coefficient involves a degree of empirical setting and requires further validation using real damage data. Future work should focus on integrating physically based rainfall thresholds, dynamic vulnerability indicators, and multi-source data such as UAV imagery and socio-economic datasets to improve both temporal sensitivity and physical interpretability.