Xi Ao Town is situated in the northernmost part of Longchuan County in Heyuan City, Guangdong Province, China. Belonging to the mountainous region of northern Guangdong, it lies approximately 106 km northwest of Longchuan County’s administrative center. The town encompasses a jurisdictional area of 143 km² ( Figure 1 ). The terrain is characterized by medium-low mountains with significant relief and steep slope gradients. Xi Ao experiences a typical subtropical monsoon climate, featuring concentrated heavy precipitation during the summer months. With the town comprising 12 villages and one neighborhood committee, supporting a population of approximately 23,000 residents. In recent decades

extensive slope cutting for housing construction has significantly increased susceptibility to geological disaster. This vulnerability was first demonstrated in 2022 when heavy rainfall across Longchuan County triggered numerous landslides, flash floods, and debris flows. Xi Ao Town was among the most severely affected townships during this event.

This study primarily utilized data from geologic hazard risk assessment and geologic hazard investigation and evaluation projects conducted by the Guangdong Provincial Geological Hazard Prevention and Control Technology Center. The core dataset was obtianed from the 2022 Geological Hazard Fine Investigation Project in Longchuan County, Heyuan City, Guangdong Province. This included comprehensive field survey data and relevant historical records. Data for the known and hidden hazard points were extracted from this investigation. To ensure positional accuracy and research rigor, the geographical coordinates of all actual risk and hidden hazard points were verified and reprocessed using Global Positioning System (GPS) technology. The evaluation system incorporated key susceptibility influencing factors, including elevation, slope gradient, slope aspect, lithology, distance to faults, distance to roads, distance to rivers, distance to residential areas. Specific datasets for these factors were obtained from publicly available resources provided by the Guangdong Provincial Geological Environment Monitoring Station ( http://dzj.gd.gov.cn/dzzhfzjszx/ ).

The occurrence of geological disasters is characterized by suddenness, rapidity, and uncertainty and involves numerous influencing factors. These factors can be broadly categorized into two groups: internal factors (primarily topography, geological structure, engineering rock properties, and slope structure) and external factors (notably triggering agents such as rainfall and human activity) ( Jiao et al., 2022 ). Based on comprehensive analysis of geological hazard development conditions specific to Xi Ao Town and adhering to the indicator selection guidelines outlined in the “Technical Requirements for Informatization of Geological Hazard Investigation and Evaluation of Geological Hazard Risks in Towns and Streets of Guangdong Province (1:10,000)”, this study selected the following eight evaluation indices for geological hazard susceptibility: slope, slope direction, elevation, stratigraphic lithology, distance from fracture zones, distance from roads, distance from water systems, distance from residential areas as the evaluation indexes of geologic hazard susceptibility. These factors comprehensively represent the primary internal predisposing conditions and key external anthropogenic triggers relevant to Xi Ao. Table 1 lists the specific data sources, and the classification schemes used for each index. The spatial distribution of these eight susceptibility indicators across the study area are shown in Figures 2-9 .

1) Slope

The slope gradient critically controls the occurrence and evolution of geological hazards by governing the stability and kinematic potential of the surface materials. Steeper slopes experience amplified gravitational stresses, increasing susceptibility to the destabilization of rock and soil masses. This frequently causes landslides, rockfalls, and related hazards. Concurrently, steep gradients accelerate surface runoff velocity, intensify hydraulic erosion and compromise slope integrity. These slopes typically feature reduced vegetation coverage and diminished soil water retention capacity. Rapid saturation occurs under intense rainfall, significantly elevating risks of landslides and debris flows.

2) Slope direction

The slope aspect (direction) induces differential solar exposure, which significantly alters the physical properties of rock and soil, thereby modulating slope stability and hazard potential ( Yu et al., 2023 ). South-facing (sunny) slopes exhibit heightened hazard susceptibility compared to north-facing (shady) slopes owing to distinct microclimatic effects: enhanced convective activity promotes orographic precipitation, increasing hydrological triggers for mass movements; pronounced diurnal thermal fluctuations accelerate physical weathering, reducing rock cohesion; persistent drying-rewetting cycles degrade material integrity through expansive stresses. These mechanisms collectively increase the failure probability on sun-exposed slopes.

3) Elevation

Elevation exerts multi-scale controls on hazard genesis by integrating topographic, pedogenic, and ecological influences. As a primary descriptor of terrain configuration (mountain ranges, valleys, and plains), it correlates strongly with weathering intensity and regolith thickness, groundwater dynamics, soil moisture regimes, vegetation communities, and land-use patterns ( Xie et al., 2023 ). In Xi Ao Town, elevations range from 156.79 m to 1,293.24 m ( Figure 7 ). These data were classified into five discrete intervals using zonal statistics and the Jenks Natural Breaks algorithm, optimizing within-group homogeneity for susceptibility modeling.

4) Stratigraphic lithology

Stratigraphic lithology fundamentally controls slope stability through its physical and mechanical properties ( Wu et al., 2016 ). The rock composition and structure dictate weathering susceptibility, strength anisotropy, and stress redistribution. These properties determine the failure modes: massive igneous rocks favor planar sliding, whereas unconsolidated sediments enable debris flow. Xi Ao’s Early Jurassic formations (predominantly sandstones interbedded with shales) exhibit pronounced differential weathering that is conducive to instability.

5) Distance from fracture zones

Regions with frequent development of tectonic activities imply an unusually active process of accumulation and release of internal stresses in the Earth’s crust, a phenomenon that greatly alters the stability of the geological structure. In such a region, the strong mechanical effects of mutual extrusion, tension or shear of crustal plates lead to significant deformation of the surface and underground rock layers, forming a series of complex geotectonic forms such as fracture zones and folded mountain ranges. These tectonic zones are characterized by tectonic stress concentration, rock fragmentation and active crustal movement, which make the physical and mechanical properties of rocks fragile and weaken the shear strength, thus greatly increasing the probability of occurrence of the geologic disasters ( Wu et al., 2021 ). According to the buffer zone analysis, the geological tectonic buffer distance of Xi Ao Town was divided into five categories, as shown in Figure 6 .

6) Distance from roads

Xi Ao Town is in the low hills of northern Guangdong; the villages are mainly distributed in rivers and valleys, and the house buildings are generally built adjacent to the mountains. During the construction of roads in mountainous areas, excavation of mountain slopes frequently occurs, which directly affects the original stability of mountain slopes, increases the risk of geological disasters, and lurks in landslides, collapses, and other safety hazards ( Wang et al., 2020 ). In the process of road construction and operation, the disturbing effect on the slope land on both sides of the road may destroy the balance and stability of the slope soil body, thus inducing geological disasters. According to the distance from the roads, the road slope is affected by different degrees, and the road buffer distance in Xi Ao Town was divided into five categories, as shown in Figure 7 .

7) Distance from water system

The cutting effect of rivers and gullies creates suitable spatial conditions for the occurrence of geological hazards and becomes the main occurrence of geological hazards. The cutting depth and density of rivers and gullies in Hoshi Au Town significantly affect the distribution density of geological hazards. The slopes on both sides of the river are subject to the river’s infiltration, scouring and side erosion all year round, which weakens the friction between the soil bodies, anti-slip ability declines, and cracks are formed; thus, that soil body loses its support, destroys the balance and triggers geological disasters ( Li et al., 2022 ). The closer the water system is to the slope, the greater the influence on the stability of the slope, and the more likely geological disasters will occur.

8) Distance from the residential area

The study area is a low mountainous and hilly area in northern Guangdong. Villages in mountainous areas are mainly distributed along rivers and valleys, and houses are mostly built on the mountains. Unreasonable excavation or slope

cutting at the foot of mountain slopes leads to the loss of the original stability of the soil body of mountain slopes, which creates favorable conditions for the occurrence of geologic hazards. The degree of influence on geological hazards also varies. depending on the distance from the settlement.

IVM is a statistical method, based on the Bayesian probability model, which is mainly based on the informativeness of the influencing factors to quantitatively measure the relevance of each influencing factor to the research object ( Ruan & Huang 2001 ). According to the classification results of each evaluation index, the number of raster and the number of geological hazards in different classification states of each evaluation index were counted using the GIS software, and the information quantity value of each evaluation index in Xi Ao Town was calculated by applying formula (1), and then the value of each evaluation index was assigned.

$I={\displaystyle \underset{i=1}{\overset{n}{\sum }}I\left({\chi }_i,H\right)}={\displaystyle \underset{i=1}{\overset{n}{\sum }}\ln \frac{N_i/N}{S_i/S}}$ (1)

I denotes the value of informativeness, which is mainly used to reflect the possibility of destruction of evaluation cells; n represents the number of indicators, H denotes the number of risk points and hidden dangers, S denotes the number of grid cells. N denotes the total number of evaluation cells that have undergone deformation and destruction; Si denotes the number of grid cells of indicator x_i, N_i denotes the number of indicators x_i that have suffered from environmental geohazards number of grid cells. When I > 0, the greater the amount of information, the greater the probability that geological disasters may occur; when I < 0, it indicates that the indicator is unfavorable to the occurrence of geological disasters; when I = 0, it indicates that there is no connection between the indicator and the occurrence of geological disasters, that is, the indicator cannot be adopted as an indicator for the comprehensive evaluation of the susceptibility to geological disasters in Xi Ao town ( Zhu et al., 2021 ).

Artificial Neural Networks (ANNs) are highly nonlinear, ultra-large-scale continuous-time dynamical systems formed by extensively interconnected processing units (neurons) ( Shao et al., 2011 ). ANNs employ a multilayer structure incorporating bias and negative feedback mechanisms, with training systems involving both positive and negative feedback processes. During the positive feedback phase of network training, the input data are processed by the hidden layer and transmitted from the input layer to the output layer to generate preliminary prediction results. These outputs were then compared with the expected target values. When deviations occurred, the system activated a negative feedback mechanism. At this stage, the errors propagated backward along the network path. Through iterative adjustment of inter-neuronal weight coefficients, discrepancies between predictions and targets are progressively reduced or eliminated. This process continuously optimizes network performance until predetermined requirements are met. In this study, we applied neural network analysis to partition geological hazard susceptibility results for Xi Ao Town. Our network model ( Figure 10 ) comprises input, hidden, and output layers with full interlayer connectivity. Network training was achieved through continuous adjustment of weights and thresholds ( Guo et al., 2000 ).

RBF-NN is a special three-layer structure ( Figure 11 ): the input layer is responsible for receiving external input signals, and consists of multiple source points or perceptual units, which convert the data of the actual problem into a form that can be processed by the network; the hidden layer adopts RBF, and this layer performs the very important nonlinear mapping function, which transforms the data points in the input space into a new implied feature space ; the output layer is linear, and based on the result of the nonlinear transformation of the hidden layer, it directly responds to the input patterns after linear combination, which leads to

the final network output ( Gu, 2013 ). Its input/output relationship can be described by the following equation (2):

$\{\begin{array}{c}{u}_i=\exp \left[-\frac{\Vert X-c_i\Vert }{2{\sigma }_i^2}\right]\\ y={\displaystyle \underset{i=1}{\overset{n}{\sum }}{w}_{ij}{u}_i-\theta }\end{array}i=1,2,\cdots ,n;j=1,2,\cdots ,n$ (2)

X = (x1, x2, ..., xn) denotes the input data; c_i denotes the central value of the function; σ_i denotes the constant required for the study, which determines the distance of the base function to the central value; n denotes the number of neurons in the hidden layer of the network model; w_ij denotes the weight of neuron i in the hidden layer of the network model linking neuron j in the output layer of the network model; θ denotes the output layer of the network model neuron critical value; y denotes the output content of the network model.

The learning process of the RBF-NN corresponds to the training process, involving inputting training data into the network model for autonomous learning. Key characteristics of this model include its capacity for self-optimizing parameters, thereby reducing manual parameter adjustment workload and lowering research costs. This capability enables the network model to achieve enhanced performance across diverse datasets and experimental tasks. Simultaneously, it reduces zoning errors introduced by human parameter adjustments during the partitioning of geological hazard susceptibility evaluation results, while improving partitioning efficiency.

In the process of geohazard evaluation research and practice, the evaluation unit is defined as the smallest unit for geohazard susceptibility evaluation, which is the spatial carrier of geohazard danger, vulnerability, susceptibility and risk measurement, and the basis for the implementation of disaster prevention and mitigation measures and planning and management. Whether the selection and division of the evaluation unit is reasonable is crucial to the relevance, scientific natural and operability of the evaluation results. Currently commonly used evaluation units can be regular shapes, such as grid cells or grid cells divided into certain sizes, or shapes of different sizes, such as slope units, watershed units or regional units ( Gao et al., 2006 ). Considering that this study area is characterized by a small range and small amount of data, the grid cell was selected as the basic evaluation unit to evaluate susceptibility. The grid cells were in accordance with the accuracy requirements, and the size of the grid was calculated with reference to the empirical formula ( Xue et al., 2018 ):

$G_s=7.49+0.0006s-2\times {10}^{-9}\times s^2+2.9\times {10}^{-15}\times s^3$ (3)

s is the scale of the map of the study area, i.e., 1:10,000, which is calculated to be G_s = 13.3, and thus the size of the grid is determined to be 12.5 m × 12.5 m.

First, it is necessary to quantify all evaluation indicators and then input them into the network model in the form of numerical values for calculation. Only the data after quantitative processing can be inputted and calculated by the network model; therefore, it is necessary to quantify all evaluation indicators ( He, 2014 ). To comprehensively and accurately assess the geohazard susceptibility of each evaluation unit, it is necessary to consider the specific state of each influencing factor within the unit and its interaction relationship, and to calculate the comprehensive information quantity under the combined situation of various evaluation factors ( Zhu et al., 2021 ). The information quantity method was used to analyze and obtain the information quantity calculation results of eight geological hazard susceptibility indicators in Xi Ao Town and assign their values ( Table 2 ).

According to the spatial distribution pattern and informativeness analysis of each indicator in the evaluation index system, the eight selected influencing factors were causally related to susceptibility to geological disasters in Xi Ao Town. According to the 300 geological disaster hidden spots collected in Xi Ao Town, the information value was calculated, and the indicators were graded using the equal spacing or natural breakpoint method to determine the magnitude of the influence of each evaluation indicator and sub-interval on the geological disaster. The information value and variance of the information value of the evaluation indicators of geological disaster susceptibility in Xi Ao Town are shown in Table 3 .

By analyzing the informativeness values, the susceptibility to geological hazards in Xi Ao Town was found to be most influenced by four factors: distance from settlements, distance from roads, elevation and slope. Based on the variance of the informativeness values, it can be concluded that the four indicators of distance from settlements, elevation, slope, and distance from roads play a significant role in influencing susceptibility to geological hazards in Xi Ao Township. Finally, the ratio of the informativeness variance of the eight evaluation indicators of Xi Ao town to the total informativeness variance was used to indicate the contribution of each evaluation indicator in the evaluation of the susceptibility to geological hazards in Xi Ao town: the higher the contribution, the higher the influence of the indicator on the degree of occurrence of geological hazards. As shown in Table 4 , the three indicators of elevation, distance from settlements and slope had a greater influence on the degree of geological disaster occurrence, and stratigraphic lithology had the least influence.

The geological disaster susceptibility zoning assessment was conducted by integrating the informativeness Method and the ANN Model. As the ANN model requires prior training before application, the evaluation process is divided into two parts: 1) training the neural network model and 2) susceptibility assessment and zoning ( Gao et al., 2006 ).

1) Neural Network Model Training

Training the neural network model serves as a foundational step in geological disaster susceptibility assessment. Parameter optimization is a prerequisite to ensure model reliability, which can be achieved either through standalone calibration procedures or integrated with susceptibility zoning workflows. The optimization process involves five critical stages: Construction of optimization indicator layers: Integrate geospatial parameters derived from the Information Value Model (IVM); Grid-based spatial discretization: Divide the study area into 12.5 × 12.5 m resolution grids using GIS; Data extraction: Compile training datasets from historical disaster inventories and environmental variables; Data standardization: Normalize input variables (mean = 0, SD = 1) to eliminate dimensional heterogeneity; Model optimization: Implement a hybrid genetic algorithm-backpropagation approach to minimize prediction errors.

2) Susceptibility zone

Following model validation, the optimized ANN is deployed for susceptibility zoning through a parallelizable workflow: indicator layer development: establish multi-criteria evaluation layers; grid system implementation: maintain spatial consistency with the training phase using identical grid specifications; predictor variable extraction: execute batch processing of spatial data through Python-GIS interfaces; data normalization: apply min-max scaling to ensure comparability across heterogeneous parameters; susceptibility classification: generate probabilistic susceptibility maps (0 - 1 scale) and classify zones using quantile thresholds.

After getting the quantized training data, use the RBF neural network object created in MATLAB and set up the training parameters, then call the function to train the model, after setting up the training parameters, you can point to “start training” ( Dong et al., 2024 ). The simulation training of the RBF neural network model used in this study is the process of calling functions that have been integrated into the neural network toolbox of the MATLAB software ( Figure 12 ). The length of this simulation training is closely related to the size of the data volume of the input training samples and the settings of the parameters before training, especially the number of neuron increments, the maximum number of neurons allowed, the error of the neural network, and the expansion speed of the model have a greater impact. The input training sample data in this study accounted for 1% of the total data volume of Xi Ao Town, that is, 973 grid cells.

The spread of the newrbe function of the RBF-NN model has a close relationship with its training accuracy, and it needs to be repeatedly debugged until the model operation results reach the optimization in the actual running process of the model. The debugging of the model operation needs to be set according to a certain rule of different sizes of the expansion speed to obtain the optimal training data, the expansion speed of the model operation is generally in the range of 0.2 - 5, and the value of the expansion speed is the number of the first term of the isotropic series with the first term of 0.2 and the common difference of 0.2. At the end of the model training, the training error control and the training accuracy analysis graph ( Figure 13 ) were made by comparing the written down values of the spreading with the corresponding values of the training accuracy.

By analyzing the table corresponding to the model expansion speed value and training accuracy, it is concluded that only when the expansion speed of the neural network model is 0.4, the accuracy of the neural network model training is the highest and at the peak. Therefore, in this study, the neural network model expansion speed of 0.4 was selected as the final parameter of training to complete the training of the network model. After the training of the neural network model was completed, the index data of Xi Ao Town, which is similar to Xi Ao Town in terms of natural environment, and the finalized parameters of the neural network model were used to carry out the evaluation of the susceptibility of Xi Ao Town to geologic hazards and partitioning, and finally, the results of the evaluation were superimposed on the registered historical geologic hazard sites to verify the accuracy of the neural network model.

Partitioning of the geohazard susceptibility evaluation results constitutes the final stage of the assessment. This process involves two key components: model computation and conversion of the evaluation results into spatial layers. Using the trained RBF-NN model and processed index data, we implemented MATLAB’s integrated RBF-NN function to calculate and output susceptibility values for Xi Ao Town. These numerical outputs directly correspond to susceptibility levels according to Guangdong Province’s “Technical Requirements for Informatization of Geological Disaster Risk Survey and Evaluation (1:10,000) in Townships (Streets) (Trial Implementation)”. Xi Ao Town is classified as high susceptibility, medium susceptibility, low susceptibility, and non-susceptibility, and the corresponding output result values were 4, 3, 2, and 1. The output result values were obtained through the calculation of the RBF-NN model and were imported into the GIS for the calculation of the output results and imported into the GIS for conversion to derive the geohazard susceptibility zoning results of Xi Ao Town ( Table 5 , Figure 14 ).

1) High susceptibility zone

The high susceptibility area of geological hazards in Xi Ao Town covers 1.14 square kilometers (0.76% of the township’s area), primarily distributed in the northern part of Guihu Village and the northern/southeastern parts of Yongan Village within high-altitude hilly areas. These areas feature elevation above 321.1 meters, slopes of 55˚ - 75˚, and sunny/semi-sunny aspects. Vulnerable to collapse and rockfall disasters when affected by Guangdong’s “Dragon Boat Water” continuous rainfall and typhoon events.

2) Medium-susceptibility zone

Medium susceptibility areas cover 13.86% of Xi Ao Town, widely distributed across mid-low altitude gentle slopes (156.6 - 321.1m elevation, ≤15˚ slopes, semi-sunny aspects) within 50 - 100 m of roads and 50m of rivers. These zones accumulate slope toe materials and require preventive monitoring against hazards triggered by heavy rainfall.

3) Low susceptibility zone

Low susceptibility zones account for 84.52% of the township, mainly occurring at lower elevations with moderate slopes, shady aspects, and significant distances from roads/rivers. A geologically stable environment presents minimal risk but necessitates continuous stability monitoring for early warning.

4) Non-susceptibility zone

Non-susceptibility areas (0.87% of townships) occur in remote locations far from settlements/rivers/roads with minimal human impact and stable geological conditions, serving as designated safe zones and hazard avoidance sites.