Landslide dynamic hazard prediction based on precipitation variation trend and backpropagation neural network

Abstract

The assessment of landslide hazards is crucial for disaster prevention and mitigation, but it has not considered the dynamic influencing factors that trigger landslides. The timeliness and practical value of the assessment results still need to be further improved. This study constructed a dynamic landslide hazard assessment system using information value model, dynamic precipitation data, and Backpropagation Neural Network (BPNN) model. Taking the Qingjiang Reservoir landslide in Changyang County, Hubei Province, China as an example, based on dynamic precipitation data and the BPNN model were used to develop a dynamic landslide hazard prediction model, and the temporal assessment and spatial distribution results of slope unit hazards in the study area from the 1980s to the 2010s, 2025, and 2030 were evaluated and predicted. It is predicted that the percentage of very high and high areas in 2025 and 2030 will be 50.5% and 57.5% respectively.

Keywords:

• Landslides
• dynamic hazard assessment
• precipitation variation trend
• BPNN
• remote sensing

1. Introduction

With the continuous expansion of urban scale and the impact of extreme weather, landslides have become exceedingly common and extremely harmful natural disasters worldwide (Petley 2012), significantly impacting people’s production and livelihood (Froude and Petley 2018). Determining the range of landslide-prone areas holds significant meaning in reducing economic losses and casualties brought upon by such disasters (Arabameri et al. 2019; Pardeshi et al. 2013; Peruccacci et al. 2023). From 1940 to 2020, it is estimated that 14,394 people died of landslides in China (Zhang et al. 2023). According to the statistical data of China’s Ministry of Natural Resources (https://www.mnr.gov.cn/), there were 5659 geological disasters such as landslides, collapses, and debris flows in 2022, resulting in a total of 90 deaths, 16 missing people, 34 injured people, and direct economic losses of 1.5 billion yuan. Among them, 3919 landslides occurred. Under these circumstances, conducting a landslide catastrophe investigation and assessment system as well as enhancing and promoting research in disaster prevention and mitigation remains a topic of keen interest for various countries.

Carrara, A. (Carrara et al. 1995) analyzed the landslide susceptibility of slope units in small watersheds in Italy and defined the landslide susceptibility assessment. Landslide susceptibility assessment is the basis of landslide hazard assessment and landslide risk assessment, which reflects the spatial probability of the slope evolving into a landslide under the action of its basic environmental factors (Griffiths et al. 2002; Guzzetti et al. 2005; Huang et al. 2020; Huang et al. 2021; Jun et al. 2016; Li et al. 2020; Wang et al. 2024). The research system of landslide susceptibility assessment is widely used to solve the threat of landslides to people’s lives, property safety and social and economic development (Baeza and Corominas 2001; Gokceoglu 2002). Landslide hazard assessment is based on the assessment of landslide susceptibility, considering the dynamic influence of landslide induced factors on landslide, and predicting the possibility of landslide hazard (Fell et al. 2008; Tyagi et al. 2022). In recent years, with the gradual improvement of the landslide assessment system, the research on landslide hazard assessment has also increased year by year. As a key step in predicting landslides and formulating disaster prevention and mitigation policies, landslide hazard assessment has become a subject of in-depth study by scholars (Singh and Kumar 2018; Tan et al. 2021; Versain et al. 2019). Moreover, it serves as a significant component of international disaster risk reduction strategies, playing a vital role in disaster management. Through this approach, decision-makers can assess potential hazards, thereby reducing human and property losses (Hader et al. 2022; Shahzad et al. 2022; Skrzypczak et al. 2021). In the subtropical monsoon climate area, a short period of heavy precipitation is easy to cause landslides (Zeng et al. 2023). However, from the content of landslide hazard assessment, over the years, the existing research mainly focuses on the selection of relatively stable geological environment factors and triggering factors for landslide hazard assessment (Basharat et al. 2023; Wang et al. 2014; Yang 2017). At present, landslide hazard assessment based on precipitation factors is mainly based on landslide susceptibility assessment, and the calculated precipitation thresholds (Huang et al. 2022) or extreme precipitation in different precipitation return periods (Kawagoe et al. 2010) is the disaster-causing factor, and the result is a static assessment under specific precipitation conditions. Because the factors that trigger landslide disasters such as climate and precipitation have dynamic trends, the hazard of landslide disasters will also change dynamically (Sui et al. 2022; Tang et al. 2017). It is necessary to add the evolution process of precipitation dynamic changes to the prediction of landslide hazard.

Significant progress has been achieved in the field of landslide hazard assessment due to the continual advancements in Geographic Information System (GIS), Remote Sensing (RS), and Global Positioning System (GPS) (Pan et al. 2021; Wang et al. 2022). With the deepening of research, various statistical models based on GIS technology are becoming more and more diversified, which can be summarized as heuristic models, mathematical statistical models and machine learning models. The heuristic models and mathematical statistical models that have been widely used are the expert scoring method, analytic hierarchy process, information value method, frequency ratio method, etc. Yang used the expert scoring method (Yang et al. 2023), Panchal & Shrivastava (2022) and Zangmene (Zangmene et al. 2023) used the analytic hierarchy process to assign weights to different disaster-causing factors and draw landslide hazard maps. Afungang (Afungang et al. 2017) used the information value model to link the geological environment factors with previous landslides and assessed the hazard of landslides that may occur in the future in Bamenda highlands. Wubalem and Meten (Wubalem and Meten 2020) used the information value model to provide a more accurate prediction of landslide hazard in an area of northwest Ethiopia. Khan (Khan et al. 2019) studied the impact of disaster-causing factors on the spatial distribution of landslides and used the frequency ratio method to draw a landslide susceptibility zoning map in northern Pakistan. With the advancement of computer technology, machine learning models have gradually been used to study landslide hazard assessment. Lee selected 17 factors that have a greater impact on the occurrence of landslides and used two methods of support vector machine and artificial neural network to predict the susceptibility of landslides in the study area (Lee & Lee et al. 2017). Based on the support vector machine method, Lee analyzed the landslide susceptibility of two different regions in South Korea (Huang and Zhao 2018; Lee, Hong, et al. 2017). Niu used BPNN to construct an intelligent safety early warning system (Niu 2020), and Mehrabi used a multilayer perceptron neural network to predict landslide susceptibility in a certain area of Italy (Mehrabi 2022). Kim used random forest and boosted tree models to analyze and verify the susceptibility of landslides in the Pyeong-Chang area (Kim et al. 2018). Through these machine learning approaches, researchers can conduct decision analysis on the probability of landslide occurrences and effectively enhance the reliability of landslide hazard assessment (Karakas et al. 2023; Yu and Chen 2020; Zhao et al. 2022).

The traditional mathematical statistical model mainly analyzes the historical landslide, determines influencing factors, and calculates the probability of landslide occurrence based on these influencing factors; the machine learning model learns complex rules from data without clarifying the correlation between various influencing factors, and predicts landslides based on training samples (Zou et al. 2023). On this basis, the mathematical statistical model and machine learning model are used to assessment the hazard of landslides, which can not only predict the relationship between the occurrence of landslides and various influencing factors but also accurately predict the hazard of landslides. However, the method of combining mathematical statistics with machine learning for landslide hazard assessment is not widely used in current research and application.

Based on the above analysis, to identify potentially hazardous areas for landslide disasters under the influence of dynamic factors and assist decision-makers in better preventing and responding to landslide disasters (Xu et al. 2019; Yu et al. 2022), this paper fully considers the dynamic impact of climate change on the hazard of landslide. Located in Changyang County, Hubei Province, this study examines landslide disasters along the Qingjiang River. Based on the assessment of landslide susceptibility using the information value method, the precipitation from the 1980s to 2010s and the predicted precipitation in 2025 and 2030 were introduced as triggering factors, and the dynamic prediction model of landslide hazard was constructed by BPNN. The dynamic hazard of landslides in the study area in the past few decades was analyzed to predict the dynamic hazard of landslides in the next 5-10 years. The research findings not only be used as a reference for the system of geological disaster hazard assessment in mountainous areas but also offer scientific decision-making and guidance for effectively preventing and mitigating mountainous landslide disasters in Changyang County.

2. Study area

Along the middle and lower reaches of the Qingjiang River, the research area is situated in Hubei Province, China. It has a strip-like distribution along the mainstream of the Qingjiang River, spanning 42 km east to west and 12 km north to south, with a total area of 504 km². Over 90% of the research area consists of mountainous terrain, characterized by well-developed valleys, and the overall direction of rivers and mountain ranges is predominantly east-west. The research area is situated at an elevation ranging from 141 to 1302 meters above sea level, with a general topography showing higher elevations in the north-south direction and lower elevations in the east-west direction, as illustrated in Figure 1. The stratigraphic lithology of the study area is mainly composed of limestone of the Lower Triassic Daye Formation (T₁d), limestone of the Upper Permian Dalong Formation (P₃d), and mud shale of the Wu Jiaping Formation (P₃w). The Carboniferous (C_1-2) and Devonian (D) sandstone and siltstone strata are secondary. The Quaternary is mainly a colluvial deposit (Q₄^{dl + col}), which is composed of gravel, sandy soil, and clayey soil fragments. Water resources are abundant in the research area which belongs to the subtropics monsoon climate. The average annual precipitation is approximately 1366.2 mm, with over 75% of the total amount falling from April to September. Seasonal heavy rainfall is the main triggering factor of landslides in Qingjiang Reservoir.

Figure 1. Study area location and distribution of landslides.

Through the high-resolution remote sensing interpretation, unmanned aerial vehicle photography interpretation, and field investigation of the Hubei Geological Bureau, a total of 43 landslides occurred in the Qingjiang Reservoir from the 1980s to 2010s, of which 23 were fording landslides, which account for 53.5% of all the landslides in the area. Influenced by the overall slope structure distribution with a predominantly dip slope on the left bank and a predominantly reverse slope on the right bank of the Qingjiang River segment in the research area, landslides mainly occur on dip slope and skew slope. The number of landslides gradually increases from west to east, with a higher density along the Qingjiang River coast from Yu Xiakou Town to Ziqiu Town (Figure 1). The landslides in the research area are mainly accumulation landslides, with small to medium scales. The material source is primarily composed of cohesive soil, gravel soil mixed with fractured limestone blocks, and most of them are traction-type soil landslides.

3. Methods

Figure 2 illustrates the research technical process employed in this study. Firstly, a static susceptibility assessment index system for landslides is established using the information value model. Using 13 influencing factors including elevation difference, slope degree, slope aspect, slope structure, total curvature, lithology, distance to fault, distance to river, distance to road, normalized difference vegetation index (NDVI), topographic wetness index (TWI), stream power index (SPI), and depth of overburden. The assessment is conducted on individual slope units, and a susceptibility zoning map is generated. Secondly, dynamic precipitation factors are introduced considering the geological and meteorological-hydrological conditions of the study area. The landslide dynamic hazard assessment is calculated by BPNN. Lastly, the accuracy of both susceptibility and hazard assessment results is verified through ROC and field investigation and validation.

Figure 2. Flow chart of this study.

3.1. Determine the mapping units

Before the landslide susceptibility and hazard assessment, the study area should be divided into reasonable mapping units. As the fundamental units for data extraction and index assignment, mapping units reflect both homogeneity within units and heterogeneity between units. Hence, the precision of mapping units is heavily reliant on the accuracy of input data and the resulting assessment outcomes (Sun et al. 2020). Grid units, slope units, geomorphologic units, and watershed units are examples of frequently used mapping units (Ba et al. 2018). The grid unit is the most commonly used landslide disaster unit, which has the characteristics of constant shape, easy calculation, and easy sampling (Reichenbach et al. 2018). However, each grid unit often contains a variety of different geological structures, because there is no direct physical relationship between geomorphological information and grid units (Chang et al. 2023). From the point of view of topography, the slope unit can well contain the topography conditions of the study area. Each slope unit has both internal identity and different characteristics from other slope units, which well reflects the relationship between the mapping units and the topography (Liu et al. 2022). Topography features play a crucial role in the formation and occurrence of landslides, and slope units capture the topographic characteristics of the study area effectively. The study area is a medium-low mountain landform. Qingjiang River passes through its territory, forming a unique gully landform area. The mountains on both sides of the valley are steep, and the slopes are mostly steep. Therefore, the slope units are selected as the mapping units of this paper. Each slope unit exhibits internal homogeneity and external heterogeneity when compared to other slope units, thus reflecting the relationship between mapping units and terrain features. In the division of slope units, the current representative division methods are the hydrological analysis method and curvature watershed method (Carrara et al. 1995; Guzzetti et al. 1999; Kai et al. 2020; Wang et al. 2017)

In this paper, the hydrological analysis method with a good division effect and moderate workload is used to divide the slope unit. This method takes the ridge line and the valley line as the boundary of the catchment area, and takes the middle of the ridge line and the valley line as the slope unit. This method considers both the flow movement and the topographic factors so that it can better reflect the spatial distribution of the slope unit in the gully landform area. It avoids the defect that the valley line cannot be identified in the plain and river areas. The digital elevation model (DEM) (https://search.asf.alaska.edu/#/) with a spatial resolution of 12.5 m is used to divide the slope unit in ArcGIS software. Positive and negative topographic techniques were used to extract ridge and valley lines, while hydrological analysis techniques were used to identify slope units. The resulting watersheds and reverse watersheds were merged to acquire slope units. Finally, slope units were adjusted based on mountain shadows, remote sensing images, slope aspect, and lithology to obtain a more realistic slope unit.

3.2. Factors causing landslides

Based on the unique gully landform characteristics of the study area, based on the basic types and spatial and temporal distribution of landslides, this paper divided the influencing factors into static and dynamic parts from the availability (Table 1), operability and scientificity, and the static susceptibility and dynamic hazard assessment system of landslides are established respectively.

Table 1. Source of landslide influencing factors in this research.

Data classification	Factors	Data source	Data type	Assignment method of slope units
Landslides data	Landslides inventory	Department of Geological Survey
Static influencing factors	Elevation Difference	DEM calculation	Continues	Majority
	Slope Degree		Continues	Majority
	Slope Aspect		Categorical	Majority
	Curvature		Continues	Majority
	Slope Structure	Field investigation	Categorical	Majority
	Distance to fault	Geological map of Changyang County	Continues	Average value
	Lithology		Categorical	Majority
	Depth of overburden	Field investigation	Continues	Majority
	TWI	DEM calculation	Continues	Majority
	SPI		Continues	Majority
	NDVI	Landsat-8 (202107)	Continues	Majority
	Distance to river	Geological map of Changyang County	Continues	Average value
	Distance to road	Topographic map of Changyang County	Continues	Average value
Dynamic influencing factors	Average annual precipitation in 1980s-2010s, 2025 and 2030	Data Centre for Resource and Environmental Sciences, Chinese Academy of Sciences	Continues	Majority

3.2.1. Static influencing factors

The damage of landslides was heavily affected by topography, engineering geology, hydrology and human engineering activities. According to the characteristics of the study area, the landslide static influencing factors are divided into four categories and 13 specific indicators, which are used for landslide susceptibility mapping.

The selection analysis of each influencing factor is as follows: (1) topography: the greater the elevation difference and slope degree, the higher the sliding speed of the landslide formed, and the greater the impact on slope degree damage. The slope aspect has a local influence on triggering slope instability, such as the effects of sunlight (vegetation and cut slope building) and precipitation infiltration (soil moisture) (Chawla et al. 2019), and most landslides in the area always occur in the south. Total curvature reflects the geometry within slopes, which is a direct relation towards the efficiency of precipitation infiltration. Slope structure refers to the difference between slope direction and inclination of the bedrock of the slope body, and the density of landslides is highest on dip slope and skew slopes. (2) engineering geology: due to the more fragmented rock and soil in the range affected by the fault fracture zone, the number of landslides is relatively large (Das et al. 2023). The higher mechanical strength of rocks and soils that make up the landslide body, the less likely the landslide will occur. At the same time, the thicker the loose accumulation body on slopes, the obvious stress concentration inside slopes, and the lower stability of slopes. (3) hydrology: TWI is an indicator that can be utilized to demonstrate effect of the terrain on runoff direction and accumulation, reflecting the spatial distribution characteristics of soil moisture. SPI can represent distribution and rate of water flow, with lager values representing more erosive force of water flow in the area. NDVI is a measure of vegetation density on slopes, which plays a crucial role in soil strength. The closer slopes are to rivers, the more likely slopes toe erosion will occur, thereby increasing the probability of slope failure (Yan et al. 2022). (4) human engineering activities: the high and steep cut slopes caused by road construction are the most significant factors impacting slope stabilization.

After processing, each of various influencing factors were projected into the Word Geodetic System 1984 (WGS84) coordinate system and formatted into 12.5 m resolution grid data. This generated multi-source spatial data that could be used in spatial overlay and modeling manipulation, preparing for subsequent conversion into slope units. When using slope units as mapping units, each influencing factor can only have one value within a slope unit. Therefore, different statistical methods need to be adopted for continuous and categorical influencing factors (Chang et al. 2023; Sun et al. 2023). Except for distance to road, fault and river, which are the average distances for each grid within the slope unit, the remaining influencing factors are the majority number of the influencing factors of each grid within the slope unit. Finally, the natural break method (Wubalem 2020) is used in classifying each influencing factor.

3.2.2. Dynamic influencing factors

In the research area, there were 14, 12, 8, and 9 landslides that occurred from the 1980s to 2010s, respectively. Based on completing the static susceptibility assessment of landslides, the dynamic impact factor system is formed by introducing dynamic precipitation factors, in conjunction with the static susceptibility influencing factors, to be used for mapping the dynamic hazard of landslides. Precipitation data was downloaded from the Data Centre for Resource and Environmental Sciences, Chinese Academy of Sciences (Xu 2017). The type of data is average annual precipitation (year/mm), and precipitation data for the periods of 1980s-2010s and 2010-2022 are downloaded.

The downloaded vector data were formatted into 12.5 m resolution raster data in ArcGIS and projected in the WGS84 coordinate system, and the average annual precipitation data from 2010 to 2022 were extracted based on the random points generated by ArcGIS to form the extraction points, and the time series forecasting analysis (Liu et al. 2011) was performed by Statistical Product and Service Solutions (SPSS) software to obtain the average annual precipitation in 2025 and 2030. Then, the predicted precipitation data for the study area in 2025 and 2030 is obtained using Inverse Distance Weighted (IDW) (Sheng et al. 2021) interpolation in ArcGIS software in Figure 3. The majority of the precipitation data raster was assigned to slope units, and finally, the precipitation factors were categorized into five classes using the natural break method. The precipitation data from the 1980s-2010s in the study area were used to perform the dynamic landslide hazard analysis for the past decades, and the precipitation data from 2025 and 2030 predicted to be obtained were used to predict the dynamic landslide hazard for the next 5-10 years.

Figure 3. Flow chart of average annual precipitation forecasts for 2025 and 2030.

3.3. Information value model

The formation of geological disasters is affected by many factors. Information value model (Lin et al. 2021) reflects how the most disaster-causing factors and their subdivision intervals are combined in a certain geological environment. Specifically, it is achieved by comparing the rate of occurrence of geological hazards with the frequency of regional geological hazards under the action of a certain factor in a specific assessment unit. The assessment steps and basic principles of the information value method are as follows (Ji et al. 2023):

1. Information value formula can be expressed for each factor under different conditions as follows:

(1) $$\begin{array}{c}{I}_{\mathit{\text{Aj}}}=\mathrm{\text{ln}}\hspace{0.17em}\frac{N_j/N}{S_j/S} \left(j=1,2,3,\dots ,\mathrm{n}\right)\end{array}$$(1)

In the formula, I_Aj represents amount of information of landslide disaster under the condition of factor A in j, N_j is the number of slope units of landslide disaster distribution of factor A in j state, and N is the number of slope units with landslide disaster distribution in the survey area. S_j is the number of slope elements containing A in j state, and S represents number of total slope units in investigation area.

• Under the combined effect of various state factors, total landslide hazard information of a slope unit can be defined by following formula:

(2) $$\begin{array}{c}{I}_i=\sum _{i=1}^n\mathrm{\text{\hspace{0.17em}ln}}\hspace{0.17em}\frac{\mathit{\text{Nj}}/N}{\mathit{\text{Sj}}/S} \left(j=1,2,3,\dots ,\mathrm{n}\right)\end{array}$$(2)

In the formula, I_i indicates total geological disaster information for a slope unit, and n is the number of influencing factors. The sum of information from a slope unit is used as a comprehensive index to evaluate whether a landslide disaster occurs. The greater the value, the more susceptible the slope unit is.

3.4. BPNN model

In this study, BPNN in an artificial neural network is used to help realize the landslide risk assessment. The BPNN was initially proposed by Rumelhart and McClelland as a multilayer feedforward neural network using an error backpropagation algorithm for training (Pham et al. 2019; Rumelhart et al. 1986; Zhang et al. 2021). The BPNN is segmented into three layers: input layer, hidden layer, and output layer. Neurons in neighboring layers are connected, but neurons in every layer are not continuous. A BPNN can accurately represent any continuous function. During the forward pass, the learning signal enters the input layer, undergoes data processing in hidden layer, and produces output in output layer. When the output does not match the desired result, network enters backpropagation of error mode. After propagating backward to the hidden layer, the error sequentially returns to the input layer. During the error backpropagation process, the error is evenly distributed among the hidden layer units, and the weights of each neuron are modified based on the error signal. This process of forward and backward error propagation continues until the output reaches the desired result, and the training stops. A learning process (including training and verification) and a prediction procedure comprise the BPNN model.

1. Establishment of the model

   MATLAB software is used to program the BPNN model in this paper. Three-layer topology was chosen by the BPNN model in this study: an input layer, a hidden layer, and an output layer. Assume that there are m neurons in the input layer, n neurons in the hidden layer, and l neurons in the output layer. The number of landslide dynamic influencing factors (m = 14) is the number of neurons in the input layer of the BPNN model. Using f(x) as the activation function of the hidden layer and g(x) as the activation function of the output layer, the mathematical equation of BPNN can be expressed as:

(3) $$\begin{array}{c}{Y}_i=f\left(\sum _{i=1}^m{w}_{\mathit{\text{ij}}}{X}_i+b_j\right)\end{array}$$(3) (4) $$\begin{array}{c}f\left(x\right)=\frac{e^x-e^{-x}}{e^x+e^{-x}}\end{array}$$(4) (5) $$\begin{array}{c}{Y}_k=g\left(\sum _{j=1}^n{w}_{\mathit{\text{jk}}}{Y}_i+b_k\right)\end{array}$$(5) (6) $$\begin{array}{c}g\left(x\right)=x\end{array}$$(6)

In the above formula, Y_i represents the output of the i neuron in the hidden layer, X_i is the output value of the i neuron in the input layer, w_ij represents the connection weight between the input layer and the hidden layer, and b_j is the threshold of the j neuron in the hidden layer. Y_k is the output of the k neuron in the output layer, w_jk represents the connection weight between the hidden layer and the output layer, and b_k is the threshold of the k neuron in the output layer. f(x) is the hyperbolic tangent function, and g(x) is the linear function.

• Determine the number of hidden layer neurons

The number of hidden layer neurons is usually determined by the following two formulas: (7) $$\begin{array}{c}n=2m+1\end{array}$$(7) (8) $$\begin{array}{c}n=\sqrt{m+l}+a\end{array}$$(8) (9) $$\begin{array}{c}\mathit{\text{NRMSE}}=\frac{\frac{1}{b}\sum _{i=1}^b{\left(c_i-{\widehat{c}}_i\right)}^2}{c_{\mathrm{\text{max}}}}\end{array}$$(9)

In the above formula, a is an arbitrary natural number between 1 and 10, b is the number of samples in the training set, c_i is i actual value, ${\widehat{c}}_i$ is i predicted value, and c_max is the maximum value in the actual value. By constantly trying to adjust the value of a, the value range of the number of hidden layer neurons n is determined to be [5,29], and the normalized root mean square error (NRMSE) of the training set is calculated to evaluate the performance of the model. The number of hidden layer neurons with the smallest NRMSE is selected as the optimal value.

• 3. Sample data processing

    There are landslides in all 32 slope units in the study area; therefore, to create a sample dataset, the study examined 32 slope units with landslide points (positive samples) and 32 slope units without landslides (negative samples). The information value of each sample was then normalized. To create random sequences, the sample dataset is sorted randomly. The ratio of 70% to 30% is used to randomly select the training set and testing set. The last 19 sets of data are designated as testing data, whereas the first 45 sets are designated as training data.

(10) $$\begin{array}{c}{X}_i=\frac{x_i-x_{\mathrm{\text{min}}}}{x_{\mathrm{\text{max}}}-x_{\mathrm{\text{min}}}}\end{array}$$(10)

In the formula: x_max and x_min are the maximum and minimum values of the amount of information of each group of samples, X_i is the amount of information of the i samples, and X_i is the amount of information of the i samples after normalization.

When the model achieves satisfactory results with minimal error, the corresponding year’s dynamic landslide hazard is predicted, thus obtaining the dynamic hazard zone. The steps of using the information value and BPNN for landslide dynamic hazard assessment are illustrated in Figure 4.

Figure 4. Flow chart of BPNN based on landslide dynamic hazard assessment.

4. Results

4.1. Division of assessment units and factor classifications

According to the scope of the research area was separated into 200 slope units using a 12.5 m DEM and the extent of the research area, which is founded on the hydrological analysis method and integrated with the field research. 43 landslides in this area are distributed in 32 slope units (Figure 5). The morphological rationality of slope units in the study area was tested by the morphological index method (Li et al. 2023): (11) $$\begin{array}{c}F=\frac{L^2}{4\pi S}\end{array}$$(11)

Figure 5. Division of slope units in research area.

In the formula: L represents the perimeter of the slope unit (m); s represents the slope unit area (m²). When the morphological index is 1-3, it shows that the divided slope unit fits the controlled geomorphological boundaries such as the ridge line and valley line in the study area, which is consistent with the actual topography. The distribution of slope unit shape index in the study area is shown in Table 2, of which 98% of the slope unit shape index is between 1-3, which indicates that the overall shape of the slope unit in the study area is relatively regular, which can meet the requirements of the next landslide susceptibility assessment and hazard assessment.

Table 2. Statistical table of slope unit morphological index distribution.

Morphological index	Number of slope units	Percentage of total slope units (%)
1 -3	196	98
3 -5	2	1
>5	2	1

The primary data for assessment and prediction of landslide susceptibility and hazard, the influencing factor parameters extracted from 200 slope units were taken as the original data. To identify the 13 influencing factors in the static landslide susceptibility assessment index system and the introduced dynamic precipitation influencing factors, the natural break method was applied. This classification is intended for subsequent assessment and prediction of landslide susceptibility and hazard. Figure 6 illustrates the distribution of topography, engineering geology, hydrology, and human engineering activities factors within the slope units, which are part of the static influencing factor system for landslides. Figure 7 displays the average annual precipitation from the 1980s to the 2010s. Figure 8 represents the average annual precipitation for the years 2025 and 2030, which were predicted using a time series approach.

Figure 6. The landslide static influencing factor system: (a) elevation difference, (b) total curvature, (c) slope degree, (d) slope aspect, (e) slope structure, (f) depth of overburden, (g) distance to fault, (h) distance to road, (i) distance to river, (j) NDVI, (k) TWI, (l) SPI,(m) lithology, (n) legend of lithology.

Figure 7. Average annual precipitation in the 1980s-2010s: (a) PRE-1980s; (b) PRE-1990s; (c) PRE-2000s; (d) PRE-2010s.

Figure 8. Average annual precipitation forecast: (a) PRE-2025; (b) PRE-2030.

4.2. Implementation of the model

To produce two models assessing landslide susceptibility and hazard, the information value model is based on the static influencing factor system for landslides. The information value of 13 influencing factors was calculated separately, and the susceptibility zoning map was generated accordingly. To establish and train the BPNN model, we utilized the information value of the landslide dynamic influencing factor system. This model is applied to the entire study area by ArcGIS, producing dynamic hazard zoning maps for historical periods (1980s-2010s) as well as for future predictions (2025 and 2030).

4.2.1. Landslide susceptibility assessment

The research area includes 43 landslides and 200 slope units, with 32 slope units exhibiting landslide occurrences. The information value of the 13 static influencing factors is calculated using the information value model, as shown in Table 3. In some categories, there are slope units that have not experienced landslides. To ensure proper computation, the number of landslide-prone slope units in these categories is assigned a value of 0.01.

Table 3. The information on static influencing factor.

Factor	Class	Number of landslides unit by classification	Number of slope units by classification	Number of slope units with landslides/Total slope units with landslides	Number of slope units/ Total slope units	Information value
Elevation Difference	0-179m	19	90	0.59	0.45	0.277
	179-324m	6	41	0.19	0.21	−0.089
	324-475m	5	30	0.16	0.15	0.041
	475-661m	2	23	0.06	0.12	−0.610
	661-1161m	0.01	16	0.0003	0.08	−5.545
Slope Degree	0-16.73°	9	24	0.28	0.12	0.852
	16.73-26.71°	17	39	0.53	0.20	1.002
	26.71-37.52°	6	45	0.19	0.23	−0.182
	37.52-50.34°	0.01	69	0.0003	0.35	−7.007
	50.34-79.87°	0.01	23	0.0003	0.12	−5.908
Slope Aspect	North	1	23	0.03	0.115	−1.303
	North-East	0.01	40	0.0003	0.20	−6.461
	East	3	17	0.09	0.085	0.098
	South-East	7	26	0.22	0.13	0.520
	South	13	28	0.41	0.14	1.065
	South-West	3	22	0.09	0.11	−0.160
	West	3	18	0.09	0.09	0.041
	North-West	2	26	0.06	0.13	−0.732
Curvature	concave	2	6	0.06	0.03	0.734
	convex	30	194	0.94	0.97	−0.034
Slope Structure	Reverse slope	3	46	0.09	0.23	−0.897
	Transverse slope	6	59	0.19	0.30	−0.453
	Skew slope	11	57	0.34	0.29	0.187
	Dip slope	12	38	0.38	0.19	0.680
Distance to fault	0-500m	1	16	0.03	0.08	−0.940
	500-1000m	3	21	0.09	0.11	−0.113
	1000-1500m	3	12	0.09	0.06	0.446
	1500-2000m	4	26	0.13	0.13	−0.039
	>2000m	21	125	0.66	0.63	0.049
Lithology	A	8	18	0.25	0.09	1.022
	B	2	9	0.06	0.05	0.329
	C	0.01	15	0.0003	0.08	−5.481
	D	9	73	0.28	0.37	−0.261
	E	0.01	1	0.0003	0.01	−2.773
	F	13	84	0.41	0.42	−0.033
TWI	0.83-4.04	1	54	0.03	0.27	−2.156
	4.04-5.27	23	128	0.72	0.64	0.116
	5.27-6.7	8	18	0.25	0.09	1.022
SPI	−4.38-2.06	0.01	1	0.0003	0.01	−2.773
	2.06-3.43	18	36	0.56	0.18	1.139
	3.43-4.7	14	138	0.44	0.69	−0.456
	4.7-6.41	0.01	25	0.0003	0.13	−5.991
NDVI	−0.002-0.015	8	18	0.25	0.09	1.022
	0.015-0.022	9	48	0.28	0.24	0.159
	0.022-0.028	7	66	0.22	0.33	−0.411
	0.028-0.034	4	37	0.13	0.19	−0.392
	0.034-0.05	4	31	0.13	0.16	−0.215
Distance to river	0-200m	4	37	0.13	0.19	−0.392
	200-400m	6	64	0.19	0.32	−0.535
	400-800m	12	78	0.38	0.39	−0.039
	800-1000m	5	10	0.16	0.05	1.139
	>1000m	5	11	0.16	0.06	1.044
Distance to road	0-50m	2	4	0.06	0.02	1.139
	50-100m	17	39	0.53	0.20	1.002
	100-150m	8	31	0.25	0.16	0.478
	150-200m	2	13	0.06	0.07	−0.039
	>200m	3	113	0.09	0.57	−1.796
Depth of overburden	5.1-6.91m	7	25	0.22	0.13	0.560
	6.91-8.04m	8	68	0.25	0.34	−0.307
	8.04-9.46m	5	26	0.16	0.13	0.184
	9.46-10.78m	7	62	0.22	0.31	−0.349
	10.78-15.74m	5	19	0.16	0.10	0.498

Figure 9 illustrates the classification intervals and information quantities for each influencing factor. Taking into account the characteristics and distribution of landslide disasters in the Qingjiang Reservoir, the elevation difference, slope structure, and lithology are used as examples for explanation.

Figure 9. The classification information value of each influencing factor and the scale ratio of slope units with land-slides: (a) elevation difference, (b) slope degree, (c) slope aspect, (d) total curvature, (e) slope structure, (f) distance to fault, (g) lithology, (h) TWI, (i) SPI, (j) NDVI, (k) distance to river, (l) distance to road, (m) depth of overburden.

1. Spatial characteristics of landslide distribution are largely determined by the topography. As shown in Figure 9(a), landslides mainly develop in the elevation difference classification of 0-179m, accounting for 59% of all slope units with landslides. This classification provides an information value of 0.277, indicating that landslides are primarily distributed in low-lying areas along the Qingjiang River.

2. Slope structure is an important indicator reflecting the slope topography and overall stability control. As depicted in Figure 9(e), slope units with landslides are predominantly distributed in skew slope and dip slope. Among them, dip slopes have the highest occurrence, accounting for 38% of slope units in this classification, providing an information value of 0.68.

3. Lithology serves as the material basis for geological hazards. The higher the mechanical strength and integrity of the lithology, the lower the potential for geological disaster occurrence, and vice versa. As indicated in Figure 9(g), landslides are mainly spread in lithologies such as limestone, dolomitic limestone interbedded with shale and siltstone rock formation, which collectively account for 93% of all slope units with landslides. This indicates that landslide disasters occur in both clastic and carbonate rocks, typically in weaker lithologies.

The comprehensive information value of the research area was analyzed. ArcGIS software was utilized to overlay and analyze the factors, and the results were assigned to slope units for landslide susceptibility assessment in Qingjiang Reservoir. The natural break method was used to classify the results of susceptibility assessment results in four grades: low, moderate, high, and very high. Figure 10 shows the overall characteristics of landslide susceptibility at different levels. The comparison between the susceptibility zones and actual disaster distributions in Table 4 indicates a high correlation. No landslides were detected in slope units classified as low or moderate-low susceptibility zones. As the susceptibility level increases, the number of actual landslide disasters increases, with 52% of slope units belonging to very high and high susceptibility zones. While the western and central areas in the study region have fewer documented landslides, the model evaluated them as very high susceptibility areas, indicating a high potential for landslide development in these regions.

Figure 10. Result of landslide susceptibility on the basis of information value model.

Table 4. Landslide susceptibility zoning result.

Susceptibility zoning	Number of slope units	Percentage of total slope units (%)	Number of landslides	Percentage of total landslides (%)
Low	24	12	0	0
Moderate	72	36	0	0
High	50	25	5	11.63
Very High	54	27	38	88.37

4.2.2. Landslide dynamic hazard prediction

According to different years, six dynamic hazard assessment models of the landslide were established respectively. The input layer is the amount of information value of each slope unit in the assessment system of dynamic influencing factors. The output layer is output ‘1’ when there is a landslide disaster, and output ‘0’ when there is no landslide disaster. Figure 11 is the structure of the BPNN model used in this paper.

Figure 11. Structure of the BPNN model for landslide dynamic hazard prediction.

Figure 12. Landslides hazard for different ages: (a) 1980s; (b) 1990s; (c) 2000s; (d) 2010s.

1. Dynamic hazard prediction of historical landslides

 The information value of precipitation factors for each decade was calculated and four BPNN models were established accordingly (Table 5). The sample dataset consisted of 64 slope units. The positive and negative samples, as well as the output layer results, varied depending on the occurred of landslides in each decade. Using the natural break method, the results of the hazard assessment were grouped in four levels: low, moderate, high, and very high (Figure 12). The distribution of landslide hazard areas varied significantly across different years, reflecting the influence of annual precipitation changes.

Table 5. The information of precipitation.

Factor	Class	Number of slope units with landslides by classification	Number of slope units with landslides	Number of slope units by classification	Information value
	1843-1863mm	2	13	45	−0.380
	1864-1880mm	5	13	62	0.216
PRE-1980s	1881-1894mm	0.01	13	37	−5.483
	1895-1912mm	5	13	39	0.679
	1913-1938mm	1	13	17	−0.099
	1011-1046mm	2	12	34	−0.019
	1047-1078mm	5	12	57	0.380
PRE-1990s	1079-1106mm	4	12	46	0.371
	1107-1134mm	1	12	38	−0.824
	1135-1178mm	0.01	12	25	−5.011
	1000-1043mm	1	8	33	−0.278
	1044-1070mm	4	8	41	0.892
PRE-2000s	1071-1093mm	2	8	43	0.151
	1094-1117mm	1	8	46	−0.610
	1118-1157mm	0.01	8	37	−4.997
	1000-1039mm	3	8	33	0.821
	1040-1065mm	1	8	43	−0.542
PRE-2010s	1066-1088mm	3	8	55	0.310
	1089-1115mm	1	8	49	−0.673
	1116-1161mm	0.01	8	20	−4.382
	1388-1399mm	0.01	32	1	−2.773
	1399-1403mm	6	32	46	−0.204
PRE-2025	1403-1406mm	11	32	83	−0.188
	1406-1409mm	10	32	54	0.146
	1409-1418mm	5	32	16	0.669
	1399-1402mm	3	32	36	−0.652
	1402-1404mm	3	32	42	−0.806
PRE-2030	1404-1406mm	6	32	36	0.041
	1406-1408mm	11	32	54	0.241
	1408-1412mm	9	32	32	0.564

Figure 13 shows the relation observed between number of slope units and average precipitation under different levels of hazards. In the 1980s, the majority of slope units were assigned to very high and high hazards, while in the 2010s, the largest number of slope units were assigned to low and moderate hazard classes. In the 1980s, with 14 landslides and an average precipitation of 1882.85 mm. The lowest number of slope units fell into the low and moderate hazard classes, while the highest number belonged to the very high and high hazard classes, with a total of 132 units. During the 1990s, there were 12 landslides with an average precipitation of 1086.905 mm, of which 106 slope units belonged to very high and high hazard classes. During the 2000s, 8 landslides occurred with an average precipitation of 1084.155 mm and 88 slope units were in the high and very high hazards. In the dataset for the 2010s, there were 9 landslides, with 78 slope units classified as low and moderate hazards, and 122 slope units classified as very high and high hazards. As average annual precipitation decreases, number of slope units that are segmented as very high and high hazard levels also decreases.

Figure 13. Slope unit numbers for different hazard areas in each model and the average annual precipitation in each decade.

• 2. Dynamic hazard prediction of landslide in the future

   Based on the predicted precipitation data layer, the information value of precipitation factors for 2025 and 2030 was calculated (Table 5). Due to the inability to determine the number of landslides in future years, all 32 slope units with historical landslides have been treated as positive samples, and all 32 slope units without historical landslides have been treated as negative samples during hazard prediction. Two BPNN models were established, and the natural break method was used to classify hazard levels in four classes: low, moderate, high, and very high (Figure 14 and Table 6). Predictions indicate that in 2025 and 2030, the proportion of high and very high areas will be 50.5% and 57.5%, respectively. The studied area’s share of very high and high hazard zones for landslides exhibits a shifting tendency with variations in precipitation, suggesting that future landslide occurrences in Qingjiang Reservoir may be more likely. As precipitation increases, there will also be a progressive increase in the occurrence probability of landslides.

Figure 14. Landslide hazard prediction for different years: (a)2025; (b)2030.

Table 6. Landslide hazard zoning result in 2025 and 2030.

Hazard zoning	Number of slope units in 2025	Percentage of total slope units in 2025 (%)	Number of slope units in 2030	Percentage of total slope units in 2030 (%)
Low	52	26	32	16
Moderate	47	23.5	53	26.5
High	77	38.5	65	32.5
Very High	24	12	50	25

4.3 Validation of the model

4.3.1. ROC validation

To further validate the reliability of landslide susceptibility and hazard assessment results, the model was evaluated using ROC. Area under curve (AUC) is an objective quantitative measure used to assess the modeling accuracy (Arabameri et al. 2017). The diagonal line of ROC, also known as baseline, has an AUC of 0.5 to 1, the more its value converges to 1, the better the model’s predictive accuracy. The 86 non-landslide disaster points selected randomly in this research to match the number of existing landslide disaster points were used as testing datasets for ROC analysis (Figure 15). Landslide susceptibility assessment with an AUC value of 0.783, while the AUC values for historical dynamic hazard assessment were 0.713, 0.835, 0.783, and 0.905, with the BPNN model for the 2010s showing the best predictive accuracy. The model results demonstrated that the landslide hazard assessment models for different periods fell into the satisfactory classification and showed good consistency with the actual occurrence of disasters. In the paper, these accuracy indicators show that the BPNN model applied is reliable and could provide dynamic prediction of landslide hazard in the next 5-10 years.

Figure 15. ROC curve for models.

4.3.2. Field investigation and validation

Landslide susceptibility zoning results area are displayed in Figure 10: very high susceptibility areas are more densely distributed on the north bank of the Qingjiang River, while they are more sparsely distributed on the south bank. This distribution is closely related to factors such as geological environmental conditions, slope structure, and human engineering activities. Through field investigations, as shown in Figure 16, it was found that the lithological characteristics of the very high susceptibility area are primarily composed of nodular limestone, limestone interbedded with weak thin-layered carbonaceous mudstone, dolomitic limestone interbedded with shale, and siltstone, which are weak rock formations. These formations are mainly distributed in the downstream structural slope areas on both sides of the Qingjiang River. Moreover, slope cutting for housing construction, road construction and other activities in the area of study have also created a significant number of slopes with exposed faces.

Figure 16. The fieldwork process. (a): very high susceptible areas; (b), (d), (f) and (g): some typical weak rock mass; (c): slope-cutting behind the house; (e): slope-cutting to build the road; and (h): cracks behind the house.

Furthermore, Figure 17(a) shows some historically identified regions with very high landslide hazards predicted by the BPNN model. Figure 17(b)–(i) displays eight typical water-related landslides of the research area, and under the combination of precipitation and frontal erosion by the Qingjiang River, the high-hazard zones will likely remain dangerous in the future.

Figure 17. Combined landslides dynamic hazard prediction and field survey map, (a): predicted very high hazard areas, (b)-(i): pictures of typical landslides.

5. Discussion

Assessment of landslide hazard is an effective method to address the threat caused by geological hazards. This method can predict the areas where landslide disasters are likely to occur based on historical data, topography, and triggering factors. It enables the rapid and accurate determination of the approximate location of potential landslides, providing essential information for the prevention and management of potential landslides. Various methods have been used to model the susceptibility and hazard assessment of landslides, among which the information value model and the BPNN model have been widely applied in the field of geological disaster assessment. In this research, the information value model was used for static landslide susceptibility assessment, and the MATLAB software was used to write the code to construct the BPNN model and combined with the information value model for predicting the occurrence of landslides in the precipitation susceptible areas of the Qingjiang Reservoir. Finally, the reliability of susceptibility mapping and historical hazard mapping was validated using the ROC and field survey validation.

This work aims to create some optimizations in comparison with earlier research to increase the rigor of the research process and the reliability of the research findings.

1. Mapping units

   Slope units created from high-resolution spatial data were used as mapping units in the study area. Topographic and geomorphologic features are significant in the formation and occurrence of landslides (Liu et al. 2022; Sun et al. 2020). Traditional mapping units, such as grid cells, are the most commonly used landslide disaster units due to their characteristics like constant shape, ease of computation and sampling. Nevertheless, the method is independent of topographic and geomorphologic features and may contain multiple different geological structures. The slope unit is better able to reflect the ground morphology characteristics in the terrain spatial unit division compared with the grid unit, because the slope unit, whose principle is based on the ridge line and valley line to divide, is the unit of geohazard occurrence. Compared with other mapping units, it can better represent the topographic and geomorphic features in a small area (Ba et al. 2018). The study area is a middle-low mountainous landform, and the mainstream of Qingjiang River passes through its territory, forming a unique gully geomorphologic area, with steep mountains on both sides of the river valley and slopes that are mostly steep-slope terrain. GIS-based slope units are able to do more accurate mapping of landslide hazards utilizing landslide survey data and high-resolution spatial data (Zhao et al. 2021). Therefore, the slope units used in this study reflect the relationship between the assessment units and the unique valley topography and geomorphology of the study area.

2. Modeling methods

   Modeling methods for landslide hazard assessment can be summarized as heuristic, mathematical statistical and machine learning models The traditional heuristic model assigns the weight of landslide influencing factors subjectively, resulting in relatively low accuracy in landslide hazard prediction, while the accuracy of mathematical statistical model in landslide hazard prediction is higher than that of the heuristic model (Aditian et al. 2018; Chen et al. 2018). The information value model is a commonly used mathematical statistical model. Based on the information entropy theory, it can calculate the weight of various geological disaster factors and reflect their contribution to the occurrence of landslide disasters (Figure 9). Farooq and Akram (Farooq and Akram 2021) used the information value to predict the susceptibility of landslides in mountainous areas, and the established model has high accuracy. Niu established a new method for landslide prediction in soil-rock contact zones based on information value model (Niu et al. 2024), which can better solve the relationship between landslide risk and various factors. The machine learning model can better solve the complex relationship between landslide hazard and various influencing factors. BPNN has the ability to learn and identify the complex nonlinear relationship between data and can learn the rules and characteristics of landslide evolution from a large amount of data. Xu (Xu et al. 2015) applied the BPNN model to the prediction of landslide susceptibility in the Three Gorges Reservoir of China, and the prediction accuracy reached 88%. Huang et al. (2020) applied heuristic, mathematical statistics and machine learning models to landslide susceptibility prediction respectively. The results show that the machine learning model represented by BPNN model is superior to the traditional heuristic model and mathematical statistics model in prediction accuracy and has better prediction performance.

   In this paper, the combination of information value model and BPNN model is applied to the prediction of landslide dynamic hazard. The verification results of the model show that the combination of these two methods has high prediction accuracy in landslide hazard prediction. The combination of information value model and BPNN model can make more comprehensive use of data information, and improve the accuracy and reliability of landslide hazard prediction (Chen et al. 2020; L. Huang et al. 2021).

3. Dynamic precipitation

   In previous studies, the triggering factors for landslide occurrences were often based on precipitation return periods (Jiménez-Perálvarez et al. 2017; Li et al. 2019; Tsunetaka 2021), and landslide hazard zonation was obtained by predicting the landslide instability probability under different return periods of extreme precipitation, which cannot provide hazard zonation for non-extreme precipitation conditions in the future. Due to the variable climatic conditions, this type of research is unable to obtain the hazard zoning under normal precipitation in the future. In contrast with previous research, this study established the dynamic influencing factors system by combining average annual precipitation and static influencing factors. Based on historical average annual precipitation predicted future actual average annual precipitation and conducted dynamic hazard prediction for landslides in the next 5-10 years, a reliable landslide hazard zoning map is made. The research results indicate that landslide disasters in the Qingjiang Reservoir are influenced by changes in precipitation intensity, and the spatial distribution of hazards for each year shows obvious differences, reflecting the dynamic changing process year by year. The paper’s research approach can generate the trend of spatial location change of landslide hazard in the future and assist governments in formulating long-term risk management strategies.

4. Sample dataset

   In the process of landslide hazard prediction, different numbers of landslide sample datasets were used according to different years. There are 43 landslides spread over 32 slope units in this study area. As shown in Figure 5, a slope unit may incorporate a variety of landslides because these landslides have the same formation mechanism and occur in the same slope unit. In the dynamic hazard prediction, a fixed number of positive and negative samples were selected by choosing 32 slope units with landslides and 32 slope units without landslides. Different numbers of positive and negative samples were used depending on the number of landslides that occurred each year. Therefore, each time the BPNN was used for prediction, the output data in the training and testing states were different. The advantage of doing this is that compared to the use of fixed landslide catalogs in previous studies, the hazard zonation maps in this study clearly show the changes in landslide hazard with precipitation, reflecting the dynamic changing process of landslide disasters for each year. We employed field investigations, remote sensing images, and geological surveys to guarantee the accuracy and consistency of the 43 sets of data samples related to landslides. The performance of the BPNN model in this study indicates that it can learn from limited data and offer good predictions, even though the number of samples of 43 landslides in a small area may be regarded as less than certain studies in a big area (Achour et al. 2021; Qasimi et al. 2023). The spatial distribution of landslides is faithfully represented by the model. The efficacy of the BPNN model in predicting landslide hazard, even in a small area with few landslide occurrences, is demonstrated by the results of ROC validation and field investigation cross-validation in our study.

The research method of this paper is devoted to predicting the hazard of landslides through precipitation, to achieve environmental safety and sustainable regional development. The suitable models in various study regions differ because of variances in landslide number, geographic location, and climate conditions. We are also aware of the research methods’ limitations, such as the model’s applicability, the kind of landslide, and the duration of the precipitation data. For the purpose of further enhancing the accuracy of landslide dynamic hazard assessment and prediction, future research will be conducted in the following directions: gather more precipitation data over various time spans in areas with the same geological background conditions; explore novel approaches and techniques; and integrate various mathematical statistical models and machine learning models. The framework for risk management and early warning for landslide disasters incorporates the landslide dynamic hazard assessment model.

6. Conclusion

In this study, we constructed a reliable landslide dynamic hazard prediction model in a landslide-prone area by combining information model with BPNN model. The theoretical significance of this study is to calculate information value of the relationship between influencing factors and the occurrence of landslides, and to establish a dynamic hazard assessment index system based on the average annual precipitation. Then, develop a landslide dynamic hazard prediction model based on MATLAB and BPNN. The research method provides new ideas and methods for the establishment of geological disaster hazard prediction based on mathematical statistical models and machine learning models, and further enriches the theoretical system of geological disaster hazard assessment.

The practical significance is that the dynamic hazard prediction of mountain landslide disasters is made, and the landslide hazard zoning results in the next 5-10 years are obtained. The results show that the accurate prediction of the very high hazard area can be realized according to the precipitation factors that trigger the occurrence of mountain landslides, which provides a focused and targeted scientific prevention and control area for disaster prevention and mitigation. On the one hand, it can effectively reduce the potential threat of landslides to people’s personal safety and property safety. On the other hand, it can also provide a reference for the dynamic hazard prediction of landslides in mountainous areas, so as to determine the hazard range of regional landslides and provide scientific basis for understanding global landslide disasters.

Acknowledgements

We the authors thank the editors and anonymous reviewers for their comments and suggestions, which are of great help to the writing of this article.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The 12.5 m spatial resolution DEM used in this study was downloaded from https://search.asf.alaska.edu/#/; Rainfall precipitation is downloaded from https://www.resdc.cn/DOI/.