Big data-based geological disaster risk evaluation method and system

Abstract

The invention discloses a geological disaster risk evaluation method and system based on big data, and the method comprises the following steps: S1, obtaining multi-source geological disaster data, including geological structure data, landform data, meteorological and hydrological data, human activity data and historical disaster data; s2, performing space-time alignment and standardization processing on the multi-source geological disaster data, and constructing a geological disaster space-time data set; s3, constructing a geological disaster risk evaluation model based on the deep learning network, wherein the geological disaster risk evaluation model comprises a feature extraction module, a space-time correlation module and a risk prediction module; s4, calculating a weight coefficient of each evaluation index by adopting a dynamic weight distribution method; s5, calculating a geological disaster risk index through the geological disaster risk evaluation model, and generating a risk grade zoning map; according to the invention, more intelligent geological disaster risk evaluation of the urban area is realized.

Description

Technical Field

[0001] This invention relates to the field of data analysis, specifically to a method and system for geological disaster risk assessment based on big data. Background Technology

[0002] Geological disasters are characterized by their suddenness and destructive power, which can not only damage the ecological environment but also cause casualties and significant economic losses. The geological disaster risks in special urban areas are even more unique and dangerous. These areas have complex topography, high density of human activities, and diverse and dynamic risk factors. Once a geological disaster occurs, it can easily trigger a chain reaction, posing a serious threat to the safety of urban operations and the lives and property of residents. Therefore, the accuracy and timeliness of geological disaster risk assessment in such areas are required to be higher.

[0003] While current mainstream geological hazard risk assessment methods have attempted to incorporate multi-source data, significant shortcomings remain: insufficient multi-source data processing capabilities, frequent spatiotemporal scale misalignment and large data volume differences among different data types, and a lack of systematic spatiotemporal alignment and standardization, resulting in inconsistent data quality and difficulty in forming reliable assessment datasets, particularly inadequate for adapting to the fragmented characteristics of small-scale, multi-data sets in specific urban areas; limited applicability of assessment models, with traditional models often employing single-scale feature extraction and static weight allocation, failing to capture the temporal series changes of meteorological and hydrological data and the spatiotemporal distribution patterns of historical disasters, leading to overfitting or prediction bias, and failing to accurately reflect the coupling effects of risk factors in specific urban areas; and insufficient timeliness of risk assessment and early warning, with most methods employing static assessment models, unable to incorporate new monitoring data to update risk results in real time, and lacking effective dynamic optimization mechanisms for model parameters, resulting in delayed early warnings and difficulty in addressing the suddenness and dynamism of risks in specific urban areas. Therefore, a geological hazard risk assessment method based on big data is proposed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a geological hazard risk assessment method based on big data, comprising the following steps:

[0005] S1. Acquire multi-source geological disaster data, including geological structure data, topographic data, meteorological and hydrological data, human activity data, and historical disaster data;

[0006] S2. Perform spatiotemporal alignment and standardization on the multi-source geological disaster data to construct a spatiotemporal dataset of geological disasters;

[0007] S3. Construct a geological hazard risk assessment model based on a deep learning network. The geological hazard risk assessment model includes a feature extraction module, a spatiotemporal correlation module, and a risk prediction module.

[0008] S4. Calculate the weight coefficients of each evaluation indicator using a dynamic weight allocation method;

[0009] S5. Calculate the geological hazard risk index using the geological hazard risk assessment model and generate a risk level zoning map;

[0010] S6. Dynamically update and issue early warnings based on real-time monitoring data for risk assessment results.

[0011] Furthermore, in step S2, the spatiotemporal alignment and normalization process includes:

[0012] Geological structural data are coded for fault density and lithological type.

[0013] Slope, aspect, and elevation features are extracted from topographic data;

[0014] Calculate rainfall intensity and hydrological parameters from meteorological and hydrological data;

[0015] Classify human activity data by building density and land use type;

[0016] Statistical analysis of disaster types and frequency of occurrence based on historical disaster data;

[0017] The multi-source geological disaster data with different resolutions are unified to the same spatiotemporal scale through spatiotemporal interpolation methods;

[0018] A standardized processing method based on geographic grids is adopted to divide the study area into regular grid units, each of which contains complete multi-source geological hazard data features;

[0019] The geographic grid-based standardization process is achieved by calculating the difference between the original data value and the mean of the data features, and then dividing by the standard deviation of the data features. Specifically, it is implemented through the following process:

[0020] ;

[0021] Where X is the original data value. Let σ be the mean of the data features, and X be the standard deviation of the data features. norm This is the standardized value.

[0022] Furthermore, in step S3, the process of constructing the geological hazard risk assessment model includes:

[0023] The feature extraction module uses a convolutional neural network to extract fault distribution features from geological structural data, topographic relief features from topographic data, rainfall distribution features from meteorological and hydrological data, and engineering construction features from human activity data.

[0024] The spatiotemporal correlation module uses a gated loop unit to capture the time-series variation characteristics of meteorological and hydrological data and the spatiotemporal distribution patterns of historical disaster data;

[0025] The risk prediction module integrates the extracted fault distribution features, topographic relief features, rainfall distribution features, and engineering construction features through a fully connected layer, and maps the integrated features into a risk index through linear transformation and activation function.

[0026] Furthermore, the geological hazard risk assessment model is optimized by combining prediction error and model complexity. First, the mean square error between the predicted and actual values ​​is calculated. Then, a regularization term for the model parameters is added to prevent overfitting. Specifically, optimization is performed using the following loss function:

[0027] ;

[0028] Where N is the number of samples, y i This represents the true risk value. To predict the risk value, M represents the number of parameters in the geological hazard risk assessment model. These are the parameters for the geological hazard risk assessment model. This is the regularization coefficient.

[0029] Furthermore, the feature extraction module employs a multi-scale convolutional neural network structure, including:

[0030] The first convolutional layer uses small-scale convolutional kernels to extract local terrain detail features from the terrain data;

[0031] The second convolutional layer uses large-scale convolutional kernels to extract regional geological structural features from geological structural data;

[0032] By fusing local terrain details extracted from the first convolutional layer and regional geological structure features extracted from the second convolutional layer through a feature pyramid network, a comprehensive feature representation for geological hazard risk assessment is formed.

[0033] Furthermore, in step S4, the dynamic weight allocation method includes:

[0034] The weights of geological stability indicators are calculated based on geological structural data;

[0035] The weights of rainfall impact indicators are calculated based on meteorological and hydrological data;

[0036] The weights of the impact indicators of engineering activities are calculated based on human activity data;

[0037] The weights of disaster recurrence probability indicators are calculated based on historical disaster data;

[0038] An importance matrix of evaluation indicators is constructed, and the initial weights of each evaluation indicator are determined by a combination of expert scoring and entropy methods.

[0039] The weights are dynamically adjusted by introducing a time decay factor and a spatial variation factor;

[0040] The weight calculation is achieved by multiplying the initial weight by the time decay factor and the spatial variation factor, specifically through the following process:

[0041] ;

[0042] Among them, W i (t) represents the dynamic weight of the i-th evaluation index at time t, W i0 Let α be the initial weight of the i-th evaluation index, α be the time decay factor, and β be the spatial variation factor of the i-th evaluation index.

[0043] Furthermore, the method for calculating the spatial variation factor includes:

[0044] Calculate the spatial variation characteristics of fault density based on geological structural data;

[0045] Calculate the spatial distribution characteristics of slope variation based on topographic data;

[0046] Calculate the spatial variation characteristics of rainfall based on meteorological and hydrological data;

[0047] Calculate the spatial distribution characteristics of building density based on human activity data;

[0048] The spatial variation factor of each evaluation index is calculated based on spatial heterogeneity. This is achieved by adding 1 to the ratio of the standard deviation to the mean of each evaluation index. The formula for calculating the spatial variation factor is as follows:

[0049] ;

[0050] Where σ is the standard deviation of the i-th evaluation index within the study area. Let be the mean of the i-th evaluation index within the study area.

[0051] Furthermore, in step S5, the calculation process of the geological hazard risk index includes:

[0052] Features are extracted from multi-source geological hazard data and a preliminary risk score is obtained through a geological hazard risk assessment model;

[0053] A weighted comprehensive evaluation is performed by combining dynamic weights calculated based on different data types;

[0054] The scores are mapped to the [0,1] interval using the sigmoid function to form a standardized risk index. This is achieved by calculating the weighted sum of the scores of each evaluation indicator and their dynamic weights, and then transforming the result using the sigmoid function. The formula for calculating the standardized risk index is as follows:

[0055] ;

[0056] Where R is the geological disaster risk index, W i S is the dynamic weight of the i-th evaluation index. i Let b be the score for the i-th evaluation indicator, b be the bias term, and n be the number of evaluation indicators.

[0057] Furthermore, in step S6, the dynamic update and early warning process includes:

[0058] Real-time acquisition of new meteorological and hydrological data to update rainfall impact assessments;

[0059] Real-time acquisition of new human activity data to update the impact assessment of engineering activities;

[0060] The geological hazard risk index was recalculated based on updated multi-source geological hazard data;

[0061] Set a risk threshold, and automatically trigger an early warning when the geological disaster risk index exceeds the risk threshold;

[0062] The parameters of the geological hazard risk assessment model are updated based on a sliding time window mechanism.

[0063] A geological hazard risk assessment system based on big data includes:

[0064] The data acquisition module is used to acquire multi-source geological disaster data, including geological structure data, topographic data, meteorological and hydrological data, human activity data, and historical disaster data.

[0065] The data processing module is used to perform spatiotemporal alignment and standardization processing on the multi-source geological disaster data to construct a spatiotemporal dataset of geological disasters;

[0066] The model building module is used to build a geological hazard risk assessment model based on a deep learning network. The geological hazard risk assessment model includes a feature extraction module, a spatiotemporal correlation module, and a risk prediction module.

[0067] The weight calculation module is used to calculate the weight coefficients of each evaluation indicator using a dynamic weight allocation method.

[0068] The risk calculation module is used to calculate the geological hazard risk index through the geological hazard risk assessment model and generate a risk level zoning map.

[0069] Update the early warning module to dynamically update and issue early warnings based on real-time monitoring data of risk assessment results.

[0070] The beneficial effects of this invention are reflected in:

[0071] In data processing, five types of multi-source data—geological structure, topography, meteorology and hydrology, human activities, and historical disasters—are integrated to avoid the limitations of single data sources. Simultaneously, spatiotemporal interpolation is used to unify the spatiotemporal scale of the data. Geographic grid division and standardization processes eliminate differences in data magnitude and spatiotemporal misalignment, ensuring the completeness and comparability of data across grid units, laying the foundation for accurate evaluation. The evaluation model employs a multi-scale convolutional neural network and a feature pyramid network to fuse features, combined with gated recurrent units to capture the time-series changes in meteorological data and the spatiotemporal patterns of historical disasters. Further optimization using a loss function containing mean squared error and regularization terms ensures both accurate predictions and avoids overfitting, improving generalization. The system adapts to different scenarios; the weight allocation combines expert scoring and entropy methods to determine the initial scientific weights, and then dynamically adjusts them after processing with time decay factors and spatial variation factors, which can adapt to different time and space scenario changes and enhance the pertinence of the evaluation; the risk assessment uses the sigmoid function to map the weighted score to the [0,1] interval to generate a standardized risk index, and outputs a risk level zoning map to intuitively present the regional risk differences. At the same time, it acquires new meteorological, hydrological, and human activity data in real time, updates the model parameters and recalculates the risk index with a sliding time window mechanism, and automatically triggers early warnings by setting risk thresholds, thus buying time for disaster prevention and mitigation. Attached Figure Description

[0072] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0073] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0074] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0075] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0076] like Figure 1 As shown, a geological hazard risk assessment method based on big data includes the following steps:

[0077] S1. Acquire multi-source geological disaster data, including geological structure data, topographic data, meteorological and hydrological data, human activity data, and historical disaster data;

[0078] S2. Perform spatiotemporal alignment and standardization on the multi-source geological disaster data to construct a spatiotemporal dataset of geological disasters;

[0079] S3. Construct a geological hazard risk assessment model based on a deep learning network. The geological hazard risk assessment model includes a feature extraction module, a spatiotemporal correlation module, and a risk prediction module.

[0080] S4. Calculate the weight coefficients of each evaluation indicator using a dynamic weight allocation method;

[0081] S5. Calculate the geological hazard risk index using the geological hazard risk assessment model and generate a risk level zoning map;

[0082] S6. Dynamically update and issue early warnings based on real-time monitoring data for risk assessment results.

[0083] Furthermore, in step S2, the spatiotemporal alignment and normalization process includes:

[0084] Geological structural data are coded for fault density and lithological type.

[0085] Slope, aspect, and elevation features are extracted from topographic data;

[0086] Calculate rainfall intensity and hydrological parameters from meteorological and hydrological data;

[0087] Classify human activity data by building density and land use type;

[0088] Statistical analysis of disaster types and frequency of occurrence based on historical disaster data;

[0089] The multi-source geological disaster data with different resolutions are unified to the same spatiotemporal scale through spatiotemporal interpolation methods;

[0090] A standardized processing method based on geographic grids is adopted to divide the study area into regular grid units, each of which contains complete multi-source geological hazard data features;

[0091] The geographic grid-based standardization process is achieved by calculating the difference between the original data value and the mean of the data features, and then dividing by the standard deviation of the data features. Specifically, it is implemented through the following process:

[0092] ;

[0093] Where X is the original data value. Let σ be the mean of the data features, and X be the standard deviation of the data features. norm The value is the standardized value;

[0094] This system addresses the challenges of dense human activity, complex geological conditions (such as the presence of both natural terrain and man-made structures), diverse data sources, and significant resolution differences in special urban areas. It performs targeted feature extraction on five data categories: geological structures (e.g., the distance between faults and engineering facilities), topography (e.g., steep slopes around residential areas), meteorology and hydrology (e.g., localized heavy rainfall in cities), human activities (e.g., subway construction, building density), and historical disasters (e.g., past small landslides, ground subsidence). This allows for precise capture of core risk factors in these areas. Furthermore, it unifies the spatiotemporal scale of data at different resolutions through spatiotemporal interpolation. Combined with geographic grid-based partitioning and standardized formulas, it eliminates differences in data volume, ensuring that each small grid (e.g., 100m×100m, adaptable to the needs of refined urban management) within special urban areas contains complete and comparable multi-source data features. This effectively solves the problems of data fragmentation and inconsistent scales in special urban areas, providing a reliable data foundation for constructing spatiotemporal datasets tailored to specific urban scenarios and improving the accuracy of risk assessment. It also reduces assessment errors caused by poor data adaptability, better meeting the needs of geological disaster risk prevention and control in special urban areas for refined data.

[0095] Taking the geological hazard risk assessment of the area within 500m of Metro Line 3 (a special area type: dense underground engineering and terrain transition zone, involving residential areas with steep slopes and metro station construction areas) as an example, the specific handling process is as follows:

[0096] Targeted processing of various types of data (adapted to the specific characteristics of urban areas):

[0097] Geological structure data processing: Collect 1:10,000 engineering geological survey reports for the area, focusing on the fault density within 100m of the subway tunnel (as fault activity may affect tunnel safety), and code the lithology of the tunnel (e.g., moderately weathered sandstone is 1, strongly weathered mudstone is 2, mudstone is easily softened when exposed to water and requires special attention).

[0098] For example, within a construction area grid (100m × 100m) of a certain station on Metro Line 3, there are two faults with a total length of 150m within a 100m radius around the tunnel. The fault density is 150m / (0.1km × 0.1km) = 15000m / km. 2 The tunnel passes through rock formations mainly composed of strongly weathered mudstone, coded as 2.

[0099] Topographic data processing: Based on 0.5m resolution urban drone aerial DEM data, the slope, aspect, and elevation of steep slopes in the area were extracted (the slope of slopes around residential areas directly affects residents' safety). Within the grid of the construction area, the slope of the eastern slope adjacent to the residential area is 35° (slopes with a slope >30° in the city need to be monitored closely), the aspect is southeast (rain erosion is more concentrated during the rainy season in summer), and the average elevation of the grid is 235m (the elevation difference with the surrounding residential area is 12m, indicating potential for landslides).

[0100] Meteorological and hydrological data processing: Data from street-level meteorological stations deployed by the city meteorological bureau (500m resolution, adapted to local rainfall differences in the city) were collected to calculate the "short-term heavy rainfall intensity" of the area (urban flooding and slope instability are often related to short-term rainfall); simultaneously, combined with urban groundwater level monitoring data, the groundwater level depth in the construction area was calculated. The maximum 1-hour rainfall in the above grid over the past 5 years was 45mm (the standard for short-term heavy rainfall is ≥20mm per hour), and the rainfall intensity = 45mm / 1h = 45mm / h; the groundwater level depth in the construction area was 3.2m (excessive groundwater level can easily lead to lithological softening).

[0101] Human activity data processing: Based on the "building survey data" and "construction log" from the urban planning department, the building density (the consequences of disasters are more severe in densely built-up areas of the city) and subway construction intensity (such as monthly excavation volume) within the grid are statistically analyzed. Within the aforementioned construction area grid, the surrounding building area is 6200 m², and the building density is 6200 m². 2 / (100m × 100m) = 62%; The monthly excavation volume for subway construction is 8000m³ (high excavation intensity easily disturbs the geological body); The land use type is classified as "construction area + residential land", and the code is 4 (custom code: 1-cultivated land, 2-forest land, 3-pure residential, 4-construction + residential).

[0102] Historical disaster data processing: The city's emergency management bureau compiled a "small-scale urban disaster ledger" (distinct from large-scale regional disaster records), and statistically analyzed the types of past geological disasters (such as ground subsidence and small landslides) and their frequency of occurrence within the grid. Within the aforementioned construction area grid, there were two instances of small-scale ground subsidence (maximum subsidence of 5cm) caused by construction in the past 10 years, with no landslide records. The disaster frequency is 2 times / 10 years = 0.2 times / year.

[0103] Spatiotemporal interpolation with unified scale (addressing differences in urban data resolution):

[0104] The data resolution in this area varies significantly: subway construction monitoring data (such as groundwater level) has a resolution of 10m, UAV DEM data has a resolution of 0.5m, and street-level meteorological data has a resolution of 500m. To meet the requirements of "small-scale fine evaluation of special urban areas," an "inverse distance weighted interpolation method" was adopted to uniformly interpolate the 500m resolution meteorological data and the 10m resolution monitoring data to a grid scale of 100m×100m (matching urban community management units), ensuring that all data correspond to the same grid and avoiding the problem of "mismatch between rainfall data and slope data in the construction area" caused by different scales.

[0105] Geographic grid-based standardization (eliminating differences in the magnitude of city data):

[0106] Standardized calculations are performed using "maximum hourly rainfall intensity of the above-mentioned construction area grid" and "slope gradient" as examples:

[0107] Rainfall intensity standardization: Assume the original dataset of the maximum hourly rainfall intensity for all 100m×100m grids along Metro Line 3 is as follows: The mean of the set was calculated to be μ = 25 mm, and the standard deviation was σ = 12 mm; the original rainfall intensity of the grid in the construction area was X = 45 mm. Substituting these values ​​into the formula:

[0108] ;

[0109] Slope standardization: Assume the mean μ = 18° and the standard deviation σ = 10° of the original slope data set for all grids along the line; the original slope X = 35° of the grid in the construction area. Substitute these values ​​into the formula:

[0110] ;

[0111] After standardization, the rainfall intensity and slope gradient data for the construction area grid are 1.67 and 1.7, respectively. Similarly, the building density (original value 62%, assuming regional mean μ=35% and standard deviation σ=18%) can be calculated. Other data, such as groundwater level and depth, are standardized, and each 100m×100m grid forms a dataset with complete standardized features that is adapted to the evaluation of special urban areas.

[0112] In step S3, the process of constructing the geological hazard risk assessment model includes:

[0113] The feature extraction module uses a convolutional neural network to extract fault distribution features from geological structural data, topographic relief features from topographic data, rainfall distribution features from meteorological and hydrological data, and engineering construction features from human activity data.

[0114] The spatiotemporal correlation module uses a gated loop unit to capture the time-series variation characteristics of meteorological and hydrological data and the spatiotemporal distribution patterns of historical disaster data;

[0115] The risk prediction module integrates the extracted fault distribution features, terrain undulation features, rainfall distribution features, and engineering construction features through a fully connected layer, and maps the integrated features into a risk index through linear transformation and activation function.

[0116] The feature extraction module employs a convolutional neural network (CNN) to accurately extract core features such as fault distribution from geological structural data, topographic relief from topographic data, rainfall distribution from meteorological and hydrological data, and engineering construction from human activity data, avoiding the subjectivity and limitations of manual feature extraction. The spatiotemporal correlation module uses a gated recurrent unit (GRU) to effectively capture the temporal dynamic features of meteorological and hydrological data (such as the temporal variation of rainfall intensity) and the spatiotemporal distribution patterns of historical disaster data, solving the problem that traditional models struggle to handle spatiotemporal coupling relationships. The risk prediction module fuses multiple extracted features through a fully connected layer and maps them to a risk index through linear transformation and activation function, achieving efficient integration of multi-source features. At the same time, the model is optimized by combining the loss function of prediction error (mean squared error) and model complexity (regularization term), ensuring prediction accuracy while preventing overfitting, improving the model's generalization ability and reliability in geological disaster risk assessment in special urban areas (such as subway lines and residential areas with steep slopes), and providing efficient and stable model support for accurate assessment.

[0117] Taking the geological hazard risk assessment of urban residential areas with steep slopes (a special type of area: complex terrain and dense population, prone to landslides) as an example, the specific process of model construction and application is explained as follows:

[0118] Applications of the Feature Extraction Module (CNN):

[0119] Multi-source data in a 100m×100m grid for the residential area were selected as input, including: geological structure data (fault distribution raster map, resolution 10m), topographic data (slope / elevation raster map, resolution 5m), meteorological and hydrological data (rainfall intensity distribution map for the past year, resolution 50m), and human activity data (building distribution / slope protection engineering raster map, resolution 10m).

[0120] For topographic data, i.e. slope raster map: Convolutional layer 1 is constructed using a 3×3 small-scale convolutional kernel (stride 1, activation function ReLU) to extract local topographic details such as "35°+ steep slope edge" and "20°- gentle slope area", and the output feature map size is 98×98×64 (64 channels).

[0121] For geological structural data, namely fault distribution raster map: a 5×5 large-scale convolution kernel (stride 1, activation function ReLU) is used to construct convolution layer 2, extract regional geological structural features such as "dense fault zone at the bottom of the slope" and "fault-free stable zone", and output feature map size is 96×96×64;

[0122] Repeat the above convolution operation on the rainfall distribution and building distribution data respectively to obtain the rainfall distribution feature map (96×96×64) and the engineering construction feature map (98×98×64).

[0123] After dimensionality reduction by a 2×2 max pooling layer, the four types of feature maps are flattened into vectors: fault distribution feature vector F1 (dimension 1×36864), topographic relief feature vector F2 (dimension 1×36864), rainfall distribution feature vector F3 (dimension 1×36864), and engineering construction feature vector F4 (dimension 1×36864).

[0124] Applications of the Space-Time Relational Module (GRU):

[0125] Monthly time-series data of the residential area from 2020 to 2024 were selected as input, including: meteorological and hydrological data (maximum monthly rainfall intensity Xt, t=1 to 48 months) and historical disaster data (number of landslides per month Yt). The GRU module was set with 32 hidden layer neurons and 2 input dimensions (rainfall intensity + number of disasters).

[0126] Reset door calculation: ,in (Input → Reset gate weights, dimension 32×2) (Hidden layer → Reset gate weights, dimension 32×32) are the training parameters. , The bias term (dimension 32×1) is σ, where σ is the sigmoid function. The hidden layer state at time t-1 (initially h0 is a vector of all zeros);

[0127] Update gate calculation: The parameter definitions are the same as those for the reset gate;

[0128] Candidate hidden state: , For element-wise product, For activation functions;

[0129] Hidden layer state: .

[0130] After 48 time steps of training, the GRU captured the temporal pattern of "rainfall intensity > 50 mm / month from June to August → number of landslides ≥ 2 times" and the spatiotemporal pattern of "high and steep slopes + high rainfall areas → high incidence of landslides", and output the spatiotemporal correlation feature vector F5 (dimension 1×32).

[0131] Applications of the risk prediction module (fully connected layer):

[0132] Feature fusion: Concatenate F1, F2, F3, F4 (1×36864 each) with F5 (1×32) to obtain the fused feature vector. (Dimension 1 × (36864 × 4 + 32) = 1 × 147488);

[0133] Linear transformations and mappings: Fully connected layer 1 (1024 neurons, ReLU activation) Transformation: (W1 is a 1024×147488 weight matrix, b1 is a 1×1024 bias); Fully connected layer 2 (1 neuron, sigmoid activation) maps F6 to a risk index:

[0134] (W2 is a 1×1024 weight vector, b2 is the bias), and finally outputs the risk index of a certain grid in the residential area (e.g., 0.85, which is high risk).

[0135] Applications of model optimization (loss function):

[0136] 1000 grid samples (N=1000) were selected from this residential area, y i The actual risk value for the grid (based on historical disaster loss calibration, such as 0.9 = extremely high risk, 0.2 = low risk). The model's predicted values ​​represent the number of model parameters. The regularization coefficient λ = 0.001.

[0137] Mean square error term: After a certain round of training, the calculated value was 0.028;

[0138] Regularization term: The calculated value is 0.011;

[0139] Total loss: .

[0140] By optimizing the parameters through backpropagation, L was gradually reduced to below 0.015. The final model achieved an accuracy of 97% on the training set and 94% on the test set, effectively avoiding overfitting and ensuring the reliability of the evaluation across different grids in the residential area.

[0141] The geological hazard risk assessment model is optimized by combining prediction error and model complexity. First, the mean square error between the predicted and actual values ​​is calculated. Then, a regularization term is added to the model parameters to prevent overfitting. Specifically, optimization is performed using the following loss function:

[0142] ;

[0143] Where N is the number of samples, y i This represents the true risk value. To predict the risk value, M represents the number of parameters in the geological hazard risk assessment model. These are the parameters for the geological hazard risk assessment model. The regularization coefficient is used.

[0144] The model optimizes the geological hazard risk assessment model by constructing a loss function based on the mean square error between the predicted and actual values ​​and the regularization term of the model parameters. This not only accurately quantifies the deviation between the model's prediction results and the actual risk situation through the mean square error, ensuring the model's prediction accuracy in geological hazard risk assessment in special urban areas (such as residential areas with steep slopes and areas with dense underground engineering), meeting the high requirements for risk assessment accuracy in such areas, but also limits the model parameter size through the regularization term. This effectively avoids the problem of decreased generalization ability in new areas or when data changes due to overfitting the training data (such as only adapting to historical data of a specific residential area). At the same time, it balances the model's prediction performance and complexity, allowing the model to have stable scenario adaptability while ensuring assessment accuracy, providing a reliable and universal model optimization solution for geological hazard risk assessment in special urban areas.

[0145] Taking the optimization of the geological hazard risk assessment model for a "residential area with steep slopes in the eastern part of the city (containing 1500 100m×100m grids, which has experienced multiple landslide disasters in the past and requires refined risk assessment)" as an example, the specific application process is as follows:

[0146] Determine the core parameters of the loss function:

[0147] Sample size N: Historical data (including geological structure, rainfall, historical disasters and other multi-source data) from 1500 grids in the residential area are selected as training samples. Each grid corresponds to one set of "input data - actual risk value", so N=1500;

[0148] True risk value Based on historical landslide disaster losses within the grid (such as building damage rate and evacuation range) and engineering survey results (such as slope stability level), the calibration rule is as follows: "No disaster record and stable slope = 0.1, occasional small landslide = 0.3, frequent landslide = 0.7, major landslide has occurred = 0.9", for example:

[0149] Grid 1 (slope gradient 38°, two minor landslides have occurred): y1=0.3;

[0150] Grid 500 (slope angle 22°, no disaster record and with support works): y500=0.1;

[0151] Grid 1200 (slope angle 45°, one major landslide occurred): y1200=0.9;

[0152] Predicted risk value Based on the output of the previously constructed CNN+GRU risk assessment model, the following are some grid prediction values ​​after a certain round of training:

[0153] Grid 1 Predicted Values ;

[0154] Grid 500 Predictions ;

[0155] Grid 1200 Predicted Values ;

[0156] Number of model parameters M: This model contains a CNN feature extraction layer (3 convolutional layers, totaling 2.4 × 10⁻⁶). 6 (Number of parameters), GRU spatiotemporal correlation layer (64 hidden neurons, totaling 1.2 × 10⁻⁶). 4 (Number of parameters), fully connected prediction layer (2 layers, totaling 8.5 × 10⁻⁶). 5 (Number of parameters), the total number of parameters is M = 3.262 × 10-1 6 ;

[0157] Regularization coefficient λ: Considering the complexity of the data in this residential area (high dimensionality of multi-source data but uneven sample distribution), we take λ=0.0005 (to avoid over-regulation leading to underfitting of the model, or under-regulation failing to prevent overfitting).

[0158] Calculate the loss function value step by step:

[0159] Step 1: Calculate the mean squared error term (to measure prediction bias):

[0160] According to the loss function formula, the mean squared error term is: .

[0161] First calculate 1500 grid cells ,For example:

[0162] Grid 1: ;

[0163] Grid 500: ;

[0164] Grid 1200: ;

[0165] Sum the squared deviations of all grids to obtain .

[0166] Substitute into the formula to calculate the mean square error term: .

[0167] Step 2: Calculate the regularization term (to control model complexity):

[0168] The regularization term formula is ,in This is the sum of squares of all model parameters (this value is stable at 2800 after parameter updates during model training).

[0169] Substituting λ=0.0005, we calculate: 0.0005×2800=1.4.

[0170] Step 3: Calculate the total loss L

[0171] The total loss is the sum of the mean squared error term and the regularization term, i.e.:

[0172] L = 0.00064 + 1.4 = 1.40064;

[0173] Verify the optimization effect of the loss function:

[0174] Without regularization: The mean squared error of the model on the training set can be reduced to 0.0003, but on the 300 newly added test grids (not involved in training) in the residential area, the mean squared error rises to 0.0032, indicating obvious overfitting (the model only "memorizes" the training grid data and cannot accurately predict the risk of the new grid).

[0175] After adding a regularization term: the parameters are adjusted through backpropagation using the Adam optimizer, causing the total loss L to gradually decrease. By the 80th training iteration, L drops to 0.42. At this point, the mean squared error of the training set is 0.0007, and the mean squared error of the test set is 0.0008. The overfitting problem is significantly improved. The model can not only accurately evaluate the risk of this steep slope residential area, but also be transferred to similar slope residential areas in the western part of the city (mean squared error of the test set is 0.0009), meeting the scenario adaptation requirements for risk assessment in special urban areas.

[0176] The feature extraction module employs a multi-scale convolutional neural network structure, including:

[0177] The first convolutional layer uses small-scale convolutional kernels to extract local terrain detail features from the terrain data;

[0178] The second convolutional layer uses large-scale convolutional kernels to extract regional geological structural features from geological structural data;

[0179] By fusing local terrain detail features extracted from the first convolutional layer and regional geological structure features extracted from the second convolutional layer through a feature pyramid network, a comprehensive feature representation for geological hazard risk assessment is formed.

[0180] The feature extraction module employs a multi-scale convolutional neural network combined with a feature pyramid network. Small-scale convolutional kernels accurately extract local topographic details from topographic data (such as abrupt slope changes, small depressions, and boundaries of densely built-up areas in urban special regions). Large-scale convolutional kernels effectively capture regional geological structural features from geological data (such as the distribution of large-scale fault zones and regional lithological variation trends). The feature pyramid network then achieves deep fusion of local details and regional features, avoiding the omission of key risk factors in single-scale feature extraction (such as ignoring the impact of regional faults when using only small-scale features, or missing local slope hazards when using only large-scale features). This results in a more comprehensive feature representation that better reflects the complex topography and intertwined local and regional risk factors in urban special regions (such as residential areas with steep slopes and subway lines). This provides higher-quality feature support for subsequent spatiotemporal correlation and risk prediction, further improving the accuracy and reliability of geological disaster risk assessment.

[0181] Taking the geological hazard risk assessment of "a residential area with steep slopes in the eastern part of the city (divided into 100m×100m grids, with the risk of local slope hazards and overlapping regional fault zones)" as an example, the specific feature extraction process is as follows:

[0182] Determine the input data and multi-scale CNN parameters:

[0183] Input data: Select 5m resolution digital elevation model (DEM) data of the residential area (representing topography and landforms, used to extract local details) and 10m resolution geological structure fault distribution map (used to extract regional features). The data size corresponding to each 100m×100m grid is: DEM data 20×20 pixels (100m÷5m=20), fault distribution map 10×10 pixels (100m÷10m=10).

[0184] Multi-scale CNN parameters: The first convolutional layer (small scale) uses a 3×3 convolutional kernel, stride=1, padding=1 (to ensure the feature map size remains unchanged after convolution), padding=1 means that 1 row / column of 0 pixels is filled at each edge of the input feature map, the number of output channels=64, and the activation function is ReLU; The second convolutional layer (large scale) uses a 7×7 convolutional kernel, stride=1, padding=3 (again to keep the feature map size unchanged), padding=31 means that 3 rows / columns of 0 pixels are filled at each edge of the input feature map, the number of output channels=64, and the activation function is ReLU; The Feature Pyramid Network (FPN) adopts a fusion method of "upsampling + feature addition", and the upsampling method is bilinear interpolation.

[0185] The first convolutional layer (small scale) extracts local terrain detail features:

[0186] Data input: Input 20×20 pixel DEM data (each pixel value represents the elevation of the corresponding 5m×5m area) into the first convolutional layer;

[0187] Convolution calculation: Local sliding calculations are performed on the DEM data using a 3×3 convolution kernel. Each convolution window covers 3×3=9 pixels (corresponding to an actual 15m×15m range), capturing local elevation changes, for example:

[0188] If the pixel elevation values ​​within a certain window are [230, 232, 235; 228, 230, 233; 225, 227, 230], after calculation by the convolution kernel weights (such as the edge detection kernel [[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]]), the output value is = (230 × (-1) + 232 × (-1) + ... + 230 × (-1)) + (230 × 8) = 24. This value indicates that there is a sudden change in elevation in this local area (corresponding to a local steep slope within an actual 15m × 15m range).

[0189] Output feature map: After ReLU activation, the output is a 20×20×64 local terrain detail feature map F1, where each channel corresponds to a local feature (such as local slope abrupt change, small water catchment depression, gentle area boundary, etc.).

[0190] The second convolutional layer (large scale) extracts regional geological structural features:

[0191] Data input: Input a 10×10 pixel fault distribution map (pixel value 1 represents a fault area, and 0 represents a non-fault area) into the second convolutional layer;

[0192] Convolution calculation: A large-scale sliding calculation is performed on the fault distribution map using a 7×7 convolution kernel. Each convolution window covers 7×7=49 pixels (corresponding to an actual 70m×70m range), capturing the regional fault distribution trend, for example:

[0193] If there are 12 pixels with a value of 1 in a certain window (that is, the 70m×70m range contains 12 10m×10m fault regions), after calculation by the convolution kernel weights (such as average pooling kernel, each weight = 1 / 49), the output value = 12×(1 / 49)≈0.245. This value indicates that the fault density in this region is relatively high (corresponding to a local segment of a large-scale fault zone in reality).

[0194] Output feature map: After ReLU activation, output a 10×10×64 regional geological structure feature map F2. Each channel corresponds to a regional feature (such as fault zone coverage, lithological zoning boundary, regional geological stability level, etc.).

[0195] Feature Pyramid Network (FPN) fusion features:

[0196] Upsampling matching size: To achieve the fusion of F1 (20×20×64) and F2 (10×10×64), F2 is upsampled to 20×20×64 using bilinear interpolation to obtain the upsampled feature map. (During the upsampling process, each pixel value of the original 10×10 pixels is mapped to a 2×2 region of 20×20 pixels, and the intermediate pixel values ​​are supplemented by interpolation to ensure that the regional feature trend remains unchanged.)

[0197] Feature fusion: The fusion is performed using "element-wise addition + batch normalization (BatchNorm)", with the following formula:

[0198] ;

[0199] ;

[0200] in This is a feature map of a certain channel of F1. for The corresponding channel feature map, BatchNorm is used to standardize the fused feature values ​​(to avoid large differences in feature values ​​at different scales affecting subsequent calculations).

[0201] Output comprehensive features: The final output is a comprehensive feature map of 20×20×64. After flattening, we obtain a comprehensive feature vector of 1×(20×20×64)=1×25600. This vector contains both detailed risk factors such as "local steep slope" and background risk factors such as "regional fault zone".

[0202] Fusion effect verification: Single-scale feature defects: If only F1 (local features) is used, the model will mistakenly identify "local steep slopes in fault-free areas" as high risk (ignoring the background of regional geological stability); if only F2 (regional features) is used, the model will miss "small slope hazards at the edge of fault zones" (ignoring local key risk points).

[0203] Integration of features and advantages: Adopting Subsequently, the model improved the accuracy of risk identification in areas with overlapping "regional fault zones + local steep slopes" from a single scale, accurately matching the risk characteristics of "intertwined local hazards and regional background" in urban high and steep slope residential areas, thus verifying the effectiveness of multi-scale fusion.

[0204] In step S4, the dynamic weight allocation method includes:

[0205] The weights of geological stability indicators are calculated based on geological structural data;

[0206] The weights of rainfall impact indicators are calculated based on meteorological and hydrological data;

[0207] The weights of the impact indicators of engineering activities are calculated based on human activity data;

[0208] The weights of disaster recurrence probability indicators are calculated based on historical disaster data;

[0209] An importance matrix of evaluation indicators is constructed, and the initial weights of each evaluation indicator are determined by a combination of expert scoring and entropy methods.

[0210] The weights are dynamically adjusted by introducing a time decay factor and a spatial variation factor;

[0211] The weight calculation is achieved by multiplying the initial weight by the time decay factor and the spatial variation factor, specifically through the following process:

[0212] ;

[0213] Among them, W i (t) represents the dynamic weight of the i-th evaluation index at time t, W i0 Let α be the initial weight of the i-th evaluation index, α be the time decay factor, and β be the spatial variation factor of the i-th evaluation index.

[0214] The dynamic weight allocation method first calculates the weights of corresponding evaluation indicators (geological stability indicators, rainfall impact indicators, engineering activity impact indicators, and disaster recurrence probability indicators) based on different types of data (geological structure, meteorology and hydrology, human activities, and historical disasters). Then, it comprehensively determines the initial weights through expert scoring (combining experience in disaster prevention) and entropy methods (relying on objective data laws), balancing subjectivity and objectivity to avoid the one-sidedness of a single weight determination method. Simultaneously, it introduces a time decay factor (adapting to the temporal changes of risk factors in specific urban areas, such as increased weight for rainfall during the rainy season and increased weight for engineering activities during the construction period) and a spatial variation factor (adapting to spatial differences, such as increased weight for geological stability in fault-prone areas and increased weight for engineering activities in densely built-up areas) to dynamically adjust the weights using formulas. By enabling weights to change flexibly in time and space, the traditional fixed weights can effectively solve the problem that they cannot adapt to the dynamic changes and large spatial differences of risk factors in special urban areas (such as residential areas with steep slopes). This makes the weights of evaluation indicators more in line with actual risk scenarios and improves the pertinence and accuracy of geological disaster risk assessment.

[0215] Taking the "residential area on a high and steep slope in the eastern part of the city (divided into 100m×100m grids, totaling 1500 grids, with a focus on four types of evaluation indicators: geological stability indicator A, rainfall impact indicator B, engineering activity indicator C, and disaster recurrence probability indicator D)" as an example, the specific application process is as follows:

[0216] The evaluation indicators and their corresponding data sources are determined as follows: Geological stability indicator A is calculated based on geological structure data (fault density, lithology type) and reflects the basic geological risk of the region; Rainfall impact indicator B is calculated based on meteorological and hydrological data (cumulative rainfall in the past month, frequency of short-term heavy rainfall) and reflects the risk induced by short-term rainfall; Engineering activity indicator C is calculated based on human activity data (intensity of slope support construction, amount of building demolition) and reflects the risk of human activity interference; Disaster recurrence probability indicator D is calculated based on historical disaster data (number of landslides in the past 5 years, recurrence interval) and reflects the historical risk inertia.

[0217] Calculate initial weights (A combination of expert scoring and entropy methods):

[0218] Step 1: Expert scoring (subjective weighting) ):

[0219] Five geological disaster experts were invited to rate the importance of four indicators on a scale of 1 to 5 (5 being the most important). The average scores for each indicator were obtained as follows: Indicator A: 4.2 points, Indicator B: 3.8 points, Indicator C: 3.5 points, and Indicator D: 2.5 points. The total score was calculated as 4.2 + 3.8 + 3.5 + 2.5 = 14 points. The subjective weights were then calculated by dividing the average score of each indicator by the total score: Subjective weight for Indicator A: 4.2 / 14 = 0.30, for Indicator B: 3.8 / 14 = 0.27, for Indicator C: 3.5 / 14 = 0.25, and for Indicator D: 2.5 / 14 = 0.18.

[0220] Step 2: Calculate the objective weights using the entropy method :

[0221] Based on the index data of 1500 grids, the entropy value of each index is first calculated. (in Given the normalized value of the i-th index in the k-th grid (n=1500), we obtain the entropy values ​​of each index: index A is 0.65, index B is 0.58, index C is 0.82, and index D is 0.95.

[0222] Then calculate the coefficient of variation for each indicator. The coefficients of variation are: A = 0.35, B = 0.42, C = 0.18, and D = 0.05. The sum of the coefficients of variation is 0.35 + 0.42 + 0.18 + 0.05 = 1.0. Finally, the formula is used... Calculate objective weights:

[0223] Indicator A is 0.35 / 1.0=0.35, Indicator B is 0.42 / 1.0=0.42, Indicator C is 0.18 / 1.0=0.18, and Indicator D is 0.05 / 1.0=0.05.

[0224] Step 3: Integrate the initial weights :

[0225] The initial weight of indicator A is calculated by taking the weighted average of subjective and objective weights (each accounting for 0.5). ;

[0226] Indicator B ;

[0227] Indicator C ;

[0228] Indicator D .

[0229] Determine the time decay factor α in relation to time t:

[0230] The time decay factor α reflects the decay trend of indicator weights over time. Indicators that are prone to short-term changes (such as rainfall impact indicator B and engineering activity indicator C) have smaller α values ​​(slower weight decay), while indicators that are stable in the long term (such as geological stability indicator A and disaster recurrence probability indicator D) have larger α values ​​(faster weight decay). The α value is set based on the actual situation of this residential area.

[0231] Indicator A (Geological data is relatively stable within six months, but its weight decays rapidly), Index B (Rainfall data shows significant intra-month variations, with slow weight decay), Indicator C (Engineering activities are relatively active within the quarter, and the weight decays slowly), Indicator D (Historical disaster data is stable from year to year, but the weight decays relatively quickly).

[0232] Time t is in units of "data update cycle". If the current time is July 2024 (the rainy season in this region), and the interval since the last data update (June 2024) is 1 month, then t=1.

[0233] Calculate the spatial variation factor :

[0234] Two typical grids were selected from the residential area (Grid 1: densely populated with faults and heavy rainfall during the rainy season; Grid 2: no faults and no construction activity). The average values ​​of four indicators were first calculated for the entire study area. with standard deviation :

[0235] Regional mean of indicator A (Fault density normalized value) and standard deviation ;

[0236] Regional mean of indicator B (Normalized rainfall), standard deviation ;

[0237] Regional mean of indicator C (Normalized value of construction strength), standard deviation ;

[0238] Regional mean of indicator D (Normalized recurrence probability), standard deviation .

[0239] Substitute the above values ​​into the formula for the spatial variation factor to calculate:

[0240] Indicator A Indicator B Indicator C Indicator D .

[0241] Calculate dynamic weights :

[0242] Grid 1 (Dense faults + heavy rainfall during the rainy season, requiring close attention to geological stability and the impact of rainfall):

[0243] Dynamic weights of indicator A:

[0244] ;

[0245] Dynamic weights of indicator B:

[0246] ;

[0247] Dynamic weights of indicator C:

[0248] ;

[0249] Dynamic weights of indicator D:

[0250] .

[0251] Grid 2 (no faults + no construction, low geological stability and low risk of engineering activities):

[0252] Because grid 2 has no faults (the actual value of index A is close to 0, and the standard deviation of this grid within the region is...) (Minimum value of 0.1 for the region) No construction activity (actual value of index C is close to 0, corresponding to standard deviation) Taking the minimum value in the region (0.05), recalculate the spatial variation factors of these two indicators: , .

[0253] Substituting into the formula, we get: index A ;

[0254] Indicator C ;

[0255] The spatial variation factors of indicators B and D remain unchanged, and their dynamic weights are 0.453 and 0.259, respectively.

[0256] Validation of the advantages of dynamic weights:

[0257] If fixed weights (i.e., initial weights) are used In grid 1, the weights of geological stability index A and rainfall impact index B are only 0.325 and 0.345, respectively, which fails to reflect the high risk of the combination of "dense faults + rainy season". However, under dynamic weighting, the weights of these two indices increase to 0.551 and 0.453, respectively, which better reflects the actual risk situation. For grid 2, the weights of geological stability index A and engineering activity index C are 0.325 and 0.215 under fixed weighting, which is prone to over-evaluation. Under dynamic weighting, the weights of these two indices decrease to 0.389 and 0.202, which better reflects the low-risk characteristics of this grid, namely "no faults and no construction". This fully verifies the adaptability of this method to the spatiotemporal differences of special urban areas.

[0258] The method for calculating the spatial variation factor includes:

[0259] Calculate the spatial variation characteristics of fault density based on geological structural data;

[0260] Calculate the spatial distribution characteristics of slope variation based on topographic data;

[0261] Calculate the spatial variation characteristics of rainfall based on meteorological and hydrological data;

[0262] Calculate the spatial distribution characteristics of building density based on human activity data;

[0263] The spatial variation factor of each evaluation index is calculated based on spatial heterogeneity. This is achieved by adding 1 to the ratio of the standard deviation to the mean of each evaluation index. The formula for calculating the spatial variation factor is as follows:

[0264] ;

[0265] Where σ is the standard deviation of the i-th evaluation index within the study area. Let be the mean of the i-th evaluation index within the study area;

[0266] By using four core data categories—geological structure, topography, meteorology and hydrology, and human activities—this study calculates spatial variations in fault density, slope distribution, rainfall, and building density. A spatial variation factor is then constructed using spatial heterogeneity (the ratio of standard deviation to mean plus 1). This approach accurately quantifies the spatial heterogeneity of different grids within specific urban areas (such as residential areas on steep slopes), avoiding the problem of traditional weight adjustments neglecting "multiple risk differences within small areas." It provides an objective basis for dynamic weight allocation that aligns with real-world spatial scenarios, allowing the weights of each evaluation indicator to be flexibly adjusted according to the specific environment of the grid (such as areas with dense faults versus areas without faults, areas with concentrated rainfall versus areas with little rainfall). This further improves the spatial accuracy of geological hazard risk assessment and is well-suited to the significant local risk differences in specific urban areas.

[0267] Taking the "situational area on steep slopes in the eastern part of the city (1500 100m×100m grids, characterized by uneven fault distribution, large slope differences, localized concentrated rainfall, and varying building densities)" as an example:

[0268] Determine the basic data for calculating the spatial variation factor:

[0269] Four types of core data (covering 1500 grids) were selected as the calculation benchmarks: geological structural data used "grid-internal fault density (normalized value, range 0-1, higher values ​​indicate denser faults)", topographic data used "grid-average slope (unit: °)", meteorological and hydrological data used "grid-average monthly rainfall (unit: mm)", and human activity data used "grid-average building density (unit: %)". Regional averages for each type of data were obtained by statistically analyzing the data from the 1500 grids. with regional standard deviation :

[0270] Fault density: regional mean (i.e., the average grid fault density across the entire region is 0.4), regional standard deviation (Reflecting the degree of difference in density between different grid faults);

[0271] Average slope: regional average (The average slope of the entire grid area is 25°), regional standard deviation ;

[0272] Average monthly rainfall: regional average (Average monthly rainfall across the entire grid area is 120mm), regional standard deviation ;

[0273] Building density: Regional average (Average building density across the entire grid is 30%), regional standard deviation .

[0274] Calculate spatial variation factor by type:

[0275] Calculations based on geological structural data (fault density):

[0276] Regional average of fault density Regional standard deviation Substituting into the formula, we obtain the spatial variation factor corresponding to the fault density:

[0277] ;

[0278] This value indicates that the spatial difference in fault density across the entire region is significant—some grids have dense faults (such as the grid at the bottom of the slope), while some grids have no faults (such as the grid in the gently sloping residential area).

[0279] Calculated based on topographic data (average slope):

[0280] Substitute the average slope , The spatial variation factor corresponding to the slope is obtained as follows:

[0281] ;

[0282] This value reflects that there are certain differences in slope within the region. The slope characteristics of the steep slope grid (slope > 30°) and the gentle grid (slope < 20°) are clearly distinguishable, but the degree of difference is less than that of the fault density.

[0283] Calculated based on meteorological and hydrological data (monthly average rainfall):

[0284] Substitute the average monthly rainfall , The spatial variability factor corresponding to the rainfall is obtained as follows:

[0285] ;

[0286] This value reflects the local concentration of regional rainfall. The windward slope grid (monthly average rainfall > 150 mm) receives more rainfall due to topographic lifting, while the leeward slope grid (monthly average rainfall < 90 mm) receives less rainfall, with moderate spatial differences.

[0287] Calculated based on human activity data (building density):

[0288] Substitute building density , The spatial variation factor corresponding to building density is obtained as follows:

[0289] ;

[0290] This value indicates a moderate difference in building density across the region—clearly distinguishing the intensity of human activity between the concentrated residential area grid (building density > 40%) and the evacuation area grid near the slope (building density < 15%).

[0291] The specificity of spatial variability factors is verified by combining specific grids:

[0292] Two typical grids of the residential area were selected (Grid 1: slope bottom + dense faults + concentrated rainfall; Grid 2: gentle area + no faults + sparse buildings). The local standard deviation of the corresponding data for the two grids (reflecting the degree of difference of the grid in the region) was recalculated to verify the adjustment effect of the spatial variability factor.

[0293] Grid 1 (High-risk feature):

[0294] The measured fault density of the grid is 0.8 (far higher than the regional mean of 0.4), and its corresponding local standard deviation is... (Because this grid belongs to a densely clustered fault zone, it differs little from similar grids, but differs greatly from the entire region), substituting into the formula yields... Compared to the regional average A higher value indicates a stronger spatial specificity of the fault density in the grid, and the weight of the "geological stability index" in the subsequent dynamic weighting will be further increased.

[0295] Grid 2 (Low-risk feature):

[0296] The measured fault density of the grid is 0.1 (far lower than the regional average of 0.4), with a local standard deviation of... Substituting into the formula, we get Compared to the regional average A lower value indicates that the fault density of the grid has no significant spatial risk, and the weight of the subsequent "geological stability index" will be reduced accordingly.

[0297] Verification of the advantages of spatial variation factors:

[0298] Without the spatial variability factor used in this case, the dynamic weights would ignore the spatial differences between grids—for example, grids 1 and 2 would have the same weight for the "geological stability index," failing to reflect the spatial risk differences in fault density; however, with the introduction of the spatial variability factor, the spatial differences in fault density between grids 1 and 2 would be overlooked. Make the weight of this indicator lower than that of grid 2 ( The system is highly accurate in matching the actual spatial risks of the two grids, avoiding a one-size-fits-all weight allocation, and making the evaluation results more in line with the local risk characteristics of special urban areas.

[0299] Furthermore, the calculation process of the geological hazard risk index in step S5 includes:

[0300] Features are extracted from multi-source geological hazard data and a preliminary risk score is obtained through a geological hazard risk assessment model;

[0301] A weighted comprehensive evaluation is performed by combining dynamic weights calculated based on different data types;

[0302] The scores are mapped to the [0,1] interval using the sigmoid function to form a standardized risk index. This is achieved by calculating the weighted sum of the scores of each evaluation indicator and their dynamic weights, and then transforming the result using the sigmoid function. The formula for calculating the standardized risk index is as follows:

[0303] ;

[0304] Where R is the geological disaster risk index, W i S is the dynamic weight of the i-th evaluation index. iLet b be the score for the i-th evaluation indicator, b be the bias term, and n be the number of evaluation indicators.

[0305] First, features are extracted from multi-source geological hazard data using a geological hazard risk assessment model to obtain preliminary risk scores, ensuring that the scores cover the core risk factors of special urban areas (such as residential areas with steep slopes). Then, a weighted comprehensive assessment is performed by combining dynamic weights calculated based on different data types, allowing the scores to adapt to the spatiotemporal risk differences of the grid. Next, the weighted scores are mapped to the [0,1] interval using the sigmoid function to form a standardized risk index, which not only eliminates the evaluation confusion caused by the difference in the magnitude of different indicator scores, but also makes the risk level intuitive and quantifiable. At the same time, the generated risk level zoning map can clearly show the distribution of high, medium and low risk grids in the region, adapting to the needs of refined management in special urban areas, making it easier for staff to quickly locate high-risk areas and formulate targeted prevention and control measures, thereby improving the practicality and decision support capabilities of geological hazard risk assessment.

[0306] Taking the "residential area on steep slopes in the eastern part of the city (1500 grids of 100m×100m, with the core evaluation indicators being geological stability (A), rainfall impact (B), engineering activities (C), and probability of disaster recurrence (D)" as an example":

[0307] Determine the initial risk score S i :

[0308] Preliminary risk score S i Feature extraction results from the CNN+GRU model, combined with data normalization (mapped to the [0,1] interval), reflect the actual risk level of each indicator:

[0309] Grid 1 (bottom of slope + dense faults + heavy rainfall during the rainy season + support construction in place):

[0310] The geological stability A (dense faults, lithology of strongly weathered mudstone) has an SA of 0.8;

[0311] The impact of rainfall on B (monthly average rainfall of 180 mm, exceeding the regional average by 50%) is represented by SB = 0.9.

[0312] Project Activity C (Slope Protection Construction, Monthly Excavation Volume 800m³) 3 SC=0.6;

[0313] The probability of disaster recurrence D (two small landslides in the past 5 years) has an SD of 0.4;

[0314] Grid 2 (flat area + no faults + low rainfall + no construction):

[0315] The geological stability A (no faults, lithology is moderately weathered sandstone) has an SA of 0.2;

[0316] The impact of rainfall on B (monthly average rainfall of 80 mm, 33% lower than the regional average) is SB=0.3;

[0317] For project activity C (no construction, only daily residence), SC=0.1;

[0318] The probability of disaster recurrence D (no disaster record in the past 5 years) has an SD of 0.1.

[0319] Determine dynamic weights With bias term b:

[0320] Dynamic weights Using the previously calculated results, the bias term b is calibrated using regional historical disaster data (if the weighted sum of historical high-risk grids in the region is generally low, b is adjusted to improve the index's distinguishability). Here, considering the actual situation of this residential area, b is set to -0.2.

[0321] Grid 1: WA=0.551, WB=0.453, WC=0.403, WD=0.259;

[0322] Grid 2: WA=0.389, WB=0.453, WC=0.202, WD=0.259.

[0323] Calculate the standardized risk index R:

[0324] Grid 1 (High-risk feature):

[0325] The first step is to calculate the weighted sum. :

[0326] 0.551×0.8+0.453×0.9+0.403×0.6+0.259×0.40.4408+0.4077+0.2418+0.1036=1.1939;

[0327] The second step is to substitute the values ​​into the sigmoid formula (n=4, b=-0.2):

[0328] ;

[0329] Grid 2 (Low-risk feature):

[0330] The first step is to calculate the weighted sum. :

[0331] 0.389×0.2+0.453×0.3+0.202×0.1+0.259×0.10.0778+0.1359+0.0202+0.0259=0.2598;

[0332] The second step is to substitute the values ​​into the sigmoid formula:

[0333] ;

[0334] Risk levels are determined based on the standardized risk index R (in conjunction with urban disaster prevention standards): R < 0.3 indicates low risk, 0.3 ≤ R < 0.6 indicates medium risk, and R ≥ 0.6 indicates high risk.

[0335] Grid 1 has an R1 of approximately 0.730, and is classified as a high-risk grid, marked in red on the zoning map.

[0336] Grid 2 has an R² of approximately 0.515, and is classified as a medium-risk grid, marked in yellow.

[0337] After the calculation of 1,500 grids in the entire area is completed, a color-coded risk level zoning map is generated. It can be seen that "the fault zone at the bottom of the slope" is a red high-risk zone, "the flat residential area" is a yellow medium-risk zone, and "the open area far away from the slope" is a green low-risk zone.

[0338] Without the sigmoid function, the weighted sum for grid 1 is 1.1939 and for grid 2 it is 0.2598. The values ​​lack a unified scale, making it easy to mistakenly equate "weighted sum of 1.1939" with "extremely high risk" (in reality, it needs to be judged in conjunction with the overall level of the area). However, after using the sigmoid mapping, R is uniformly in the [0,1] interval. Staff can directly and quickly determine the risk level based on the value of R, and the zoning map can clearly locate high-risk clusters. This provides a clear basis for decisions such as "prioritizing the deployment of monitoring equipment in high-risk grids" and "strengthening patrols in medium-risk areas," adapting to the refined prevention and control needs of special urban areas.

[0339] In step S6, the dynamic update and early warning process includes:

[0340] Real-time acquisition of new meteorological and hydrological data to update rainfall impact assessments;

[0341] Real-time acquisition of new human activity data to update the impact assessment of engineering activities;

[0342] The geological hazard risk index was recalculated based on updated multi-source geological hazard data;

[0343] Set a risk threshold, and automatically trigger an early warning when the geological disaster risk index exceeds the risk threshold;

[0344] The parameters of the geological hazard risk assessment model are updated based on a sliding time window mechanism;

[0345] By acquiring new meteorological, hydrological, and human activity data in real time, the system can promptly capture dynamic changes in risk factors in special urban areas (such as residential areas with steep slopes) (e.g., short-term heavy rainfall, sudden increases in construction intensity), avoiding misjudgments due to data lag in traditional static assessments. Based on updated data, the risk index is recalculated to ensure the assessment results always reflect the current actual risk status. Setting risk thresholds automatically triggers early warnings, enabling rapid response to high-risk situations and buying time for disaster prevention (e.g., advance evacuation, slope reinforcement). Simultaneously, a sliding time window mechanism updates model parameters, allowing the model to continuously adapt to new data characteristics (e.g., changes in rainfall patterns during the rainy season), preventing a decline in model accuracy over long-term use. Ultimately, this achieves "real-time, dynamic, and early warning" risk assessment, fully meeting the needs of urban special areas where risk factors are highly variable and require rapid response.

[0346] Taking the "residential area on steep slopes in the eastern part of the city (1500 grids of 100m×100m, original high-risk grid 1: R1≈0.73, risk threshold set to 0.8)" as an example:

[0347] Get new data in real time (trigger update):

[0348] New meteorological and hydrological data: The meteorological station in this residential area detected a short-term heavy rainfall of 50 mm from 14:00 to 15:00 on [Date] (far exceeding the previous monthly average hourly rainfall of 2 mm / h). Therefore, the score and dynamic weight of rainfall impact index B need to be updated. Under the new rainfall data, the rainfall intensity score for grid 1 will be adjusted from [Previous Data]. Rise to Time decay factor of rainfall impact indicators The weight was adjusted from 0.7 to 0.6 (short-term rainfall has a stronger timeliness and a slower weight decay), and the dynamic weight was adjusted from... Rise to ;

[0349] New data on human activities: Construction units report in real time that "the excavation volume for slope protection construction around Grid 1 has increased from 800m." 3 / Monthly increase to 1200m 3 " / month (50% over the original plan)", requiring an update to the score of project activity indicator C, and the increased construction intensity will cause S C From 0.6 to Dynamic weights from Fine-tuning to (Increased construction intensity enhances the importance of indicators).

[0350] Recalculate the geological hazard risk index:

[0351] Based on the updated and Using the risk index formula (b = -0.2, n = 4), recalculate the risk index for grid 1:

[0352] The first step is to calculate the updated weighted sum:

[0353] =0.551×0.8 (Indicator A remains unchanged) + 0.52×0.95 (Indicator B is updated) + 0.43×0.8 (Indicator C is updated) + 0.259×0.4 (Indicator D remains unchanged) = 0.4408 + 0.494 + 0.344 + 0.1036 = 1.3824;

[0354] The second step is to substitute the values ​​into the sigmoid formula:

[0355] ;

[0356] (If the rainfall increases by another 30mm between 16:00 and 17:00, , Recalculated (Exceeding the risk threshold of 0.8).

[0357] Triggering early warning and response measures:

[0358] when When the threshold of 0.8 is exceeded, the system automatically triggers a Level 3 warning (based on risk index: 0.8-0.9 is a Level 3 warning, and above 0.9 is a Level 2 warning):

[0359] Warning push: Real-time push of warning information to the residential area management committee, construction unit and emergency department, the content of which includes "Grid 1 (bottom area of ​​slope) risk index 0.821, reaching level three warning, it is recommended to complete slope inspection and evacuate temporary construction personnel in the surrounding area within 1 hour";

[0360] Linkage measures: The system synchronously retrieves surveillance cameras around grid 1, automatically identifies the number of construction workers, and assists emergency departments in confirming the evacuation progress; at the same time, it pushes historical disaster data of the grid (two landslides in the past 5 years) to provide a reference for on-site disposal.

[0361] Updating model parameters via a sliding time window:

[0362] Set the sliding time window size to "3 months". After adding new data each month, incorporate the latest data within the window (such as rainfall and construction data from May to July 2024) into the model training, replacing the old data outside the window (such as data from February to April 2024).

[0363] Parameter updates: Retrain the weights of the GRU spatiotemporal correlation layer (e.g., adjust the weights of rainfall time series features from 0.3 to 0.45 to adapt to the rainfall pattern of the rainy season), and the convolution kernel parameters of the CNN feature extraction layer (e.g., adjust the weights of local slope change features from 0.25 to 0.3).

[0364] Performance Verification: Before the update, the model's risk prediction error for the scenario of "short-term heavy rainfall + sudden increase in construction intensity" was 8%. After the update, the error decreased to 4%, as predicted in the original model. (actual ), updated predicted value This makes the model more closely reflect actual risks and improves its long-term adaptability.

[0365] A geological hazard risk assessment system based on big data includes:

[0366] The data acquisition module is used to acquire multi-source geological disaster data, including geological structure data, topographic data, meteorological and hydrological data, human activity data, and historical disaster data.

[0367] The data processing module is used to perform spatiotemporal alignment and standardization processing on the multi-source geological disaster data to construct a spatiotemporal dataset of geological disasters;

[0368] The model building module is used to build a geological hazard risk assessment model based on a deep learning network. The geological hazard risk assessment model includes a feature extraction module, a spatiotemporal correlation module, and a risk prediction module.

[0369] The weight calculation module is used to calculate the weight coefficients of each evaluation indicator using a dynamic weight allocation method.

[0370] The risk calculation module is used to calculate the geological hazard risk index through the geological hazard risk assessment model and generate a risk level zoning map.

[0371] Update the early warning module to dynamically update and issue early warnings based on real-time monitoring data of risk assessment results.

[0372] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A geological hazard risk assessment method based on big data, characterized in that: Includes the following steps: S1. Acquire multi-source geological disaster data, including geological structure data, topographic data, meteorological and hydrological data, human activity data, and historical disaster data; S2. Perform spatiotemporal alignment and standardization on multi-source geological disaster data to construct a spatiotemporal dataset of geological disasters; S3. Construct a geological hazard risk assessment model based on a deep learning network. The geological hazard risk assessment model includes a feature extraction module, a spatiotemporal correlation module, and a risk prediction module. S4. Calculate the weight coefficients of each evaluation indicator using a dynamic weight allocation method; S5. Calculate the geological hazard risk index using the geological hazard risk assessment model and generate a risk level zoning map; S6. Dynamically update and issue early warnings based on real-time monitoring data for risk assessment results.

2. The geological hazard risk assessment method based on big data according to claim 1, characterized in that: In step S2, the spatiotemporal alignment and normalization process includes: Geological structural data are coded for fault density and lithological type. Slope, aspect, and elevation features are extracted from topographic data; Calculate rainfall intensity and hydrological parameters from meteorological and hydrological data; Classify human activity data by building density and land use type; Statistical analysis of disaster types and frequency of occurrence based on historical disaster data; The multi-source geological disaster data with different resolutions are unified to the same spatiotemporal scale through spatiotemporal interpolation methods; A standardized processing based on geographic grids is adopted to divide the study area into regular grid units, each of which contains complete multi-source geological hazard data features; The geographic grid-based standardization process is achieved by calculating the difference between the original data value and the mean of the data features, and then dividing by the standard deviation of the data features.

3. The geological hazard risk assessment method based on big data according to claim 1, characterized in that: In step S3, the process of constructing the geological hazard risk assessment model includes: The feature extraction module uses a convolutional neural network to extract fault distribution features from geological structural data, topographic relief features from topographic data, rainfall distribution features from meteorological and hydrological data, and engineering construction features from human activity data. The spatiotemporal correlation module uses a gated loop unit to capture the time-series variation characteristics of meteorological and hydrological data and the spatiotemporal distribution patterns of historical disaster data; The risk prediction module integrates the extracted fault distribution features, topographic relief features, rainfall distribution features, and engineering construction features through a fully connected layer, and maps the integrated features into a risk index through linear transformation and activation function.

4. The geological hazard risk assessment method based on big data according to claim 1, characterized in that: The geological hazard risk assessment model is optimized by combining prediction error and model complexity. First, the mean square error between the predicted value and the actual value is calculated, and then a regularization term is added to the model parameters to prevent overfitting.

5. The geological hazard risk assessment method based on big data according to claim 3, characterized in that: The feature extraction module employs a multi-scale convolutional neural network structure, including: The first convolutional layer uses small-scale convolutional kernels to extract local terrain detail features from the terrain data; The second convolutional layer uses large-scale convolutional kernels to extract regional geological structural features from geological structural data; By fusing local topographic details extracted from the first convolutional layer and regional geological structural features extracted from the second convolutional layer through a feature pyramid network, a comprehensive feature representation for geological hazard risk assessment is formed.

6. The geological hazard risk assessment method based on big data according to claim 1, characterized in that: In step S4, the dynamic weight allocation method includes: The weights of geological stability indicators are calculated based on geological structural data; The weights of rainfall impact indicators are calculated based on meteorological and hydrological data; The weights of the impact indicators of engineering activities are calculated based on human activity data; The weights of disaster recurrence probability indicators are calculated based on historical disaster data; An importance matrix of evaluation indicators is constructed, and the initial weights of each evaluation indicator are determined by a combination of expert scoring and entropy methods. The weights are dynamically adjusted by introducing a time decay factor and a spatial variation factor; The weight calculation is achieved by multiplying the initial weight by the time decay factor and the spatial variation factor.

7. The geological hazard risk assessment method based on big data according to claim 6, characterized in that: The method for calculating the spatial variation factor includes: Calculate the spatial variation characteristics of fault density based on geological structural data; Calculate the spatial distribution characteristics of slope variation based on topographic data; Calculate the spatial variation characteristics of rainfall based on meteorological and hydrological data; Calculate the spatial distribution characteristics of building density based on human activity data; The spatial variation factor of each evaluation index is calculated based on spatial heterogeneity, which is achieved by adding 1 to the ratio of the standard deviation to the mean of each evaluation index.

8. The geological hazard risk assessment method based on big data according to claim 1, characterized in that: In step S5, the calculation process of the geological hazard risk index includes: Features are extracted from multi-source geological hazard data and a preliminary risk score is obtained through a geological hazard risk assessment model; A weighted comprehensive evaluation is performed by combining dynamic weights calculated based on different data types; The initial risk score is mapped to the [0,1] interval by using the sigmoid function to form a standardized risk index. This is achieved by calculating the weighted sum of the scores of each evaluation indicator and the dynamic weights, and then transforming it using the sigmoid function.

9. The geological hazard risk assessment method based on big data according to claim 1, characterized in that: In step S6, the dynamic update and early warning process includes: Real-time acquisition of new meteorological and hydrological data to update rainfall impact assessments; Real-time acquisition of new human activity data to update the impact assessment of engineering activities; The geological hazard risk index was recalculated based on updated multi-source geological hazard data; Set a risk threshold, and automatically trigger an early warning when the geological disaster risk index exceeds the risk threshold; The parameters of the geological hazard risk assessment model are updated based on a sliding time window mechanism.

10. A geological hazard risk assessment system based on big data, wherein the system is applied in any one of the methods described in claims 1-9, characterized in that: include: The data acquisition module is used to acquire multi-source geological disaster data, including geological structure data, topographic data, meteorological and hydrological data, human activity data, and historical disaster data. The data processing module is used to perform spatiotemporal alignment and standardization on multi-source geological disaster data to construct a spatiotemporal dataset of geological disasters. The model building module is used to build a geological hazard risk assessment model based on a deep learning network. The geological hazard risk assessment model includes a feature extraction module, a spatiotemporal correlation module, and a risk prediction module. The weight calculation module is used to calculate the weight coefficients of each evaluation indicator using a dynamic weight allocation method. The risk calculation module is used to calculate the geological hazard risk index through the geological hazard risk assessment model and generate a risk level zoning map; Update the early warning module to dynamically update and issue early warnings based on real-time monitoring data of risk assessment results.

Images