Geological disaster risk assessment method and device based on remote sensing big data

Abstract

The invention relates to the field of geological disaster risk assessment, and discloses a geological disaster risk assessment method and device based on remote sensing big data, and the method comprises the following steps: S1, collecting disaster trigger factors related to multi-source data and geological disasters; s2, establishing a disaster risk distribution model in the source region; s3, constructing a disaster risk distribution model of the target area in the target area based on the limited observation data; s4, calculating a transmission mapping relation between the source area distribution and the target area distribution by using a mapping method between disaster distributions; s5, predicting disaster risk dynamic distribution in the target area through the transmission mapping in the step S4; and S6, dynamically updating the disaster risk distribution model according to the actual observation data of the target area, and generating a disaster risk map. Through the source-target area distribution mapping based on the optimal transmission theory, the technical effect of accurately predicting the disaster risk distribution of the target area under the condition of data scarcity is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster risk assessment, and in particular to a geological disaster risk assessment method and device based on remote sensing big data. Background Art

[0002] As a type of natural disaster that seriously threatens the safety of human life and property, geological disasters are characterized by complex spatial distribution, diverse triggering factors, and significant dynamic changes. With the rapid development of remote sensing technology, a large amount of remote sensing big data (such as SAR images, optical images, DEM, etc.) provides an important data basis for the assessment and prediction of geological disaster risks. However, the existing geological disaster risk assessment methods still have many shortcomings, such as insufficient risk prediction accuracy in data-scarce areas, limited ability to capture dynamic changes, and the inability to effectively migrate differences in disaster distribution between regions, which makes it difficult to meet the actual needs of dynamic disaster assessment and early warning.

[0003] In the existing technology, most risk assessment methods are based on single-region static modeling, usually through simple interpolation, regression analysis or a single statistical method to model disaster distribution, resulting in inaccurate assessment results in data-scarce areas. In addition, traditional methods cannot fully integrate the multi-source characteristics of remote sensing big data, and lack the ability to comprehensively model dynamic changes over time. Especially in the absence of historical data in the target area, the deviation between the assessment results and the actual disaster risk distribution is large. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a geological hazard risk assessment method and device based on remote sensing big data, which solves the technical problems in the existing technology that the geological hazard risk distribution is difficult to accurately model, the dynamic prediction ability is insufficient, and the cross-regional risk distribution cannot be effectively migrated under the condition of data scarcity.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a geological disaster risk assessment method based on remote sensing big data, comprising the following steps:

[0006] S1. Collect multi-source data, including remote sensing data, meteorological data and ground monitoring data, and extract disaster triggering factors related to geological hazards;

[0007] S2. Establishing a disaster risk distribution model in the source area, wherein the distribution model is based on the characteristic weights of disaster triggering factors and characterizes the distribution of disaster risks in spatial location and time dimensions;

[0008] S3. Construct a disaster risk distribution model for the target area based on limited observation data in the target area;

[0009] S4, using a mapping method between disaster distributions to calculate a transmission mapping relationship between source area distribution and target area distribution;

[0010] S5, predicting the dynamic distribution of disaster risks in the target area through the transmission mapping in step S4;

[0011] S6. Dynamically update the disaster risk distribution model based on the actual observation data of the target area and generate a disaster risk map.

[0012] Preferably, the remote sensing data in step S1 includes optical image data, SAR / InSAR data and LiDAR data, the meteorological data includes rainfall, evaporation and temperature data, and the ground monitoring data includes GNSS data, soil moisture data and historical disaster records.

[0013] Preferably, the disaster risk distribution model of the source area in step S2 is a risk distribution model based on probability density, and the risk distribution is obtained by normalizing after weighted fitting of the multidimensional characteristics of disaster triggering factors.

[0014] Preferably, in the step S3, the disaster risk distribution model of the target area is constructed by using limited sparse sampling points, and the distribution weights of the sparse samples are generated using the kernel density estimation method, and the distribution of the target area is fitted.

[0015] Preferably, the transmission mapping relationship between disaster distributions in step S4 is solved based on the optimal transmission theory, and the transmission mapping is obtained by minimizing the transmission cost between disaster distributions and satisfies the following conditions: the disaster risk distribution in the source area is transmitted to the disaster risk distribution in the target area through the mapping.

[0016] Preferably, the transmission cost is obtained by calculating the Wasserstein distance between the disaster risk distribution in the source area and the disaster risk distribution in the target area, and constructing the transmission path on the time series by a dynamic optimal transmission method.

[0017] Preferably, the prediction of dynamic distribution of disaster risk in step S5 comprises the following steps:

[0018] S51, generating a dynamic risk distribution of a target area based on the transmission map and the disaster risk distribution model of the source area;

[0019] S52. Use multi-temporal input data combined with time series analysis to predict the dynamic evolution of disaster risk over time;

[0020] S53. Generate a disaster risk map of the target area to identify areas with different risk levels.

[0021] Preferably, the dynamic update of the disaster risk distribution model in step S6 includes the following process:

[0022] A. Based on the newly added observation data of the target area, the dynamic path of the distribution mapping is optimized to make the updated disaster risk distribution closer to the actual distribution;

[0023] B. Use time series data to constrain dynamic transmission paths so that the distribution evolution satisfies the quality conservation conditions and the transmission cost minimization objectives.

[0024] Preferably, the training process of the disaster risk assessment model in step S6 adopts the following loss function:

[0025] Data fitting loss, which is used to minimize the error between the model prediction results and the observed data;

[0026] Distribution loss, which is used to minimize the Wasserstein distance between the source area disaster risk distribution and the target area disaster risk distribution;

[0027] By dynamically adjusting the loss function weight, the prediction accuracy and distribution matching degree of the model are balanced.

[0028] A geological disaster risk assessment device based on remote sensing big data, comprising:

[0029] The data processing module is used to collect and pre-process remote sensing data, meteorological data and ground monitoring data, extract trigger factors related to disaster risks, and generate multi-dimensional feature data of trigger factors;

[0030] Distribution modeling module, used to construct source area disaster risk distribution model and target area disaster risk distribution model;

[0031] The transmission mapping module is used to calculate the mapping relationship between the disaster distribution in the source area and the target area and generate the transmission path;

[0032] a risk prediction module, for generating a dynamic disaster risk distribution of a target area based on the transmission map;

[0033] An optimized update module is used to dynamically update the disaster risk distribution model based on newly added observation data in the target area;

[0034] The display module is used to generate a disaster risk map of the target area and display the risk levels and dynamic change trends of different areas.

[0035] The present invention provides a method and device for assessing geological disaster risks based on remote sensing big data. It has the following beneficial effects:

[0036] 1. The source-target area distribution mapping based on the optimal transmission theory of the present invention achieves the technical effect of accurately predicting the disaster risk distribution in the target area under the condition of data scarcity. Compared with the technical solutions in the prior art that rely on simple interpolation or regression analysis for risk prediction, it solves the shortcomings of being unable to effectively deal with cross-regional distribution differences and low mapping accuracy.

[0037] 2. The present invention constructs a disaster risk distribution model with spatiotemporal dynamic characteristics through dynamic integration and processing of multi-source data, combined with kernel density estimation and time series modeling, achieving the technical effect of high-resolution and dynamic prediction.

[0038] 3. The present invention uses real-time observation data to dynamically correct the disaster risk distribution, achieving the technical effect of real-time adjustment of the risk prediction model and improving the model's robustness and accuracy. Compared with the technical solutions of static modeling or single prediction in the prior art, it solves the technical problems of its difficulty in adapting to environmental changes and poor model adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the process steps of the present invention;

[0040] Figure 2 This is a schematic flow chart of step S5 of the present invention;

[0041] Figure 3 It is a schematic diagram of the device architecture of the present invention. DETAILED DESCRIPTION

[0042] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Please refer to the attached Figure 1 The present invention provides a method for assessing geological disaster risk based on remote sensing big data, comprising the following steps:

[0044] S1. Collect multi-source data, including remote sensing data, meteorological data and ground monitoring data, and extract disaster triggering factors related to geological disasters; the remote sensing data in step S1 includes optical imaging data, SAR / InSAR data and LiDAR data, the meteorological data includes rainfall, evaporation and temperature data, and the ground monitoring data includes GNSS data, soil moisture data and historical disaster records.

[0045] Specifically, first, step S1 of the present invention is used to obtain trigger factor data related to geological disasters from multi-source data to provide basic data support for subsequent risk assessment distribution modeling. The results of step S1 will be directly used as input data for the construction of the disaster risk distribution model.

[0046] In this embodiment, the multi-source data mainly includes remote sensing data, meteorological data and ground monitoring data. By integrating and processing these data, disaster triggering factors can be extracted, such as terrain characteristics, hydrological characteristics, surface deformation characteristics and vegetation coverage characteristics. Triggering factors are important parameters that affect the occurrence of geological disasters. Reasonable extraction and processing of these factors is crucial for subsequent distribution modeling.

[0047] In general, remote sensing data may include optical image data (such as Landsat images), SAR / InSAR data, and LiDAR data. Optical image data provides surface reflectivity information and can be used to extract terrain and vegetation features; SAR / InSAR data is used to analyze surface deformation features; LiDAR data generates a surface elevation model through high-precision laser ranging, which is used to extract detailed terrain features such as slope and aspect. As an option, high-resolution remote sensing images (such as Sentinel-2 data with a resolution better than 10 meters) can be used to improve the accuracy of feature extraction.

[0048] In a possible implementation, the meteorological data includes rainfall, evaporation, temperature and wind speed data. These data are usually obtained through weather station monitoring or from global meteorological models (such as ERA5 reanalysis data). Specifically, the accumulated rainfall It is an important hydrological factor that determines the occurrence of geological disasters. Its calculation formula is:

[0049]

[0050] in, Indicates time The accumulated rainfall in is the cumulative time window, Indicates time The instantaneous rainfall intensity. Cumulative time window The choice can be set according to regional characteristics, usually 24 hours to 72 hours.

[0051] The ground monitoring data is usually obtained by monitoring the ground deformation information through GNSS stations, or by obtaining soil moisture and groundwater level changes through sensors. It can be expressed as:

[0052]

[0053] in, Indicates time In Location The deformation speed, is the change in ground height, is the time interval. The height change data can be calculated using InSAR technology.

[0054] In a possible embodiment, the vegetation coverage feature is extracted by calculating the National Difference Vegetation Index (NDVI). The calculation formula of NDVI is as follows:

[0055]

[0056] in, represents the reflectivity in the near-infrared band, Represents the reflectance of the red band. NDVI is used to assess the density and health of vegetation. Its value range is between [-1,1]. The larger the value, the denser the vegetation coverage.

[0057] In data preprocessing, data from different sources are first time-aligned and spatially resampled. In general, time alignment is based on hours or days, and the spatial resolution is based on the highest resolution data. For example, for remote sensing images, they are uniformly resampled to a resolution of 10 meters. In some embodiments, a multi-temporal data fusion method can be used to take the temporal characteristics of the data as input variables to better capture the dynamic evolution trend of disaster triggering factors.

[0058] As an option, to reduce data noise, a variety of denoising algorithms can be used, such as Gaussian filtering, bilateral filtering or time series filtering. In specific implementations, the coherent speckle noise of SAR data can be removed by an adaptive filtering algorithm, while the outliers of meteorological data can be corrected by a time series interpolation algorithm.

[0059] S2. Establish a disaster risk distribution model in the source area. The distribution model characterizes the distribution of disaster risks in spatial location and time dimensions based on the characteristic weights of disaster triggering factors. The disaster risk distribution model in the source area in step S2 is a risk distribution model based on probability density. The risk distribution is obtained by normalizing the multi-dimensional characteristics of disaster triggering factors after weighted fitting.

[0060] Specifically, the construction of the source area disaster risk distribution model is based on the trigger factor data extracted in step S1, with the aim of establishing a mathematical model that can accurately reflect the distribution characteristics of disaster risks in time and space. By modeling and analyzing the historical data of the source area, a reliable risk distribution reference can be provided for subsequent steps, and cross-regional distribution migration and dynamic prediction can be supported. The implementation of this step is highly related to the trigger factor data in the aforementioned step S1, and its processing results are directly used as input for the construction of the distribution model, while providing a mapping basis for the target area modeling in step S3.

[0061] In this embodiment, the source area disaster risk distribution model describes the spatial location and time distribution characteristics of disaster risk by constructing a probability density distribution. In general, the risk distribution model is expressed as:

[0062]

[0063] in, Indicates the spatial location of the source region and time The risk distribution density function is represents the spatial extent of the source region, It is a disaster risk function, which is used to comprehensively describe the contribution of triggering factors to disaster risk.

[0064] In one possible implementation, the disaster risk function The definition of is based on the trigger factor data extracted in step S1, which is specifically expressed as:

[0065]

[0066] in, Indicates The trigger factor is located at and time The value of For the The weight coefficient of each trigger factor indicates its impact on disaster risk. is a random error term that describes the impact of unmodeled factors.

[0067] In general, the weight of the trigger factor It can be determined by regression analysis or machine learning algorithms. In some embodiments, the least squares method is used to fit , in order to minimize the error between the model prediction value and the observed risk data. Specifically, the objective function of the least squares method is:

[0068]

[0069] in, Indicates that the source area is at the sampling point The actual disaster risk value, is the number of sampling points.

[0070] As an option, weight The determination of can also be optimized by regularization methods, such as introducing L2 regularization terms to prevent overfitting. The objective function after regularization is expressed as:

[0071]

[0072] in, is the regularization coefficient, which is used to balance the fitting error and model complexity.

[0073] In some embodiments, in order to enhance the nonlinear expression ability of the model, the trigger factor data can be processed by introducing nonlinear transformation. Specifically, the risk function Can be expanded to:

[0074]

[0075] in, is a nonlinear transformation function, such as a logarithmic function, an exponential function, or a polynomial function.

[0076] In data preprocessing, all trigger factor data need to be standardized to eliminate the impact of dimensional differences on the model. In general, standardization methods include maximum and minimum value normalization or Z-score normalization. In a possible implementation, the calculation formula for maximum and minimum value normalization is:

[0077]

[0078] in, is the normalized trigger factor value, and Respectively The risk distribution model is validated by comparative analysis of historical disaster data in the source area. In some embodiments, the model performance is evaluated by calculating the mean square error between the distribution model prediction value and the actual observation value.

[0079] Specifically, the risk distribution model constructed in this embodiment has the following characteristics:

[0080] Ability to comprehensively assess disaster risks based on multiple triggering factors;

[0081] Describing risk distribution in the form of probability density functions provides a solid mathematical foundation for cross-regional mapping and dynamic modeling;

[0082] The accuracy and adaptability of the model are enhanced through parameter fitting and verification.

[0083] As an extension, the risk distribution model constructed in this step can be further combined with the time series analysis method to extract the spatiotemporal variation characteristics of disaster risks and provide dynamic input for subsequent steps.

[0084] S3. Constructing a disaster risk distribution model for the target area based on limited observation data in the target area; in the step S3, the disaster risk distribution model for the target area is constructed by limited sparse sampling points, and the distribution weights of the sparse samples are generated using the kernel density estimation method, and the distribution of the target area is fitted.

[0085] Specifically, this step is closely related to the construction of the source area distribution model in the aforementioned step S2, and its results will provide input for the source-target area risk migration in the distribution mapping and directly connect with the mapping calculation in step S4.

[0086] In this step, the data characteristics and distribution modeling of the target area need to fully consider the problem of data scarcity. Generally, the target area lacks complete risk data or observation data. Therefore, it is necessary to approximate the distribution modeling through limited sampling points while maintaining the smoothness and accuracy of the distribution.

[0087] In this embodiment, the disaster risk distribution model of the target area is established by the probability density estimation method of sparse sampling points. Specifically, the target area distribution is expressed as:

[0088]

[0089] in, Indicates that the target area is at position and time The risk distribution density is is the number of sampling points, represents the weight of the sampling point, is the Dirac function, which is used to represent the sampling point position of the target area.

[0090] In general, the weight of the sampling point It is calculated by the kernel density estimation method. Kernel density estimation is done by defining a kernel function The sampling points are weighted, and the choice of kernel function directly affects the smoothness of the distribution. In one possible implementation, the distribution density of the target area is calculated by the following formula:

[0091]

[0092] in, is the kernel bandwidth, which indicates the scale of the kernel function, is the spatial dimension, It is a kernel function. Commonly used kernel functions include Gaussian kernel, Epanechnikov kernel or uniform kernel.

[0093] Specifically, the Gaussian kernel function is defined as:

[0094]

[0095] in, , represents the standardized input variable, is the Euclidean distance.

[0096] As an option, the core bandwidth The selection of can be determined by a cross-validation method to minimize the estimation error. In some embodiments, the bandwidth The calculation formula is:

[0097]

[0098] in, is the standard deviation of the sampling points, is the number of sampling points, For dimension.

[0099] In a possible implementation, the distribution density of the target area must also meet the boundary convergence condition. Specifically, the integral of the distribution density should meet the following constraints:

[0100]

[0101] in, is the spatial extent of the target area.

[0102] Generally, the data sources of sparse sampling points include weather stations, known high-risk areas in remote sensing images, or historical disaster data. In some embodiments, to improve the representativeness of sparse sampling points, a segmentation algorithm of remote sensing images can be combined to automatically select high-risk areas as sampling points.

[0103] In this embodiment, in order to further improve the accuracy of the target area distribution, the sparse sampling points can be distributed and expanded by using an interpolation algorithm. As an option, the interpolation algorithm includes an inverse distance weighted interpolation method (IDW), a Kriging interpolation method, or a piecewise linear interpolation method.

[0104] In data preprocessing, the sampling point data of the target area needs to be standardized to ensure the consistency of the distribution characteristics with the source area. Generally, the standardization process is the same as the method used in step S2, including maximum and minimum value normalization or Z-score normalization.

[0105] In some embodiments, to ensure the dynamic distribution of the target area, the sampling point data can be expanded into multi-phase features in combination with the time series analysis method. Specifically, the sliding window method can be used to calculate the change rate of the time series to supplement the time dynamic features of the target area.

[0106] In summary, the construction of the disaster risk distribution model in the target area is achieved through probability density estimation and interpolation methods of sparse sampling points, which provides key input for subsequent distribution mapping and dynamic prediction.

[0107] S4. Calculate the transmission mapping relationship between the source area distribution and the target area distribution by using the mapping method between disaster distributions; the transmission mapping relationship between disaster distributions in step S4 is solved based on the optimal transmission theory, and the transmission mapping is obtained by minimizing the transmission cost between disaster distributions, and meets the following conditions: the disaster risk distribution of the source area is transmitted to the disaster risk distribution of the target area through the mapping. The transmission cost is obtained by calculating the Wasserstein distance between the disaster risk distribution of the source area and the disaster risk distribution of the target area, and the transmission path on the time series is constructed by the dynamic optimal transmission method.

[0108] Specifically, the core of step S4 is to use the source area disaster risk distribution model constructed in the above step S2 and the target area disaster risk distribution model constructed in step S3 to establish a mapping relationship between the two through the optimal transmission theory. The result of the distribution mapping is used to migrate the distribution information of the source area to the target area to achieve dynamic prediction of disaster risks in the target area. The implementation of this step requires combining the risk distribution characteristics of the source and target areas, and at the same time minimizing the transmission cost between distributions through the optimization of the mapping path, so as to provide input for subsequent dynamic prediction of disaster risks.

[0109] In this embodiment, the distribution mapping based on the optimal transmission theory is implemented through the following process. In general, the source area distribution model is expressed as ,The target area distribution model is expressed as The goal of distribution mapping is to find the distribution from the source region Distribution to target area The optimal transmission path is determined to minimize the transmission cost between the two distributions.

[0110] In one possible implementation, the transmission cost is measured by the Wasserstein distance (also called the optimal transmission distance). The Wasserstein distance is defined as:

[0111]

[0112] in, is the order of distance, Indicates from position To location The transmission cost, represents the set of all joint distributions that satisfy the following conditions:

[0113]

[0114] Right now The marginal distributions of are the source distributions and target distribution

[0115] As an option, the Wasserstein-2 distance is used in this embodiment ( ) as a transmission distance metric. Specifically, the Wasserstein-2 distance has better geometric interpretation and can directly reflect the transmission cost between risk distributions.

[0116] In order to solve the Wasserstein distance, the optimal transmission mapping is used in this embodiment. , whose goal is to map the distribution of the source region Transformed into the distribution of the target area, satisfying:

[0117]

[0118] in, express The inferred distribution under , which is defined as:

[0119]

[0120] in, Represents the mapping function The Jacobian of . Mapping function The optimization goal

[0121] is to minimize the following transmission costs:

[0122]

[0123] In general, mapping The solution can be achieved through an iterative algorithm. In one possible implementation, the Sinkhorn regularization method can be used to regularize the Wasserstein distance to improve computational efficiency. The optimization goal after regularization is:

[0124]

[0125] in, is the regularization parameter, is the Kullback-Leibler divergence, which is used to constrain the smoothness of the joint distribution.

[0126] Specifically, this embodiment combines the dynamic optimal transmission theory and further expands the mapping process to adapt to the distribution of dynamic changes in time. The goal of dynamic optimal transmission is to build a continuous transmission path in the time dimension. and velocity field , so that:

[0127]

[0128] At the same time, the following mass conservation constraints are satisfied:

[0129]

[0130] in, Indicates time The distribution density of Indicates time In some embodiments, terrain constraints and physical factors can be introduced to enhance the robustness of the mapping process. The physical rationality of the path can be further optimized by combining the terrain slope, river network and other characteristics of the target area. Specifically, the following terrain constraints can be added to the objective function:

[0131]

[0132] in, is the terrain slope function, is the elevation distribution of the target area, and is a weight parameter used to adjust the impact of the constraint. In the post-processing of distribution mapping, it is necessary to verify whether the mapping results conform to the risk characteristics of the target area.

[0133] Generally, the Wasserstein distance between the mapped target distribution and the actual target distribution can be calculated for verification. If the mapping result does not conform to the actual situation, it can be recalculated by adding regularization constraints or adjusting the source distribution model.

[0134] Please refer to the attached Figure 2 S5, predicting the dynamic distribution of disaster risk in the target area through the transmission mapping in step S4; the prediction of the dynamic distribution of disaster risk in step S5 includes the following steps:

[0135] S51, generating a dynamic risk distribution of a target area based on the transmission map and the disaster risk distribution model of the source area;

[0136] S52. Use multi-temporal input data combined with time series analysis to predict the dynamic evolution of disaster risk over time;

[0137] S53. Generate a disaster risk map of the target area to identify areas with different risk levels.

[0138] Specifically, this step aims to provide key data support for disaster assessment and early warning by predicting the disaster risk distribution in the target area at different time points. The implementation of this step relies on the basic distribution information generated by the source-target area mapping, and further combines the time dynamic characteristics to predict and visualize the risk field. The obtained dynamic risk distribution results also provide input for subsequent dynamic optimization.

[0139] In this embodiment, the dynamic prediction of disaster risk in the target area is based on the distribution evolution in the time series. In general, the predicted dynamic distribution field is expressed as:

[0140]

[0141] in, Indicates the spatial location of the target area and time The distribution of disaster risk, is the disaster risk distribution in the source area, is the optimal transmission map generated in step S4. As an option, in order to capture the dynamic change characteristics of disaster risk in the target area over time, this embodiment combines the time series analysis method. Specifically, the dynamic distribution prediction of the target area can be modeled as:

[0142]

[0143] in, represents the basis functions on the space, represents the basis function in time, and are model parameters, is the number of basis functions selected.

[0144] In order to enhance the nonlinear prediction capability of the model, this embodiment further introduces a deep learning model to fit and predict the dynamic distribution. In one possible implementation, a combined model based on a convolutional neural network (CNN) and a long short-term memory network (LSTM) can be used to perform spatiotemporal joint modeling of the dynamic distribution of the target area. CNN is used to extract spatial distribution features, and LSTM is used to capture time series characteristics. The specific structure is as follows:

[0145] The input layer receives the distribution map generated in step S4

[0146] The convolutional layer extracts the spatial characteristics of the risk distribution in the target area;

[0147] The LSTM layer performs time series modeling for multi-phase distribution;

[0148] The output layer generates the dynamic distribution of the target area at future time points.

[0149] Alternatively, the loss function of a deep learning model can be defined as a weighted sum of distribution similarity and data fitting error:

[0150]

[0151] in, represents the mean square error between the model prediction results and the observed data, represents the Wasserstein distance between the source region distribution and the target region distribution, is the weight coefficient. In the post-processing of dynamic distribution prediction, this embodiment visualizes the prediction results to generate a disaster risk map of the target area. In general, the risk map is obtained by comparing the predicted distribution density values. Generate threshold values.

[0152] For example, you can The values ​​are divided into three levels: high risk, medium risk and low risk, and marked with different colors. The generation of risk maps can be combined with terrain data to make high-risk areas more intuitive on the map.

[0153] S6. Dynamically update the disaster risk distribution model according to the actual observation data of the target area and generate a disaster risk map. The dynamic update of the disaster risk distribution model in step S6 includes the following process:

[0154] A. Based on the newly added observation data of the target area, the dynamic path of the distribution mapping is optimized to make the updated disaster risk distribution closer to the actual distribution;

[0155] B. Use time series data to constrain dynamic transmission paths so that the distribution evolution satisfies the quality conservation conditions and the transmission cost minimization objectives.

[0156] The training process of the disaster risk assessment model in step S6 adopts the following loss function:

[0157] Data fitting loss, which is used to minimize the error between the model prediction results and the observed data;

[0158] Distribution loss, which is used to minimize the Wasserstein distance between the source area disaster risk distribution and the target area disaster risk distribution;

[0159] By dynamically adjusting the loss function weight, the prediction accuracy and distribution matching degree of the model are balanced.

[0160] Specifically, this step is closely related to step S5. The optimization process is directly based on the dynamic prediction results. At the same time, it combines the observation data and time series information collected in real time in the target area to dynamically adjust the model parameters and distribution characteristics to provide support for long-term monitoring and real-time warning of disaster risks.

[0161] In this embodiment, the dynamic optimization of disaster risk distribution is centered on time series updating. With the predicted distribution Comparative analysis, gradually adjust the distribution parameters , making the predicted distribution closer to the actual observation results. In general, the optimization goal is:

[0162]

[0163] in, represents the distribution model after dynamic optimization, is the observed data, are the model parameters to be optimized, and are the spatial and temporal extents of the target region, respectively. As an option, the model parameters The optimization can be done using the gradient descent method. In each optimization iteration, the gradient of the loss function to the parameters is calculated:

[0164]

[0165] in, is the learning rate, is the gradient of the loss function with respect to the parameters, Indicates the number of iterations. Specifically, in order to improve the efficiency and robustness of the optimization process, this embodiment also introduces constraints for dynamic optimal transmission. In general, the distribution update needs to satisfy the following mass conservation equation:

[0166]

[0167] in, For time The distribution density of is the corresponding transmission velocity field. During the optimization process, the smoothness of the transmission path is constrained by the following regularization term:

[0168]

[0169] This regularization term prevents excessive oscillations in the distribution update.

[0170] In post-processing, the optimized risk distribution is verified for accuracy and reliability by comparing it with actual disaster occurrence data. Generally, the verification process uses a combination of mean square error and distribution similarity metrics to ensure that the optimized distribution is consistent with the actual observed trend.

[0171] Please refer to the attached Figure 3 , a geological disaster risk assessment device based on remote sensing big data, comprising:

[0172] The data processing module is the input module of the device, which is used to receive multi-source data (including remote sensing data, meteorological data and ground monitoring data). The data processing module cleans, denoises, aligns and standardizes the raw data, and extracts the disaster triggering factors into a multi-dimensional feature matrix in a unified format, which is provided to the distribution modeling module. This module is connected to the distribution modeling module through an internal data interface to provide high-quality input data.

[0173] The distribution modeling module is used to construct the disaster risk distribution models of the source area and the target area respectively. The distribution modeling module receives the multi-dimensional feature data output by the data processing module, and constructs the high-resolution risk distribution of the source area and the sparse risk distribution model of the target area based on the probability density model and the kernel density estimation algorithm. The distribution modeling module is connected to the transmission mapping module, and the modeling results are used as input to calculate the mapping relationship between the two distributions.

[0174] The core task of the transmission mapping module is to establish a distribution mapping from the source area to the target area based on the optimal transmission theory. This module calculates the Wasserstein distance, generates the optimal transmission path and transmission mapping function, and maps the risk distribution of the source area to the target area. The transmission mapping module is connected to the risk prediction module, and the mapping results are used as input to provide a basis for the dynamic risk distribution prediction of the target area.

[0175] The risk prediction module generates a dynamic risk distribution of the target area at a future time point by combining the transmission mapping results and the time series characteristics of the target area. The module supports dual implementation methods based on physical modeling (such as time series analysis) and deep learning (such as spatiotemporal networks), and can capture the spatial and temporal evolution trends of disaster risks in the target area. The output of the risk prediction module is passed to the optimization update module through an interface for correction and optimization.

[0176] The optimization and update module dynamically corrects the risk prediction results. This module compares the predicted distribution with real-time observation data, and dynamically adjusts the parameters of the distribution model through gradient optimization, regularization constraints, and Bayesian updating methods to keep the prediction results consistent with the actual observation data. The optimization and update module forms a closed loop with the risk prediction module to ensure that the distribution model has dynamic adaptability when the disaster triggering conditions change.

[0177] The display module is the output module of the device, which is used to intuitively display the optimized disaster risk distribution as a risk map. The display module visualizes information such as high-risk areas and dynamic change trends through spatial and temporal classification of risk distribution.

[0178] First, the data processing module extracts multidimensional feature data and passes it to the distribution modeling module to generate the risk distribution of the source-target area. The transmission mapping module uses these distributions to establish a mapping relationship between distributions through optimal transmission theory, and migrates the distribution information of the source area to the target area. Based on the mapping results, the risk prediction module generates the dynamic risk distribution of the target area at a future time point, and passes the prediction results to the optimization update module for real-time adjustment and correction. Finally, the display module receives the optimized distribution data and generates a disaster risk map for user reference.

[0179] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A geological disaster risk assessment method based on remote sensing big data, characterized in that: The following steps are involved: S1. Collect multi-source data, including remote sensing data, meteorological data and ground monitoring data, and extract disaster triggering factors related to geological hazards; S2. Establishing a disaster risk distribution model in the source area, wherein the distribution model is based on the characteristic weights of disaster triggering factors and characterizes the distribution of disaster risks in spatial location and time dimensions; S3. Construct a disaster risk distribution model for the target area based on limited observation data in the target area; S4, using a mapping method between disaster distributions to calculate a transmission mapping relationship between source area distribution and target area distribution; S5, predicting the dynamic distribution of disaster risks in the target area through the transmission mapping in step S4; S6. Dynamically update the disaster risk distribution model based on the actual observation data of the target area and generate a disaster risk map.

2. The method for assessing geological hazards based on remote sensing big data according to claim 1, characterized in that: The remote sensing data in step S1 includes optical image data, SAR / InSAR data and LiDAR data, the meteorological data includes rainfall, evaporation and temperature data, and the ground monitoring data includes GNSS data, soil moisture data and historical disaster records.

3. The method for assessing geological disaster risk based on remote sensing big data according to claim 1, characterized in that: The disaster risk distribution model of the source area in step S2 is a risk distribution model based on probability density, and the risk distribution is obtained by normalizing after weighted fitting of the multidimensional characteristics of disaster triggering factors.

4. The method for assessing geological disaster risk based on remote sensing big data according to claim 1, characterized in that: In the step S3, the disaster risk distribution model of the target area is constructed by limited sparse sampling points, the distribution weights of the sparse samples are generated by the kernel density estimation method, and the distribution of the target area is fitted.

5. The method for assessing geological hazards based on remote sensing big data according to claim 1, characterized in that: The transmission mapping relationship between disaster distributions in step S4 is solved based on the optimal transmission theory. The transmission mapping is obtained by minimizing the transmission cost between disaster distributions and meets the following conditions: the disaster risk distribution in the source area is transmitted to the disaster risk distribution in the target area through the mapping.

6. A geological disaster risk assessment method based on remote sensing big data according to claim 5, characterized in that: The transmission cost is obtained by calculating the Wasserstein distance between the disaster risk distribution of the source area and the disaster risk distribution of the target area, and the transmission path on the time series is constructed by a dynamic optimal transmission method.

7. The method for assessing geological disaster risk based on remote sensing big data according to claim 1, characterized in that: The prediction of dynamic distribution of disaster risk in step S5 includes the following steps: S51, generating a dynamic risk distribution of a target area based on the transmission map and the disaster risk distribution model of the source area; S52. Use multi-temporal input data combined with time series analysis to predict the dynamic evolution of disaster risk over time; S53. Generate a disaster risk map of the target area to identify areas with different risk levels.

8. The method for assessing geological disaster risk based on remote sensing big data according to claim 1, characterized in that: The dynamic update of the disaster risk distribution model in step S6 includes the following process: A. Based on the newly added observation data of the target area, the dynamic path of the distribution mapping is optimized to make the updated disaster risk distribution closer to the actual distribution; B. Use time series data to constrain dynamic transmission paths so that the distribution evolution satisfies the quality conservation conditions and the transmission cost minimization objectives.

9. The method for assessing geological disaster risk based on remote sensing big data according to claim 1, characterized in that: The training process of the disaster risk assessment model in step S6 adopts the following loss function: Data fitting loss, which is used to minimize the error between the model prediction results and the observed data; Distribution loss, which is used to minimize the Wasserstein distance between the source area disaster risk distribution and the target area disaster risk distribution; By dynamically adjusting the loss function weight, the prediction accuracy and distribution matching degree of the model are balanced.

10. A geological disaster risk assessment device based on remote sensing big data, according to a geological disaster risk assessment method based on remote sensing big data according to any one of claims 1 to 9, characterized in that: include: The data processing module is used to collect and pre-process remote sensing data, meteorological data and ground monitoring data, extract trigger factors related to disaster risks, and generate multi-dimensional feature data of trigger factors; Distribution modeling module, used to construct source area disaster risk distribution model and target area disaster risk distribution model; The transmission mapping module is used to calculate the mapping relationship between the disaster distribution in the source area and the target area and generate the transmission path; a risk prediction module, for generating a dynamic disaster risk distribution of a target area based on the transmission map; An optimized update module is used to dynamically update the disaster risk distribution model based on newly added observation data in the target area; The display module is used to generate a disaster risk map of the target area and display the risk levels and dynamic change trends of different areas.

Images