Quaternary overburden slope landslide monitoring method and device based on AIOT

Abstract

The present invention discloses a Quaternary overburden slope landslide monitoring method and device based on AIOT, and specifically relates to the technical field of landslide monitoring. By acquiring soil shear strength and vegetation coverage data in different areas, and combining existing monitoring data and trained AI models, a comprehensive analysis and assessment of landslide risks are carried out. In the risk assessment process, the AI ​​model is used to analyze the influence of soil and vegetation, and the dependence and characteristic correlation of the landslide occurrence area are analyzed in combination with principal component analysis and fuzzy logic, so as to judge the performance ability of the existing AI model in different areas. According to the evaluation results, the area is divided into high, medium and low precision prediction areas, and the transfer learning method is applied to the low precision area to improve the prediction accuracy, so as to realize accurate early warning and prediction of landslide risks in different areas, effectively deal with the differences in landslide characteristics between regions, improve the generalization ability and prediction accuracy of the AI ​​model, and have strong practical value and technical advantages.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide monitoring, and in particular to a method and device for monitoring landslides in a Quaternary overburden slope based on AIOT. Background Art

[0002] Quaternary overburden slope landslide monitoring based on AIOT (artificial intelligence Internet of Things) refers to the use of artificial intelligence technology combined with Internet of Things devices to monitor possible landslides on the Quaternary overburden (i.e., slopes covered with loose sediments composed of mud, gravel, etc.) in real time. By installing sensors, monitoring equipment and other Internet of Things technologies, data such as slope displacement, vibration, meteorology, and soil moisture content are collected, and landslide risks are analyzed and predicted through artificial intelligence algorithms. This technology can achieve early warning and dynamic monitoring of landslides, which helps to reduce losses caused by natural disasters. In the existing technology, the slope monitoring system based on AIOT technology transmits real-time data to the central control system through wireless networks by arranging various sensors on the slope, such as soil sensors, crack monitors, temperature and humidity sensors, etc. The AI ​​algorithm analyzes these data, identifies potential landslide risks, and issues early warning signals. For example, in the construction of roads in some mountainous areas, combined with meteorological data and slope monitoring data, artificial intelligence is used to conduct risk assessments on areas where landslides may occur, so as to achieve early prediction and prevention of disasters.

[0003] The prior art has the following deficiencies: In the prior art, AI models are trained through existing landslide events and historical monitoring data to enable them to identify the characteristics and patterns of landslides. However, the causes and manifestations of landslides in different regions may vary greatly. For example, landslides in some areas may be more caused by heavy rainfall, while other areas may be caused by earthquakes or human activities (such as mining, road construction, etc.). The landslide occurrence patterns in different regions are different, which may result in the model being unable to generalize from historical data in a certain area to other areas. AI models may assume that landslide characteristics are universal during training, but the actual situation may be significantly different. In addition, due to the significant differences in slope characteristics in different regions, the causes and mechanisms of landslides may be completely different in different places. This can make predictions based on general models inaccurate, resulting in a significant decrease in accuracy, especially when data comes from multiple geographical regions, and the model has difficulty capturing differences between regions. Summary of the invention

[0004] The purpose of the present invention is to provide a method and device for monitoring landslides in a Quaternary overburden slope based on AIOT, so as to solve the shortcomings of the background technology.

[0005] In order to achieve the above object, the present invention provides the following technical solution: a method for monitoring landslides in a Quaternary overburden layer based on AIOT, comprising the following steps: Obtain soil shear strength data and vegetation coverage data in different areas through remote sensing images and soil databases; Based on the soil shear strength data and vegetation coverage data, a comprehensive analysis is conducted in combination with existing monitoring data and trained AI models to assess the landslide risks in different areas. In risk assessment, AI models are used to analyze the dependence of soil shear strength and vegetation coverage on landslide occurrence areas, and principal component analysis is used to analyze the correlation between regional characteristics and landslide occurrence. Fuzzy logic is then used to determine the performance of existing AI models in different regions. According to the judgment results, different regions are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas. For low-precision prediction areas, the transfer learning method is used to enable the existing AI model to migrate between different regions to improve the prediction accuracy of the existing AI model.

[0006] Preferably, the existing monitoring data includes meteorological data, human activity data, geological data and historical landslide data.

[0007] Preferably, for each target area, the AI ​​model predicts the probability of landslide occurrence in the area based on the input soil shear strength, vegetation coverage and environmental factors, and divides the risks into high-risk areas, medium-risk areas and low-risk areas based on the predicted probability; through the prediction results of the model, the landslide risks in different areas are spatially analyzed to generate a landslide risk heat map. The heat map represents the degree of landslide risk in each area through gradients of different colors to identify high-risk areas.

[0008] Preferably, the dependence of soil shear strength and vegetation coverage on the landslide occurrence area is analyzed to generate a landslide dependence comprehensive index, and the method for obtaining the landslide dependence comprehensive index is: A Bayesian network model representing landslide risk is constructed. Each node in the model represents a factor, and the connection between nodes is determined according to the dependency between factors. For each node, a conditional probability table is defined to represent the probability distribution of the node given its parent node. The input data includes the actual observed values ​​of each factor. The structure of the Bayesian network and the conditional probability table are used to derive the posterior probability according to the Bayesian theorem. Based on the known variables input, the Bayesian network is used to update the posterior probability of landslide occurrence as a comprehensive landslide dependency index, which is obtained by calculating the weighted sum of the conditional probability of landslide occurrence based on soil shear strength, vegetation coverage, slope and precipitation.

[0009] Preferably, the correlation between regional characteristics and landslide occurrence is analyzed by principal component analysis to generate a landslide characteristic correlation index, and the method for obtaining the landslide characteristic correlation index is: Collect observations of multiple features in the region and represent them as , where n is the number of features, and the observations of each feature in different regions are represented as a matrix X, where each column represents a feature and each row represents a region; the standardized data matrix is ​​marked as Z, where the mean of each column is 0 and the standard deviation is 1; the covariance matrix of the standardized data matrix Z is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are calculated. The eigenvalues ​​and eigenvectors are solved by equations, and the first k principal components are selected according to the size of the eigenvalues. The principal components are linear combinations of the original features, and the expression is: ;in is the weight of the nth feature in the kth principal component, and the landslide feature association index is calculated, and the expression is: ;in: is the eigenvalue of the kth principal component, is the kth principal component, is the landslide characteristic correlation index.

[0010] Preferably, the performance of the existing AI model in different areas is judged by using fuzzy logic, including: The landslide dependence comprehensive index and landslide characteristic correlation index are used as the input items of fuzzy logic and divided into different fuzzy sets respectively. The performance of existing AI models in different areas is used as the output of fuzzy logic and divided into different fuzzy sets; Formulate fuzzy rules to describe the impact of landslide dependence composite index and landslide characteristic correlation index on the performance ability of existing AI models in different regions; Through fuzzy reasoning, different areas are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas according to the performance capabilities of existing AI models in different areas.

[0011] Preferably, for low-precision prediction areas, a transfer learning method is used to enable existing AI models to migrate between different areas to improve the prediction accuracy of existing AI models, including: using the training data of the source domain to train a prediction model, using the trained model in the source domain as the preliminary model of the target domain, adjusting the source domain model on the limited training data in the target domain, and adjusting the parameters of the model to adapt it to the characteristics and task requirements of the target domain; during the training process, introducing a migration loss function to minimize the distribution difference between the source domain and the target domain; continuing to train the target model in the target domain, adjusting the hyperparameters, and improving the prediction accuracy by continuously optimizing the loss function of the target domain model; using the test data set of the target domain to evaluate the performance of the migrated model; if the prediction accuracy of the model in the target domain is improved, the transfer learning is considered successful; if the prediction accuracy of the model in the target domain is not improved, it is necessary to further adjust the model parameters or migration strategy.

[0012] The present invention also provides a Quaternary overburden slope landslide monitoring device based on AIOT, including a data acquisition module, a risk assessment module, a fuzzy logic analysis module and a transfer learning module; Data acquisition module: obtain soil shear strength data and vegetation coverage data in different areas through remote sensing images and soil database; Risk assessment module: Based on the acquired soil shear strength data and vegetation coverage data, combined with existing monitoring data and trained AI models, comprehensive analysis is performed to assess the landslide risks in different areas; Fuzzy logic analysis module: In risk assessment, AI models are used to analyze the dependence of soil shear strength and vegetation coverage on landslide occurrence areas, and principal component analysis is used to analyze the correlation between regional characteristics and landslide occurrence. Fuzzy logic is then used to determine the performance of existing AI models in different areas. Transfer learning module: Based on the judgment results, different areas are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas. For low-precision prediction areas, transfer learning methods are used to enable existing AI models to migrate between different areas to improve the prediction accuracy of existing AI models.

[0013] In the above technical solution, the technical effects and advantages provided by the present invention are: The present invention comprehensively analyzes soil shear strength, vegetation coverage, meteorological data and historical landslide data, combines principal component analysis and fuzzy logic judgment, accurately evaluates the landslide risks in different regions, and divides the regions into high, medium and low precision prediction areas. For low-precision prediction areas, the present invention adopts a transfer learning method to adjust the AI ​​model by migrating the knowledge of the source field so that it can adapt to the characteristics of the target area and improve the prediction accuracy. This technical solution combines regional characteristics with the laws of landslide occurrence, so that the model can be effectively migrated between different geographical regions, improves the accuracy and universality of landslide risk prediction, and helps to provide customized landslide disaster prevention and early warning solutions for different regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0015] Figure 1 The present invention is a flow chart of the method.

[0016] Figure 2 It is a module diagram of the device of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0018] Example 1, please refer to Figure 1 As shown, the Quaternary cover layer slope landslide monitoring method based on AIOT described in this embodiment includes: Obtain soil shear strength data and vegetation coverage data in different areas through remote sensing images and soil databases; Based on the soil shear strength data and vegetation coverage data, a comprehensive analysis is conducted in combination with existing monitoring data and trained AI models to assess the landslide risks in different areas. In risk assessment, AI models are used to analyze the dependence of soil shear strength and vegetation coverage on landslide occurrence areas, and principal component analysis is used to analyze the correlation between regional characteristics and landslide occurrence. Fuzzy logic is then used to determine the performance of existing AI models in different regions. According to the judgment results, different regions are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas. For low-precision prediction areas, the transfer learning method is used to enable the existing AI model to migrate between different regions to improve the prediction accuracy of the existing AI model.

[0019] Soil shear strength refers to the ability of soil to resist sliding when subjected to external forces, which is usually determined by the physical properties of the soil (such as density, particle size, moisture, etc.). The process of obtaining this data can be achieved by combining remote sensing image analysis with soil databases to obtain soil shear strength data for a large area.

[0020] Remote sensing images can provide large-scale, real-time and high-precision surface information. Through the analysis of remote sensing images, soil physical characteristics related to soil shear strength can be extracted, including: Soil moisture: Soil moisture is a key factor affecting shear strength. Wet soils tend to have lower shear strength. Remote sensing techniques (such as microwave remote sensing) can accurately estimate the moisture content of surface soil.

[0021] Surface temperature: Remote sensing technology can also provide relevant data on surface temperature. The relationship between soil temperature and shear strength is more complicated, especially in areas with seasonal changes. Soil temperature may affect the expansion and stability of the soil.

[0022] Soil type and composition: Through high-resolution satellite images (such as Landsat images, Sentinel images) and geological data, the type of soil and its particle size can be extracted, which further affects the calculation of shear strength.

[0023] Remote sensing technology implementation method: High-resolution satellite images: Through satellite images such as Landsat and Sentinel, you can obtain information such as soil type, vegetation coverage, and humidity in the area.

[0024] Microwave remote sensing technology: especially suitable for obtaining soil moisture (such as C-band and L-band microwave radiometers). Microwave remote sensing technology is very sensitive to the moisture of the surface soil, especially after precipitation, when the soil moisture changes greatly, thus affecting the soil shear strength.

[0025] Soil databases usually contain a large amount of regional soil information, which can supplement soil physical data that cannot be directly obtained from remote sensing images. For example: USDA Soil Database: The United States Department of Agriculture (USDA) provides detailed soil property data, including soil shear strength, particle composition, clay content, density, and more.

[0026] FAO Soil Database: The Food and Agriculture Organization of the United Nations (FAO) provides global soil information, including soil classification, physical properties, etc., suitable for large-scale soil shear strength assessment.

[0027] In order to improve the accuracy of soil shear strength, the combination of remote sensing images and soil database is particularly critical. The specific methods are as follows: Spatial interpolation: Based on some known soil shear strength data in the soil database, these data can be extrapolated to unmeasured areas through spatial interpolation methods (such as inverse distance weighted method, Kriging interpolation method, etc.).

[0028] Combining remote sensing images with ground data: By comparing and analyzing remote sensing images with ground soil sampling data, a soil shear strength prediction model is constructed. Common methods include machine learning (such as random forests, support vector machines) and regression analysis, and the actual shear strength data of ground sampling points is used to verify and calibrate remote sensing image data.

[0029] Due to the resolution limitation of remote sensing images, it is often necessary to calibrate the extracted results of soil shear strength to ensure its accuracy. The calibration process can be carried out using the following methods: Ground verification: Select ground sampling points in different areas and use standardized shear strength tests (such as direct shear tests) for verification.

[0030] Model optimization: Use machine learning algorithms (such as random forests and neural networks) to train and optimize models based on remote sensing images and actual soil data to ensure that the model can better adapt to the soil characteristics of different regions.

[0031] To conduct landslide risk assessment, it is first necessary to collect relevant data, including soil shear strength, vegetation coverage, meteorological data, geological data, historical landslide data, etc. All kinds of data need to be preprocessed and standardized so that they can be effectively applied in AI models.

[0032] Soil shear strength (such as data obtained through remote sensing images and soil databases) is one of the key factors in landslide risk assessment. This data can reflect the stability and anti-sliding ability of the soil. By obtaining soil shear strength data in different regions, detailed information about surface stability can be provided to the AI ​​model. Data sources: remote sensing images, soil databases (such as USDA, FAO, China Soil Database) and microwave remote sensing data. Data preprocessing: For soil shear strength data, it may be necessary to normalize different soil types in different regions. Fill in the blank data in the region through spatial interpolation technology.

[0033] Vegetation coverage data can be obtained through the Normalized Difference Vegetation Index (NDVI) in remote sensing images. The NDVI value can reflect the growth status and coverage of vegetation, thereby indirectly reflecting soil stability. The root system of vegetation can effectively enhance the anti-sliding ability of soil. Data source: NDVI data provided by remote sensing satellites (such as Sentinel and Landsat). Data preprocessing: According to different seasons and climate conditions, multi-temporal remote sensing images are used to monitor the dynamic changes of vegetation and reduce the error of single time point data.

[0034] Meteorological data, especially precipitation, temperature and humidity, are important factors affecting landslides. Heavy precipitation may lead to soil saturation, which in turn triggers landslides. Real-time or historical precipitation, temperature, humidity and other information can be obtained through weather station data and remote sensing meteorological data (such as meteorological satellite data).

[0035] Geological data include slope, rock and soil type, geological structure, etc. These data are crucial to understand the mechanism of landslide occurrence. Areas with steeper slopes usually have a higher risk of landslides, and different rock and soil layer types and structures may affect the occurrence of landslides.

[0036] Detailed records of historical landslide events (such as time, location, scale, cause, etc.) are an important source of data for training AI models. By analyzing historical data, AI models can learn the occurrence patterns and laws of landslides in different regions.

[0037] The collected data can be used to train AI models to identify and predict landslide risks. Common AI models include machine learning algorithms (such as random forests, support vector machines, decision trees) and deep learning models (such as neural networks, convolutional neural networks, etc.).

[0038] First, the data needs to be labeled as landslide events and non-landslide events. The landslide dataset includes the time, location, scale, and triggering factors (such as heavy rainfall, earthquake, etc.) of the landslide, while the non-landslide dataset includes instances in the corresponding area where no landslide occurred.

[0039] Input features: soil shear strength, vegetation cover, meteorological data, geological data.

[0040] Target variable: whether a landslide occurs (classification task) or not (regression task).

[0041] After data annotation, you can choose to train an AI model for learning. Commonly used AI models include: Random Forest: Suitable for classification and regression problems of high-dimensional and complex data, and can handle nonlinear relationships.

[0042] Support Vector Machine (SVM): It can classify landslides and is particularly suitable for data with complex feature space.

[0043] Convolutional Neural Network (CNN): Applicable to remote sensing image data, it can effectively extract landslide features in images.

[0044] The performance of the model is evaluated through methods such as cross-validation. Evaluation indicators include accuracy, recall, F1-score, AUC value, etc. For different regions, the AI ​​model may have different performance, so model optimization is required: Overfitting and underfitting: Avoid overfitting or underfitting by adjusting model parameters (such as tree depth, learning rate, etc.).

[0045] Feature selection and importance analysis: The most important features for landslide prediction are screened out through feature selection methods (such as L1 regularization, recursive feature elimination, etc.).

[0046] Since the landslide characteristics in different regions vary greatly, the model can be transferred between different regions through transfer learning to enhance the generalization ability of the model. Transfer learning can effectively improve the prediction accuracy in unseen areas by transferring the training knowledge from one field (source area) to another field (target area).

[0047] Once the AI ​​model is trained and optimized to a sufficient level of accuracy, it can be applied to landslide risk assessment in different regions. Landslide risk assessment is performed by comprehensively analyzing the soil shear strength, vegetation coverage, meteorological data and other information obtained, combined with the prediction results of the AI ​​model.

[0048] For each target area, the AI ​​model predicts the probability of landslides in the area based on the input soil shear strength, vegetation coverage and other environmental factors. The output is usually: Landslide probability: the model predicts the possibility of landslides occurring in the area. The value range is usually between 0 and 1. The larger the value, the higher the landslide risk.

[0049] Landslide level: Based on the predicted probability, the risk is divided into several levels, such as high risk, medium risk, and low risk areas.

[0050] The prediction results of the model can be used to analyze the spatial distribution of landslide risks in different regions and generate a landslide risk heat map. The heat map uses different color gradients to indicate the degree of landslide risk in each region, helping decision makers to quickly identify high-risk areas.

[0051] High-risk areas: Areas with a higher probability of landslides, which require detailed monitoring and early warning.

[0052] Low-risk areas: Areas with a low probability of landslides can reduce the monitoring frequency, but regular inspections are still required.

[0053] Since landslide risk changes over time, especially under the influence of external factors such as heavy rainfall and earthquakes, the risk of landslides may increase sharply. Therefore, the AI ​​model should be able to perform real-time monitoring and dynamic evaluation. Through inputs such as real-time meteorological data and soil moisture data, the model can timely update the landslide risk assessment results and provide timely warnings.

[0054] When the model detects that the landslide risk in a certain area exceeds a preset threshold, the system can automatically issue an early warning notification to remind relevant departments to take preventive measures.

[0055] Using AI models to analyze the dependence of soil shear strength and vegetation coverage on landslide occurrence areas can reveal how the two jointly affect landslide risk. Soil shear strength directly reflects the stability and anti-sliding ability of the soil. When the shear strength is low, the soil is prone to sliding, resulting in an increased risk of landslides. The AI ​​model can identify areas with weaker shear strength as high-risk areas by learning the relationship between soil shear strength and landslide occurrence in different regions. At the same time, vegetation coverage consolidates the soil through roots and enhances soil stability. The AI ​​model can identify that in areas with low vegetation coverage, the risk of landslides is generally higher, especially in cases of high precipitation. The lack of vegetation makes the soil easy to erode, leading to landslides.

[0056] The AI ​​model can comprehensively analyze the interaction of these two factors and their impact on the occurrence of landslides, thereby revealing their regional dependence. In some areas, when the vegetation coverage is high, the probability of landslides may still be low even if the soil shear strength is low, because the vegetation effectively enhances the stability of the soil. However, in another type of area, the low value of soil shear strength may not be fully compensated by the increase in vegetation coverage, especially in cases of large slopes or unfavorable meteorological conditions, the risk of landslides is still high. The AI ​​model can quantify the relative importance of soil shear strength and vegetation coverage through regional analysis, providing more accurate predictions for landslide risk assessment.

[0057] The dependence of soil shear strength and vegetation coverage on the landslide occurrence area is analyzed to generate a comprehensive landslide dependence index. The method for obtaining the comprehensive landslide dependence index is as follows: Construct a Bayesian network model to represent landslide risk. Each node in the model represents a factor (such as soil shear strength, vegetation coverage, precipitation, etc.), and the connection between nodes is determined based on the dependency between these factors. Node: Each node represents a variable, such as soil shear strength, vegetation coverage, slope, precipitation, etc. Edge: Directed edges between nodes represent conditional dependencies, for example, soil shear strength may depend on vegetation coverage, and precipitation may affect the probability of landslide occurrence.

[0058] For each node, a conditional probability table (CPT) is defined to represent the probability distribution of the node given its parent node. For example, the CPT of the soil shear strength node can be expressed as: P (Shear Strength | Vegetation Cover); the CPT of the landslide occurrence node is expressed as: P (Landslide | Shear Strength, Vegetation Cover, Rainfall, Slope); Landslide is a landslide; for multivariate models, the conditional probability table is usually estimated through historical data or expert experience.

[0059] The input data includes the actual observed values ​​of various factors (such as soil shear strength, vegetation coverage, etc.). For example: the soil shear strength is 25kPa, the vegetation coverage is 60%, the slope is 35°, and the precipitation is 200mm.

[0060] The structure of the Bayesian network and the conditional probability table are used to derive the posterior probability according to the Bayesian theorem. For landslide risk assessment, the Bayesian network is used to update the posterior probability of landslide occurrence based on the input known variables (such as soil shear strength, vegetation coverage, precipitation, etc.) as a comprehensive index of landslide dependence, that is, the weighted sum of the conditional probability of landslide occurrence by calculating soil shear strength, vegetation coverage, slope and precipitation is obtained.

[0061] The landslide dependency composite index can be used to assess the landslide risk in different areas. For example, if the composite index is close to 1, it means that the risk of landslide is high. If the composite index is close to 0, it means that the risk of landslide is low.

[0062] The correlation between regional characteristics and landslide occurrence is analyzed by principal component analysis to generate the landslide characteristic correlation index. The landslide characteristic correlation index is obtained as follows: Collect observation data of multiple features in the area, including soil shear strength, vegetation cover, slope and precipitation; express them as , where n is the number of features. The observations of each feature in different regions are represented as a matrix X, where each column represents a feature and each row represents a region.

[0063] In order to ensure that all features have consistent dimensions and prevent the large scale of some features from affecting the calculation of principal component analysis, the data needs to be standardized. The standardized data matrix is ​​Z, where the mean of each column is 0 and the standard deviation is 1.

[0064] Calculate the covariance matrix Σ of the standardized data matrix Z. The covariance matrix is ​​used to describe the linear relationship between each feature. The calculation formula is: ;in: It is the transpose of the standardized data matrix Z, m is the number of samples, and the covariance matrix Σ is an n×n matrix that reflects the relationship between features.

[0065] Calculate the eigenvalues ​​and eigenvectors of the covariance matrix Σ. The eigenvalues ​​and eigenvectors reflect the variance and direction of each principal component. The size of the eigenvalue represents the contribution of the corresponding principal component to the variance of the original data. The eigenvalues ​​and eigenvectors are solved by the equation: Σv=λv; where: v is the eigenvector and λ is the eigenvalue.

[0066] Based on the size of the eigenvalues, select the first k principal components. Usually, select the principal component that can explain the most percentage of the total variance. For example, if the first k principal components can explain more than 80% of the total variance, then you can choose the first k principal components.

[0067] The principal component is a linear combination of the original features. When calculating the principal component, each principal component can be expressed as: ;in is the weight of the nth feature in the kth principal component. The principal components are arranged from large to small according to the eigenvalues, and the weight w corresponding to each principal component reflects the contribution of the feature in the principal component. The landslide feature association index is calculated as follows: ;in: is the eigenvalue of the kth principal component, indicating the contribution of this principal component to the variance, is the kth principal component, is the landslide characteristic correlation index. The larger the landslide characteristic correlation index is, the stronger the correlation between regional characteristics and landslide occurrence is.

[0068] Fuzzy logic is used to determine the performance of existing AI models in different areas, including: The landslide dependence comprehensive index and landslide characteristic correlation index SDF are used as the input items of fuzzy logic, and the performance ability of the existing AI model in different regions is used as the output item of fuzzy logic; Landslide dependency comprehensive index: This index combines the impact of different factors (such as soil shear strength, vegetation coverage, precipitation, etc.) on landslide occurrence. Its value usually indicates the level of landslide risk. In the fuzzy logic system, the landslide dependency comprehensive index will be used as an input variable, and its value can be divided into three fuzzy intervals: low, medium, and high. For example: low risk: indicates that the possibility of landslide is small; medium risk: indicates that the landslide risk exists, but it is not particularly significant; high risk: indicates that the possibility of landslide is high and early warning needs to be strengthened.

[0069] Landslide characteristic correlation index: This index analyzes the correlation between the geographical and climatic characteristics of different regions (such as soil shear strength, vegetation coverage, slope, etc.) and the occurrence of landslides, and obtains a value that indicates the degree of influence of regional characteristics on the occurrence of landslides. SDF values ​​can also be divided into different fuzzy levels, such as: low correlation: indicating that the regional characteristics have little impact on the occurrence of landslides; medium correlation: indicating that the regional characteristics have a certain impact on the landslide, but it is not decisive; high correlation: indicating that the regional characteristics play a significant role in the occurrence of landslides and have a high correlation. Output item: the performance ability of existing AI models in different regions.

[0070] According to fuzzy logic reasoning, the performance of the AI ​​model reflects the model's prediction accuracy and reliability for landslide risk in a specific area. The output fuzzy level is usually determined based on a combination of the landslide dependency comprehensive index and the landslide feature association index. The output fuzzy level can be divided into several ranges, such as: Excellent performance: The AI ​​model can accurately identify landslide risks and provide effective early warning, adapting to changing regional characteristics; General performance: The AI ​​model can handle landslide risks in certain areas, but it is unstable or has reduced accuracy in certain areas or under special conditions; Poor performance: The AI ​​model has poor ability to predict landslides in specific areas, which may be due to insufficient model training or large differences in regional characteristics, resulting in unsatisfactory prediction results.

[0071] Fuzzification of input variables: First, the landslide dependence comprehensive index and landslide characteristic correlation index are fuzzified respectively. Through the preset fuzzy rules, the landslide dependence comprehensive index (e.g., low, medium, high) and SDF (e.g., low correlation, medium correlation, high correlation) are transformed into fuzzy sets.

[0072] Rule setting: Set a series of fuzzy logic rules to represent the relationship between landslide risk and model performance in different regions. For example, if the landslide dependency index is high and SDF is highly correlated, the model performance is excellent; if the landslide dependency index is low and SDF is low correlated, the model performance is poor; if the landslide dependency index is medium and SDF is medium correlated, the model performance is fair.

[0073] In the fuzzy reasoning process, the system calculates the output fuzzy set based on these rules, which represents the performance of the model in different areas. At this time, the fuzzy reasoning machine will synthesize the rules under different input conditions to obtain the final output result.

[0074] Through the defuzzification operation, the fuzzy logic output is converted into clear numerical values ​​or classifications to indicate the performance of the AI ​​model in different areas. For example, according to the fuzzy reasoning results, if the comprehensive judgment model performs as "excellent", the final result is a high-precision prediction. If the comprehensive judgment model performs as "average", the final result is a medium-precision prediction. If the comprehensive judgment model performs as "poor", the final result is a low-precision prediction.

[0075] Set the source domain and target domain: Source Domain: A region containing a large amount of labeled data, which usually has sufficient data to effectively train the AI ​​model. Target Domain: A region with different characteristics from the source domain, but with scarce or difficult data collection, which requires transfer learning to improve prediction accuracy. By transferring the knowledge in the source domain model, the AI ​​model in the target domain can adapt to the characteristics of different geographical regions, thereby improving prediction accuracy.

[0076] The key goal of transfer learning is to transfer knowledge from the source domain to the target domain. and target areas , and the source domain has a large amount of labeled data, while the target domain has less data. The process of transfer learning can be briefly described by the following formula: Source Domain Model is the training data from the source domain: ; The goal of the source domain is to learn a mapping function To predict landslide risk: are input characteristics of the source domain (such as soil shear strength, vegetation cover, etc.). is the output label of the source domain (such as whether a landslide occurred or not).

[0077] In the target area Due to the lack of sufficient labeled data, the goal is to transform the model trained in the source domain Migrate to the target domain and make adaptive adjustments. Set the training data of the target domain as: ;in is the input feature of the target domain, is the output label of the target domain, and i is the index in the sample dataset. The goal of transfer learning is to learn a new mapping function , which can be achieved through the model of the source domain The transfer learning is set to optimize the target domain model by introducing a transfer loss term: ;in: Represents the solution to minimize the parameter , The target area The loss function on the target domain model measures the prediction accuracy (usually mean square error or cross entropy loss), Is the source field and target areas The migration loss function between the source domain model and the target domain model measures the difference between the source domain model and the target domain model. u is a hyperparameter that adjusts the contribution of the source domain model knowledge to the adaptability of the target domain model. A larger u emphasizes the migration of source domain knowledge, while a smaller u allows the training of the target domain model to dominate.

[0078] Transfer loss function It mainly measures the inter-domain differences between the source domain and the target domain. In transfer learning, common methods include adversarial training and maximum mean difference (MMD). ,The model is able to reduce the distribution differences between the source and target domains, thereby improving the prediction accuracy of the target domain.

[0079] The effectiveness of transfer learning is determined by evaluating the predictive performance in the target domain. The transferred model is evaluated using test data from the target domain. Performance , the expression is: ; If the performance of the model in the target domain is significantly improved, the transfer learning is considered successful. If the accuracy of the target domain is still low, u can be further adjusted or the transfer strategy can be improved.

[0080] Example 2, please refer to Figure 2 As shown, the AIOT-based Quaternary overburden slope landslide monitoring device described in this embodiment includes a data acquisition module, a risk assessment module, a fuzzy logic analysis module and a transfer learning module; Data acquisition module: obtain soil shear strength data and vegetation coverage data in different areas through remote sensing images and soil database; Risk assessment module: Based on the acquired soil shear strength data and vegetation coverage data, combined with existing monitoring data and trained AI models, comprehensive analysis is performed to assess the landslide risks in different areas; Fuzzy logic analysis module: In risk assessment, AI models are used to analyze the dependence of soil shear strength and vegetation coverage on landslide occurrence areas, and principal component analysis is used to analyze the correlation between regional characteristics and landslide occurrence. Fuzzy logic is then used to determine the performance of existing AI models in different areas. Transfer learning module: Based on the judgment results, different areas are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas. For low-precision prediction areas, transfer learning methods are used to enable existing AI models to migrate between different areas to improve the prediction accuracy of existing AI models.

[0081] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be covered by the protection scope of the present application.

Claims

1. The Quaternary overburden slope landslide monitoring method based on AIOT is characterized by: include: Obtain soil shear strength data and vegetation coverage data in different areas through remote sensing images and soil databases; Based on the soil shear strength data and vegetation coverage data, a comprehensive analysis is conducted in combination with existing monitoring data and trained AI models to assess the landslide risks in different areas. In risk assessment, AI models are used to analyze the dependence of soil shear strength and vegetation coverage on landslide occurrence areas, and principal component analysis is used to analyze the correlation between regional characteristics and landslide occurrence. Fuzzy logic is then used to determine the performance of existing AI models in different regions. According to the judgment results, different regions are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas. For low-precision prediction areas, the transfer learning method is used to enable the existing AI model to migrate between different regions to improve the prediction accuracy of the existing AI model.

2. The method for monitoring landslides in the Quaternary overburden layer based on AIOT according to claim 1 is characterized by: The existing monitoring data include meteorological data, human activity data, geological data and historical landslide data.

3. The method for monitoring landslides in the Quaternary overburden layer based on AIOT according to claim 1 is characterized by: For each target area, the AI ​​model predicts the probability of landslides in the area based on the input soil shear strength, vegetation coverage and environmental factors, and divides the risks into high-risk areas, medium-risk areas and low-risk areas based on the predicted probability. Through the model's prediction results, the landslide risks in different areas are spatially analyzed to generate a landslide risk heat map. The heat map uses gradients of different colors to represent the degree of landslide risk in each area to identify high-risk areas.

4. The method for monitoring landslides in the Quaternary overburden layer based on AIOT according to claim 1 is characterized in that: The dependence of soil shear strength and vegetation coverage on the landslide occurrence area is analyzed to generate a comprehensive landslide dependence index. The method for obtaining the comprehensive landslide dependence index is as follows: A Bayesian network model representing landslide risk is constructed. Each node in the model represents a factor, and the connection between nodes is determined according to the dependency between factors. For each node, a conditional probability table is defined to represent the probability distribution of the node given its parent node. The input data includes the actual observed values ​​of each factor. The structure of the Bayesian network and the conditional probability table are used to derive the posterior probability according to the Bayesian theorem. Based on the known variables input, the Bayesian network is used to update the posterior probability of landslide occurrence as a comprehensive landslide dependency index, which is obtained by calculating the weighted sum of the conditional probability of landslide occurrence based on soil shear strength, vegetation coverage, slope and precipitation.

5. The method for monitoring landslides in the Quaternary overburden layer based on AIOT according to claim 4 is characterized by: The correlation between regional characteristics and landslide occurrence is analyzed by principal component analysis to generate the landslide characteristic correlation index. The landslide characteristic correlation index is obtained as follows: Collect observations of multiple features in the region and represent them as , where n is the number of features, and the observations of each feature in different regions are represented as a matrix X, where each column represents a feature and each row represents a region; the standardized data matrix is ​​marked as Z, where the mean of each column is 0 and the standard deviation is 1; the covariance matrix of the standardized data matrix Z is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are calculated. The eigenvalues ​​and eigenvectors are solved by equations, and the first k principal components are selected according to the size of the eigenvalues. The principal components are linear combinations of the original features, and the expression is: ;in is the weight of the nth feature in the kth principal component, and the landslide feature association index is calculated, and the expression is: ;in: is the eigenvalue of the kth principal component, is the kth principal component, is the landslide characteristic correlation index.

6. The method for monitoring landslides in the Quaternary overburden layer based on AIOT according to claim 5 is characterized by: Fuzzy logic is used to determine the performance of existing AI models in different areas, including: The landslide dependence comprehensive index and landslide characteristic correlation index are used as the input items of fuzzy logic and divided into different fuzzy sets respectively. The performance of existing AI models in different areas is used as the output of fuzzy logic and divided into different fuzzy sets; Formulate fuzzy rules to describe the impact of landslide dependence composite index and landslide characteristic correlation index on the performance ability of existing AI models in different regions; Through fuzzy reasoning, different areas are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas according to the performance capabilities of existing AI models in different areas.

7. The method for monitoring landslides in the Quaternary overburden layer based on AIOT according to claim 6 is characterized by: For low-precision prediction areas, a transfer learning method is used to enable existing AI models to migrate between different areas to improve the prediction accuracy of existing AI models, including: using the training data of the source domain to train a prediction model, using the trained model in the source domain as the preliminary model in the target domain, adjusting the source domain model on the limited training data in the target domain, and adjusting the model parameters to adapt it to the characteristics and task requirements of the target domain; during the training process, introducing a migration loss function to minimize the distribution difference between the source domain and the target domain; continuing to train the target model in the target domain, adjusting the hyperparameters, and improving the prediction accuracy by continuously optimizing the loss function of the target domain model; using the test data set of the target domain to evaluate the performance of the migrated model; if the prediction accuracy of the model in the target domain is improved, the transfer learning is considered successful; if the prediction accuracy of the model in the target domain is not improved, the model parameters or migration strategy need to be further adjusted.

8. A Quaternary overburden slope landslide monitoring device based on AIOT, used to implement the Quaternary overburden slope landslide monitoring method based on AIOT according to any one of claims 1 to 7, characterized in that: It includes data acquisition module, risk assessment module, fuzzy logic analysis module and transfer learning module; Data acquisition module: obtain soil shear strength data and vegetation coverage data in different areas through remote sensing images and soil database; Risk assessment module: Based on the acquired soil shear strength data and vegetation coverage data, combined with existing monitoring data and trained AI models, comprehensive analysis is performed to assess the landslide risks in different areas; Fuzzy logic analysis module: In risk assessment, AI models are used to analyze the dependence of soil shear strength and vegetation coverage on landslide occurrence areas, and principal component analysis is used to analyze the correlation between regional characteristics and landslide occurrence. Fuzzy logic is then used to determine the performance of existing AI models in different areas. Transfer learning module: Based on the judgment results, different areas are divided into high-precision prediction areas, medium-precision prediction areas and low-precision prediction areas. For low-precision prediction areas, transfer learning methods are used to enable existing AI models to migrate between different areas to improve the prediction accuracy of existing AI models.

Images