A geological disaster monitoring and early warning method based on a low-orbit satellite internet of things

Abstract

This invention discloses a geological disaster monitoring and early warning method based on low-Earth orbit (LEO) satellite Internet of Things (IoT). The method specifically includes: real-time acquisition of multi-source data on potential geological disaster sites via LEO remote sensing satellites, ground-based sensor nodes, and a satellite navigation system; spatiotemporal alignment and interpolation of the multi-source data, mapping it onto a three-dimensional spatiotemporal grid centered on the potential geological disaster sites to form a multi-dimensional data cube; constructing mechanistic constraints using geological disaster physical equations, and extracting spatiotemporal correlation features from the multi-dimensional data cube using a spatiotemporal graph neural network; achieving in-depth mining of risk evolution patterns through a mechanism-guided attention mechanism, and outputting a comprehensive risk level and the contribution of disaster-causing factors. This invention integrates multi-source data from LEO satellite IoT to achieve comprehensive and accurate geological disaster monitoring and early warning, improving disaster response capabilities.

Description

Technical Field

[0001] This invention relates to the field of low-Earth orbit satellite Internet of Things (LEO) technology, and in particular to a method for monitoring and early warning of geological disasters based on LEO satellite Internet of Things. Background Technology

[0002] In the crucial field of geological disaster monitoring and early warning, traditional monitoring methods have long relied primarily on ground-based sensor networks and limited high-orbit satellite remote sensing data. Ground-based sensor networks can provide high-precision local monitoring data, which is crucial for accurately understanding local geological conditions. However, they have significant limitations: their coverage is relatively limited, and deployment costs are high. This severely restricts their ability to effectively monitor large areas prone to geological disasters. Especially in remote mountainous areas with complex terrain and extremely inconvenient transportation, the deployment of ground-based sensor nodes faces numerous difficulties, and subsequent maintenance is also challenging, making it difficult to achieve comprehensive coverage of these critical areas.

[0003] While high-orbit satellite remote sensing data offers broad coverage and can acquire geological information at a macroscopic level, its high orbital altitude and imaging resolution limitations restrict its ability to monitor potential geological hazards in real time and capture detailed information. This makes it insufficient to meet the needs of early warning of geological disasters and precise analysis of risk evolution, failing to provide timely and accurate information support for disaster prevention and potentially hindering its effectiveness in responding to sudden geological disasters.

[0004] With the rapid development of technology, low-Earth orbit (LEO) satellite Internet of Things (IoT) technology is beginning to emerge. LEO satellites, with their unique advantages of low orbital altitude, high spatiotemporal resolution, and global coverage, have brought new ideas and broad possibilities to geological disaster monitoring and early warning. LEO satellites can collect multi-source data on potential geological disaster sites in real time, including remote sensing imagery, time-series data from ground sensors, and satellite positioning data. This rich and diverse data contains a wealth of spatiotemporal information, laying a solid foundation for accurate monitoring and effective early warning of geological disasters. However, processing this multi-source data currently faces many challenging issues. How to effectively integrate and fully utilize this data from different sources and of different types, and deeply explore its potential spatiotemporal correlation characteristics and risk evolution patterns, has become a critical problem that urgently needs to be solved.

[0005] Traditional methods have significant shortcomings in data processing. For multi-source data, there is a lack of effective methods for spatiotemporal alignment and interpolation, leading to spatiotemporal inconsistencies between different data sets. This inconsistency makes it difficult to integrate the data into a cohesive whole, negatively impacting subsequent extraction of useful spatiotemporal correlation features and risk analysis, thus reducing the data's usability and the accuracy of the analysis results. Furthermore, traditional methods struggle to organically integrate geologically based constraints with data-driven analysis methods, resulting in insufficient depth in exploring risk evolution patterns.

[0006] Traditional methods also have shortcomings in edge computing and cloud collaboration. Due to a lack of effective control mechanisms, edge nodes often exhibit blindness and inefficiency in data collection and processing. They cannot flexibly adjust monitoring strategies according to actual geological disaster monitoring needs and risk conditions, resulting in a lack of targeted monitoring efforts. This not only affects monitoring effectiveness but also increases system energy consumption and wastes resources. Furthermore, traditional methods lack secure and reliable evidence storage technologies for storing and disseminating critical information for early warning decisions, failing to ensure the authenticity and integrity of the information. Moreover, when issuing early warning information, they cannot adaptively select the optimal path, making it difficult to deliver warning information to relevant personnel in a timely and accurate manner, thus affecting the effectiveness of disaster early warning.

[0007] In the field of model training and optimization, traditional methods often employ centralized training, sending all data to the cloud for training. While this approach can achieve model training to some extent, it increases the cost of data transmission and storage, especially when dealing with large datasets. Furthermore, it poses a risk of data privacy breaches; unauthorized access to cloud data can severely damage geological disaster monitoring efforts. Moreover, traditional methods lack effective means for handling global early warning models, resulting in poor model adaptability to different geological environments and impacting the accuracy and reliability of geological disaster early warnings. Summary of the Invention

[0008] The purpose of this invention is to provide a geological disaster monitoring and early warning method based on low-orbit satellite Internet of Things (LES), which integrates multi-source data from LES to achieve comprehensive and accurate geological disaster monitoring and early warning, thereby solving at least one of the aforementioned problems in the prior art.

[0009] In a first aspect, the present invention provides a geological disaster monitoring and early warning method based on low-orbit satellite Internet of Things, the method specifically comprising: Real-time collection of multi-source data on potential geological disaster sites using low-orbit remote sensing satellites, ground sensing nodes, and satellite navigation systems; Multi-source data is spatiotemporally aligned and interpolated, and mapped onto a three-dimensional spatiotemporal grid centered on geological hazard points to form a multi-dimensional data cube. Mechanism constraints are constructed using physical equations of geological hazards, and spatiotemporal graph neural networks are combined to extract spatiotemporal correlation features from multidimensional data cubes. Through mechanism-guided attention mechanisms, risk evolution patterns are deeply mined, and comprehensive risk levels and the contribution of disaster-causing factors are output. After performing lightweight anomaly detection at the edge nodes, the system dynamically selects collaborative perception, local aggregation, or cloud-based deep analysis strategies based on the anomaly level and network status. The system also dynamically adjusts the sampling frequency and task priority of the edge nodes based on the global situation through the cloud. Key information for early warning decisions is generated and stored on the blockchain. When the overall risk level is determined to exceed a preset risk threshold, a multimodal communication link is triggered simultaneously to adaptively select the optimal path to release standardized early warning information.

[0010] In a second aspect, the present invention provides a computer device, comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the geological disaster monitoring and early warning method based on low-orbit satellite Internet of Things as described in any of the above methods.

[0011] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention integrates multi-source data from low-orbit satellite IoT to achieve comprehensive and accurate geological disaster monitoring and early warning, thereby improving disaster response capabilities.

[0012] 2. This invention maps multi-source data to a three-dimensional spatiotemporal grid centered on geological hazard hazard points through spatiotemporal alignment and interpolation processing, forming a multi-dimensional data cube. This effectively solves the problem of spatiotemporal inconsistency between data and provides a unified data foundation for subsequent spatiotemporal correlation feature extraction and risk analysis.

[0013] 3. This invention utilizes the physical equations of geological disasters to construct mechanistic constraints, and combines spatiotemporal graph neural networks to extract spatiotemporal correlation features from multidimensional data cubes. Through a mechanism-guided attention mechanism, it achieves in-depth mining of risk evolution patterns, accurately outputting the comprehensive risk level and the contribution of disaster-causing factors, thereby improving the accuracy and reliability of geological disaster early warning.

[0014] 4. After performing lightweight anomaly detection at edge nodes, this invention dynamically selects collaborative sensing, local aggregation, or cloud-based deep analysis strategies based on the anomaly level and network status. Furthermore, it dynamically adjusts the sampling frequency and task priority of edge nodes in the cloud based on the global situation, thereby maximizing monitoring benefits and minimizing system energy consumption, and improving the flexibility and efficiency of the entire monitoring system.

[0015] 5. This invention generates blockchain evidence for key information in early warning decision-making, ensuring the authenticity and immutability of the early warning information; when the risk exceeds the preset risk threshold, it synchronously triggers a multimodal communication link, adaptively selects the optimal path to release standardized early warning information, and improves the timeliness and coverage of the early warning information.

[0016] 6. This invention adopts a federated learning approach, using the edge nodes of geological disaster hazard points in various regions as local clients to independently train a lightweight early warning threshold generation model locally. The local model parameters are then uploaded to the cloud for aggregation to generate a global early warning model. Simultaneously, based on the basic geological and environmental parameters of each region stored in the cloud, knowledge distillation and regularization constraints are applied to the global early warning model, improving the model's generalization ability under different geological environments. This enables the generation of localized dynamic early warning thresholds, further enhancing the accuracy and adaptability of geological disaster early warning. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a geological disaster monitoring and early warning method based on a low-orbit satellite Internet of Things according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating a geological disaster monitoring and early warning method based on low-Earth orbit satellite Internet of Things (LEO) according to an embodiment of the present invention is shown below in detail: S101 collects multi-source data on potential geological disaster sites in real time through low-orbit remote sensing satellites, ground sensing nodes, and satellite navigation systems. S102 performs spatiotemporal alignment and interpolation on multi-source data and maps it onto a three-dimensional spatiotemporal grid centered on geological hazard points to form a multi-dimensional data cube. S103 utilizes the physical equations of geological hazards to construct mechanistic constraints, and combines spatiotemporal graph neural networks to extract spatiotemporal correlation features from multidimensional data cubes. Through a mechanism-guided attention mechanism, it achieves in-depth mining of risk evolution patterns and outputs comprehensive risk levels and the contribution of disaster-causing factors. S104 performs lightweight anomaly detection at edge nodes and dynamically selects collaborative perception, local aggregation, or cloud-based deep analysis strategies based on the anomaly level and network status. It also dynamically adjusts the sampling frequency and task priority of edge nodes based on the global situation through the cloud. S105 generates blockchain evidence of key information for early warning decisions and, when the overall risk level is determined to exceed the preset risk threshold, simultaneously triggers a multimodal communication link to adaptively select the optimal path to release standardized early warning information.

[0021] In this embodiment, remote sensing image data from low-orbit remote sensing satellites is used to preliminarily identify potential geological hazard areas such as surface deformation, landslides, and debris flow gullies through visual interpretation or automated image processing algorithms (such as deep learning models). These areas typically have specific topographic features, such as steep slopes and loose deposits.

[0022] Ground-based sensor nodes are deployed in areas initially identified as potential hazard zones to collect real-time time-series data on soil moisture, displacement, and stress. Analysis of this data can further confirm the presence of potential geological hazards. For example, if soil moisture continues to rise abnormally, accompanied by accelerated surface displacement, the area is likely at risk of landslides or debris flows.

[0023] By using satellite positioning data provided by the satellite navigation system, areas identified as having potential geological hazards can be precisely located, thereby determining the location of potential geological hazard points.

[0024] After determining the location of potential geological hazard sites, multi-source data on these sites is collected in real time using low-Earth orbit (LEO) remote sensing satellites, ground-based sensor nodes, and satellite navigation systems. LEO remote sensing satellites, with their low orbital altitude, can acquire remote sensing imagery data of potential geological hazard sites with high spatiotemporal resolution. This imagery data can intuitively reflect changes in the topography and geomorphology of the hazard sites. Ground-based sensor nodes, deployed near the potential geological hazard sites, can collect real-time time-series data from ground sensors, such as soil moisture, displacement, and stress, providing high-precision data support for accurately understanding local geological conditions. Satellite navigation systems can provide satellite positioning data to accurately determine the location information of potential geological hazard sites. Through the coordinated work of these three devices, comprehensive and real-time multi-source data on potential geological hazard sites can be acquired, laying the foundation for subsequent monitoring and early warning work. For example, at a potential geological hazard site in a mountainous area, LEO remote sensing satellites periodically capture images of the area, ground-based sensor nodes continuously monitor soil moisture and mountain displacement, and the satellite navigation system records the location coordinates of the hazard site in real time. This data is then aggregated and transmitted to a data processing center.

[0025] After acquiring multi-source data, spatiotemporal alignment and interpolation are performed. Data from different sources may exhibit inconsistencies in time and space; for example, the acquisition time of low-orbit remote sensing satellites may differ from that of ground-based sensors, or the different location distributions of ground-based sensors may lead to spatial data inconsistencies. Therefore, spatiotemporal alignment is necessary to unify these data across time and space. Then, appropriate interpolation methods are used to map the data onto a three-dimensional spatiotemporal raster centered on geological hazard sites, forming a multidimensional data cube. This integrates data of different types and sources into a unified data structure, facilitating subsequent analysis and processing. For instance, soil moisture data acquired at different times are aligned temporally and then interpolated onto a three-dimensional spatiotemporal raster, forming a multidimensional data cube together with other data.

[0026] Mechanism constraints are constructed using the physical equations of geological hazards, and spatiotemporal graph neural networks are combined to extract spatiotemporal correlation features from a multidimensional data cube. The physical equations of geological hazards, based on geological principles, describe the inherent laws governing the occurrence and development of geological hazards, and using them as mechanistic constraints can guide the direction of data analysis and processing. Spatiotemporal graph neural networks possess powerful data analysis and processing capabilities, enabling the extraction of spatiotemporal correlation features between data. Through a mechanism-guided attention mechanism, in-depth mining of risk evolution patterns is further achieved. In this process, the model comprehensively considers various information within the multidimensional data cube, analyzes the interactions and influences between different factors, and outputs a comprehensive risk level and the contribution of hazard-causing factors. For example, the model analyzes the spatiotemporal correlation between factors such as soil moisture, landslide displacement, and rainfall, determining that the comprehensive risk level of the current geological hazard hazard point is relatively high, and identifying rainfall as the main hazard-causing factor.

[0027] After lightweight anomaly detection at edge nodes, a collaborative sensing, local aggregation, or cloud-based deep analysis strategy is dynamically selected based on the anomaly level and network status. Edge nodes first perform preliminary anomaly detection on the collected data to determine the presence and severity of any anomalies. If the anomaly level is low and the network is stable, a local aggregation strategy can be chosen to perform preliminary data processing and storage. If the anomaly level is high or the network is unstable, a collaborative sensing strategy can be chosen to share data and perform collaborative analysis with other edge nodes. If the anomaly is complex and requires more in-depth analysis, a cloud-based deep analysis strategy is selected to upload the data to the cloud for processing. Simultaneously, the cloud dynamically adjusts the sampling frequency and task priority of edge nodes based on the overall situation. The cloud monitors the entire geological disaster monitoring area in real time, adjusting the sampling frequency of edge nodes according to the risk level and monitoring needs of different areas, increasing the sampling frequency for higher-risk areas and appropriately decreasing it for lower-risk areas. Furthermore, the task priority of edge nodes is adjusted based on the urgency and importance of the tasks to ensure that important monitoring tasks are completed first. For example, in an area with high rainfall, the cloud will increase the sampling frequency of edge nodes in that area and prioritize data analysis tasks in that area to keep abreast of changes in potential geological hazards.

[0028] Key information for early warning decisions is generated and stored on a blockchain. Blockchain technology, with its immutable and secure characteristics, ensures the authenticity and integrity of this information. This key information includes the overall risk level, the contribution of disaster-causing factors, monitoring time, and monitoring location. The blockchain storage provides strong support for subsequent auditing and traceability. When the overall risk level exceeds a preset risk threshold, a multimodal communication link is simultaneously triggered, adaptively selecting the optimal path to release standardized early warning information. The multimodal communication link includes various methods such as BeiDou short message service, satellite broadcasting, cellular networks, and terrestrial broadcasting, allowing for the selection of appropriate communication methods based on different scenarios and user needs. Simultaneously, the system adaptively selects the optimal path to release early warning information based on network conditions, user location, and other factors, ensuring that the early warning information is delivered to relevant personnel in a timely and accurate manner.

[0029] The technical principles of this embodiment include three main aspects: multi-source data fusion and processing technology, hybrid networking architecture, and intelligent edge computing and communication collaboration technology.

[0030] In terms of multi-source data fusion and processing technology, the system integrates multiple data sources such as high-resolution satellite remote sensing imagery and ground sensor data to achieve multi-source data fusion, as detailed below: (1) Multi-source data hybrid modeling: Satellite remote sensing provides macroscopic information such as topography, vegetation cover, and surface deformation over a wide range, while ground sensors can acquire high-precision microscopic data such as displacement, rainfall, and groundwater level. By combining physical mechanism models with deep learning, a hybrid prediction model is created to integrate data from different sources and in different formats to construct a unified data model, thereby reflecting the actual situation of geological hazard hazard points more comprehensively and accurately.

[0031] (2) Adaptive adjustment of early warning threshold: Based on accumulated monitoring data, the system adopts a federated learning framework to integrate the historical evolution patterns of multiple disaster points under the premise of protecting data privacy, and automatically optimizes the early warning thresholds in different regions and geological conditions, thereby significantly improving the accuracy of early warning.

[0032] In terms of hybrid networking architecture, the high-precision positioning function of the BeiDou Navigation Satellite System is utilized, combined with 4G and LEO low-orbit satellite technologies, to achieve real-time monitoring and data transmission of potential geological hazard sites. In complex terrain environments, the hybrid constellation design provides more stable signal coverage, as detailed below: (1) 4G network: It integrates a 4G module. When there is a 4G signal, the 4G network can be used first, which can realize bidirectional, high bandwidth, low latency stable transmission, while reducing terminal power consumption and increasing usage time.

[0033] (2) BeiDou Satellite Navigation: The BeiDou system adopts a unique hybrid networking mode of three types of satellites: geostationary orbit (GEO), inclined geosynchronous orbit (IGSO), and medium Earth orbit (MEO). This constellation design has excellent global coverage performance (horizontal positioning accuracy better than 10 meters globally and better than 5 meters in the Asia-Pacific region). At the same time, its nanosecond-level time synchronization capability provides high-precision positioning for industries such as power and finance.

[0034] (3) LEO low-orbit remote sensing satellites: The remote sensing satellites adopt a multi-orbit plane design, with multiple satellites deployed on each plane to form a network coverage structure. The satellites can carry a variety of sensors such as optical, SAR, hyperspectral, and infrared sensors to achieve comprehensive observation.

[0035] (4) LEO Low Earth Orbit Communication Satellites: By precisely controlling the phase difference of multiple satellites in the orbital plane, a uniformly distributed satellite constellation is formed. The satellites adopt a DCS (Data Collection System) communication system, which is designed for wide-area IoT terminal access. It has a modular, low-power architecture, can operate for long periods of time, and meets the energy constraints of satellites. The terminals use high-gain, low-elevation antennas to effectively improve signal coverage and anti-interference capabilities, ensuring high-sensitivity communication in complex environments.

[0036] Hybrid networking technology provides high spatiotemporal resolution Earth observation data and reliable communication links for geological disaster monitoring and early warning systems. Combined with ground sensor networks, BeiDou positioning systems, and other technologies, it forms a comprehensive geological disaster monitoring system, which greatly improves the accuracy and timeliness of disaster early warning.

[0037] In terms of intelligent edge computing and communication collaboration technology, the system adopts an "edge-cloud" collaborative architecture to achieve low-latency, highly reliable monitoring and early warning, as detailed below: (1) Extreme environment perception and power supply: The latest MEMS micro sensor array is adopted to integrate low power consumption and multi-parameter monitoring capabilities (displacement, tilt angle, vibration, acoustic emission, etc.). Combined with environmental energy harvesting technology (photovoltaic, vibration energy, rainwater energy, etc.), the sensing nodes can be self-powered for a long time, which greatly extends the working cycle of the equipment to more than 5 years.

[0038] (2) Self-organizing network and fault-tolerant communication: Based on the improved LoRaWAN protocol, a self-organizing and self-healing monitoring network is constructed. The system has multi-path routing capabilities and can automatically reconstruct communication paths when some nodes fail. Key data adopts distributed storage technology to ensure data integrity even under extreme conditions.

[0039] (3) Edge intelligent decision-making: A lightweight AI model is deployed on the sensor node side to achieve local data screening and anomaly detection, transmitting only key data and reducing communication load by about 85%. The system can dynamically adjust the sampling frequency based on real-time monitoring data and automatically increase the monitoring frequency when anomalies are detected, thus optimizing energy efficiency.

[0040] In some embodiments, step S102 above, which involves performing spatiotemporal alignment and interpolation on the multi-source data and mapping it onto a three-dimensional spatiotemporal grid centered on geological hazard hazard points to form a multidimensional data cube, specifically includes: Based on a unified spatial coordinate system and time reference standard, multi-source data is transformed to a unified spatiotemporal benchmark through coordinate transformation and time synchronization protocols. For remote sensing images, ground sensor time-series data and satellite positioning data in multi-source data, spatiotemporal alignment processing is performed using geometric correction, resampling interpolation and positioning techniques respectively. The three-dimensional spatial range and raster resolution are defined with geological hazard hazard points as the core, and the time slices are divided at equal intervals along the time dimension to construct a spatiotemporal raster sequence; By using spatial allocation algorithms and attribute interpolation methods, multi-source data that has undergone spatiotemporal alignment is mapped to the corresponding cells of a spatiotemporal raster sequence, generating a three-dimensional attribute field for each time slice. The three-dimensional attribute fields of each time slice are stacked to form a multidimensional data cube, and each raster cell of the multidimensional data cube is assigned a unique spatiotemporal identifier and associated with multidimensional attributes.

[0041] In this embodiment, due to the different sources of multi-source data, their spatial coordinate systems and time reference standards differ. To ensure consistency in subsequent processing, it is necessary to transform the multi-source data based on a unified spatial coordinate system and time reference standard through coordinate transformation and time synchronization protocols. For example, the original coordinate system of data collected by ground sensor nodes may differ from that of low-Earth orbit remote sensing satellites or satellite navigation systems. A coordinate transformation algorithm is used to transform the coordinates of the ground sensor node data to the same spatial coordinate system as low-Earth orbit remote sensing satellites and satellite navigation systems. Simultaneously, for time information from different data sources, a time synchronization protocol is used to unify the time reference standard of all data. For example, the times recorded by different devices are uniformly converted to Coordinated Universal Time (UTC), thereby transforming the multi-source data to a unified spatiotemporal reference and laying the foundation for subsequent spatiotemporal alignment processing.

[0042] For remote sensing imagery, ground sensor time-series data, and satellite positioning data from multiple sources, different techniques are employed for spatiotemporal alignment. For remote sensing imagery, due to potential geometric distortions during acquisition, geometric correction techniques are used. Based on known ground control points or digital elevation models, the imagery is geometrically corrected to eliminate distortions and ensure accurate correspondence with geographic information in a unified spatial coordinate system. For ground sensor time-series data, since different sensors may have different sampling frequencies, leading to inconsistencies in the time dimension, resampling interpolation techniques are used. Based on a unified time reference standard, the ground sensor time-series data is resampled to unify time intervals, achieving temporal alignment. For satellite positioning data, to improve positioning accuracy, positioning technology is used. This combines signal information from multiple satellites and data from ground reference stations to process the satellite positioning data, making its spatial location more accurate and consistent with positional information in a unified spatial coordinate system, thus completing the spatiotemporal alignment of the multi-source data.

[0043] Focusing on potential geological hazard sites, and based on actual monitoring needs and the characteristics of geological hazards, a three-dimensional spatial range and raster resolution are defined. For example, based on the scale and potential impact range of the potential geological hazard site, a suitable three-dimensional spatial range is determined, such as a cylindrical or cuboid space with a certain radius centered on the hazard site. Simultaneously, the raster resolution is determined according to the required data accuracy and level of monitoring detail, such as dividing the spatial range into cubic raster cells with certain side lengths. In the time dimension, equally spaced time slices are divided along the time axis. The time interval can be determined based on the development speed of the geological hazard and the required monitoring frequency, such as dividing a time slice every hour or every day, thereby constructing a spatiotemporal raster sequence and providing a structural framework for subsequent data mapping.

[0044] By employing spatial allocation algorithms and attribute interpolation methods, multi-source data, after spatiotemporal alignment, is mapped to corresponding cells in a spatiotemporal raster sequence. For the spatial allocation algorithm, the raster cell to which a data point belongs is determined based on its spatial location information, and the data point is assigned to the corresponding raster. For the attribute interpolation method, since each raster cell may not have a directly corresponding data point, interpolation calculations are performed based on the attribute information of surrounding data points to determine the attribute value of that raster cell. For example, for temperature data collected by ground sensors, when mapping it to raster cells, if there is no sensor in that raster cell, the temperature attribute value of that raster cell can be calculated using methods such as inverse distance weighted interpolation based on temperature data collected by surrounding sensors. This generates a three-dimensional attribute field for each time slice, which contains various attribute information of each raster cell at a specific time, such as temperature, humidity, and displacement.

[0045] The three-dimensional attribute fields of each time slice are stacked in chronological order to form a multidimensional data cube. This multidimensional data cube contains information not only in the spatial dimension but also in the temporal dimension, comprehensively reflecting the changes in various attributes of geological hazard hazard points at different times and in different spaces. Simultaneously, each raster cell in the multidimensional data cube is assigned a unique spatiotemporal identifier. This identifier accurately represents the spatial and temporal location information of the raster cell and is associated with multidimensional attributes; that is, the identifier of each raster cell corresponds to the various attribute information it contains, facilitating subsequent data querying, analysis, and processing, and providing rich data support for accurate monitoring and effective early warning of geological hazards.

[0046] Furthermore, the step of mapping the spatiotemporally aligned multi-source data to corresponding cells of the spatiotemporal raster sequence using spatial allocation algorithms and attribute interpolation methods to generate a three-dimensional attribute field for each time slice specifically includes: Determine the raster cell to be mapped in the spatiotemporal raster sequence and its corresponding time slice; Using a spatiotemporal kriging interpolation model, based on the spatiotemporal coordinates of the grid cell, the spatiotemporal variability function value between the grid cell and all spatiotemporally aligned point monitoring data is calculated. Based on the spatiotemporal variability function value, the Kriging equation system is constructed and solved to obtain the interpolation weight of each point monitoring data. The attribute values ​​of the point monitoring data are then interpolated and weighted and fused using the interpolation weight. For the spatiotemporally aligned area monitoring data, the spatial overlap between the area monitoring data and the grid cells is calculated, and the spatial overlap is used as a weight to perform spatial weighted fusion of the attribute values ​​of the area monitoring data. The interpolation weighted fusion results of point monitoring data and the spatial allocation weighted fusion results of area monitoring data are integrated to generate the fusion attribute values ​​of raster cells, and all raster cells are traversed to form the three-dimensional attribute field of the corresponding time slice.

[0047] In this embodiment, within the constructed spatiotemporal raster sequence, the target to be processed is determined according to a preset order or a random selection method; that is, the raster cell to be mapped is identified, along with the time slice to which the raster cell belongs. For example, when processing geological disaster monitoring data at a specific moment, the time slice corresponding to that moment is first determined. Then, among the numerous raster cells contained in this time slice, one is selected as the raster cell to be mapped, providing a specific processing object for subsequent data mapping operations.

[0048] For point monitoring data that has undergone spatiotemporal alignment processing, these data typically exist in the form of discrete points, recording various attribute information of geological hazard hazard points at different spatiotemporal locations. Using a spatiotemporal kriging interpolation model, the spatiotemporal variability function value between the current raster cell to be mapped and all point monitoring data is calculated based on the spatiotemporal coordinates of the raster cell.

[0049] The spatiotemporal kriging interpolation model is an extension of traditional kriging interpolation theory in the spatiotemporal dimension. By introducing a time variable and optimizing spatial correlation analysis, it forms an improved method suitable for dynamic geological monitoring scenarios. Specifically, it analyzes the spatiotemporal distribution characteristics of multi-source data (such as time-series data from ground sensors, satellite remote sensing imagery, and satellite positioning data) collected from geological hazard hazard points. By statistically analyzing the spatial coverage and density differences of different types of data, as well as the temporal sampling frequency and dynamic change patterns, it clarifies the non-uniformity and correlation characteristics of the data in the spatiotemporal dimension. The spatiotemporal kriging interpolation model also designs a spatiotemporal variogram function to quantify the spatiotemporal correlation between data points. This function must simultaneously consider the impact of spatial distance and time interval on attribute value differences. For example, by analyzing the time lag effect and spatial diffusion pattern of monitoring data in historical geological hazard events, the spatiotemporal weight allocation ratio is determined. During parameter calibration, historical monitoring datasets of geological hazard hazard points (covering different geological conditions and disaster stages) are used, and the variogram function parameters, including spatial anisotropy parameters and time decay coefficients, are optimized through cross-validation. For example, in landslide monitoring scenarios, by comparing the interpolation errors under different parameter combinations, the parameter combination that minimizes the root mean square error is selected as the default parameters of the model to ensure the model's adaptability to the target area.

[0050] The spatiotemporal variability function (SVC) value reflects the spatiotemporal variation characteristics of point monitoring data. By calculating this value, we can understand the spatiotemporal correlation between different point monitoring data and the current raster cell, providing a basis for subsequently determining interpolation weights. For example, if a point monitoring data is very close to the current raster cell in terms of spatiotemporal distribution, the SVC value between them may be small, indicating that the point monitoring data has a significant impact on the properties of the current raster cell.

[0051] Based on the calculated spatiotemporal variability function values, a set of Kriging equations is constructed. The Kriging equations are a mathematical model based on spatial statistics. By solving this set of equations, the interpolation weights of each point monitoring data point can be obtained. These weights reflect the contribution of different point monitoring data points to the attribute values ​​of the current raster cell. After obtaining the interpolation weights, the attribute values ​​of the point monitoring data are interpolated and weighted for fusion. Specifically, the attribute value of each point monitoring data point is multiplied by its corresponding interpolation weight, and then the weighted results of all point monitoring data points are summed to obtain the fusion result of the point monitoring data points on the attribute values ​​of the current raster cell. For example, if there are three point monitoring data points with attribute values ​​A, B, and C, and corresponding interpolation weights a, b, and c, respectively, then the fusion result of the point monitoring data points on the attribute values ​​of the current raster cell is a×A + b×B + c×C.

[0052] For spatiotemporally aligned areal monitoring data, which refers to a continuous data set acquired through systematic monitoring methods for a specific geographical area or geological surface, comprehensively reflecting the spatial distribution characteristics and dynamic changes of that area. Areal monitoring data is characterized by continuous spatial coverage. It involves deploying dense monitoring networks or utilizing remote sensing technology to conduct full-area or gridded sampling of target areas (such as geological hazard zones, urban surface subsidence zones, farmland soil moisture distribution zones, etc.) to obtain data reflecting spatial heterogeneity. Essentially, it expands discrete monitoring points into a continuous surface, revealing the distribution patterns and evolution trends of spatial variables (such as displacement, stress, humidity, etc.). Areal monitoring data typically exists in the form of regions, such as remote sensing imagery data of a specific region or attribute data of a certain area covered by a ground-based sensor network. The spatial overlap between the areal monitoring data and the current raster cell is calculated. This spatial overlap can be determined by calculating the proportion of the overlapping area between the raster cell and the areal monitoring data region to the total area of ​​the raster cell. The spatial overlap is then used as a weight to spatially allocate and weight the attribute values ​​of the areal monitoring data. That is, the attribute value of the areal monitoring data is multiplied by its spatial overlap with the current raster cell to obtain the fusion result of the areal monitoring data with the attribute value of the current raster cell. For example, if the attribute value of the areal monitoring data is D and its spatial overlap with the current raster cell is d, then the fusion result of the areal monitoring data with the attribute value of the current raster cell is d×D.

[0053] The interpolation-weighted fusion results of point monitoring data and the spatial allocation-weighted fusion results of areal monitoring data are integrated to obtain the fused attribute value of the current raster cell. This fused attribute value comprehensively considers the influence of point and areal monitoring data on the attributes of the current raster cell, and can more accurately reflect the geological attributes of the current raster cell at a specific time. After processing the current raster cell, all raster cells under that time slice in the spatiotemporal raster sequence are traversed in the same way, and a fused attribute value is generated for each raster cell, finally forming a three-dimensional attribute field for the corresponding time slice. This three-dimensional attribute field contains various attribute information of all raster cells under that time slice, providing an important data foundation for subsequent geological disaster monitoring and early warning analysis.

[0054] In some embodiments, step S103 above, which involves constructing mechanistic constraints using geological hazard physical equations and extracting spatiotemporal correlation features from a multidimensional data cube using a spatiotemporal graph neural network, and achieving in-depth mining of risk evolution patterns through a mechanism-guided attention mechanism to output a comprehensive risk level and the contribution of disaster-causing factors, specifically includes: Extract all grid cells of the target geological hazard hazard body from the multidimensional data cube, and calculate the physical state index of each grid cell based on the geological hazard physical mechanism model; Using each grid cell as a node, a spatiotemporal graph neural network is constructed based on the correlation between its spatial adjacency and physical state indicators; The physical state index is transformed into prior constraints of nodes and edges. These prior constraints are then injected into the attention weight calculation mechanism of the spatiotemporal graph neural network through a learnable mapping function, forming a mechanism-guided attention module. The mechanism-guided attention module performs multi-layer message passing and aggregation on the multi-dimensional temporal attributes of nodes to update the node's own state and obtain the updated node. Based on the updated nodes, the comprehensive risk level of the target geological hazard body and the contribution of each disaster-causing factor are output through classification and regression heads, respectively.

[0055] In this embodiment, the spatial extent of the target geological hazard hazard body is located from the constructed multidimensional data cube. For example, in a landslide monitoring scenario, the slope area where a landslide may occur is delineated based on historical disaster records or terrain analysis results. Subsequently, complete data records of all grid cells within this area are extracted, including multi-source attributes of each grid cell in the time dimension (such as surface displacement, soil moisture content, groundwater level, etc.). The data of each grid cell must include spatial coordinates, timestamps, and at least three different types of monitoring indicators to ensure the comprehensiveness of subsequent analysis. For example, a grid cell may simultaneously record hourly rainfall, changes in surface crack width, and deep displacement rate; these data will serve as the basis for physical state calculations.

[0056] Based on a physical mechanism model of geological hazards, monitoring data for each grid cell is processed to calculate its physical state indices. For example, in debris flow monitoring, the stability coefficient of each grid cell is calculated using an infinite slope model based on soil mechanical parameters and rainfall data; in earthquake liquefaction monitoring, the liquefaction potential index is calculated using the relationship between dynamic pore water pressure and effective stress. These indices need to reflect the instability probability of the geological body under specific conditions and must be strongly correlated with actual geological conditions (such as soil type and slope). During the calculation process, a geological expert knowledge base is introduced to calibrate the model parameters, for example, adjusting the soil internal friction angle parameter based on field sampling results to ensure the accuracy of the physical state indices. Finally, each grid cell will generate a set of physical state indices in scalar or vector form, serving as prior constraints for subsequent graph neural network analysis.

[0057] A spatiotemporal graph neural network is constructed using grid cells as nodes. Spatial adjacency is determined by both geographic distance and topographic features: if two grid cells are spatially adjacent (e.g., sharing edges or vertices), a spatial edge is established; if topographic connectivity exists (e.g., cells within the same gully), a spatial edge is established even if they are not adjacent. Temporal adjacency is defined by the continuity of time series: observations of the same grid cell at different time steps are connected sequentially to form temporal edges. Edge weights are dynamically calculated based on the correlation of physical state indicators: for example, if the stability coefficients of two adjacent grid cells change synchronously, their spatial edge weight increases; if the groundwater level of a grid cell has a lagged correlation with the rainfall of a downstream cell, its temporal edge weight is adjusted to reflect this influence. In this way, the graph structure incorporates both static geographic information and dynamic geological process associations.

[0058] An attention mechanism is introduced into the spatiotemporal graph neural network, transforming physical state indicators into prior constraints. Specifically, the physical state indicators (such as stability coefficients) of each node are converted into attention weight bias terms through a learnable nonlinear mapping function. For example, nodes with lower stability coefficients receive higher attention weights during message propagation, indicating a greater influence of the node on the surrounding area. Simultaneously, edge weights (such as terrain connectivity strength) are also converted into attention weight adjustment factors through a similar mapping function, allowing physically stronger edges to contribute more to message propagation. This design ensures that the attention mechanism prioritizes geomechanical principles while considering statistical correlations in the data. For instance, when propagating landslide risk, the model naturally focuses more on areas with high slopes and low stability coefficients, rather than solely relying on spatial proximity.

[0059] In one possible implementation, let the physical state index, such as the stability coefficient, be x. This index is then transformed into an attention weight bias term b through a learnable nonlinear mapping function f(x). The learnable nonlinear mapping function f(x) can be expressed as: b = f(x), f(x) = g(Wx + c), where W is a learnable weight matrix, c is a learnable bias term, and g is a nonlinear activation function.

[0060] Message passing and aggregation are performed at each layer of the spatiotemporal graph neural network. Within each layer, nodes collect information from neighboring nodes based on attention weights and update their own state. For example, the first layer might focus on the propagation of local physical states (such as rainfall infiltration causing a synchronous increase in soil moisture content in adjacent units), while the second layer captures regional risk evolution (such as multiple locally unstable units forming a continuous slip surface). During message passing, prior constraints on physical state indicators continuously guide attention allocation, preventing the model from falling into erroneous associations caused by data noise. After 3-5 layers of transmission, the final state of a node will comprehensively reflect its historical physical state, neighborhood influence, and global risk trend. For example, the final state of a certain grid cell might show it in a "high-risk - rainfall-dominated" mode, indicating that the risk of this cell is mainly caused by continuous rainfall and is approaching a critical instability state.

[0061] Based on the updated node states, the final results are output through classification and regression heads. The classification head divides the overall risk of the target geological hazard body into multiple levels (e.g., low, medium, high, extremely high), and is trained using a multi-class cross-entropy loss function. The regression head quantifies the contribution of each hazard-causing factor, such as outputting the weight percentage of factors like rainfall, earthquakes, and human activities on the current risk, and is trained using a mean squared error loss function. The output results must meet interpretability requirements: the risk level must be correlated with a physical state index threshold (e.g., a stability coefficient <0.8 corresponds to high risk), and the contribution of hazard-causing factors must be consistent with the attention weight distribution in the message passing process. For example, if the risk level of a landslide is "extremely high," and the regression head outputs a rainfall contribution of 70%, it indicates that the model considers continuous heavy rainfall to be the dominant factor, and this conclusion can be traced back to the high attention weight of rainfall-related edges in the graph neural network.

[0062] This embodiment solves the problem of lack of physical interpretability in traditional data-driven methods by deeply integrating prior knowledge of geomechanics into spatiotemporal graph neural networks, and significantly improves the accuracy and reliability of geological disaster early warning.

[0063] Furthermore, the calculation of the physical state indicators of each grid cell based on the geological disaster physical mechanism model specifically includes: Based on the target geological hazard type, the corresponding physical control equations are matched from a pre-built physical equation library, and the physical control equations are spatially discretized to be suitable for grid cell calculation. The parameter fields of each grid cell are extracted from the multidimensional data cube, and the parameter fields include geomechanical parameters, hydrological parameters and load parameters; Based on the discretized physical control equations, the parameter fields of all grid cells are solved simultaneously using a parallel computing architecture of a graphics processor to obtain the physical state index field.

[0064] In this embodiment, a physical equation library for geological hazards is pre-built and maintained, which stores the physical control equations for different types of geological hazards. For example: For landslide hazards, physical equations based on limit equilibrium theory are mainly used, such as simplified Bishop slice localization equations, Janbu slice localization equations, and Fredlund equations considering the strength of unsaturated soil. For example, existing techniques (Azarafza, M., Akgün, H., Ghazifard, A., Asghari-Kaljahi, E., Rahnamarad, J., & Derakhshani, R. (2021). Discontinuous rock slope stability analysis by limit equilibrium approaches-a review. International Journal of DigitalEarth, 14(12), 1918-1941.) start from the basic principles of limit equilibrium methods and discuss in detail the application of these methods in the stability analysis of discontinuous rock slopes, covering different types of slope instability mechanisms, including wedge failure, plane failure, toppling failure, and rotational failure. Various limit equilibrium methods are listed, including the Fellenius method, the simplified Bishop method, and the Janbu method. Their differences and applicable ranges in calculating the safety factor are compared. Specific application cases of these methods in different instability mechanisms are provided, demonstrating how to select the appropriate analysis method based on the geometric characteristics and geological conditions of the slope.

[0065] For debris flow disasters, physical equations based on fluid dynamics and sediment kinematics are mainly used, such as the modified Saint-Venant equations and Bingham fluid constitutive equations.

[0066] For land subsidence disasters, physical equations based on soil consolidation theory are mainly used, such as Terzaghi's one-dimensional consolidation equation and simplified forms of Biot's consolidation theory.

[0067] Once the target geological hazard type is determined, the system automatically matches one or more core physical control equations from the database. Subsequently, these continuous equations are discretized using the finite difference method or the finite volume method to form a set of algebraic equations suitable for calculations using three-dimensional grid cells. During discretization, each grid cell is treated as a control volume, and the differential terms in the equations (such as gradients and divergences) are converted into the difference form of the attribute values ​​of adjacent cells.

[0068] Specifically, based on the target geological hazard type (such as landslide, debris flow, ground subsidence, etc.), the corresponding physical control equations are automatically matched from a pre-built physical equation library for geological hazards. After matching, the continuous physical control equations are spatially discretized, dividing the target area into regular grid cells (such as a 10m×10m×1m three-dimensional grid), and the spatial derivative terms (such as gradient and divergence) in the equations are converted into a difference format based on the adjacency relationship of the grid cells. For example, the slope term in the landslide stability coefficient equation is obtained by calculating the ratio of the elevation difference between the current grid cell and the horizontal distance between adjacent cells. The discretized equations must satisfy numerical stability conditions to ensure convergence in subsequent parallel computations.

[0069] The parameter field data for each raster cell in the multidimensional data cube is extracted, including geomechanical parameters (such as soil density, internal friction angle, and cohesion), hydrological parameters (such as groundwater level, pore water pressure, and rainfall intensity), and load parameters (such as seismic acceleration and artificially applied weight). The parameter field data must cover all raster cells of the target geological hazard area and include continuous observations over time. For example, raster cells in a landslide monitoring area may simultaneously record hourly rainfall, groundwater level changes, and surface displacement data. After extraction, the parameter field is preprocessed, including data cleaning (removing outliers), missing value interpolation (such as filling in missing groundwater level data using spatiotemporal kriging), and unit standardization (such as converting seismic acceleration from g to m / s²). 2 The preprocessed parameter field needs to be stored in the form of a structured array, with each grid cell corresponding to a time series vector containing multiple parameters, providing standardized input for subsequent parallel computing.

[0070] To improve computational efficiency, a parallel computing architecture using a graphics processing unit (GPU) is employed to simultaneously solve the physical control equations for all grid cells. Specifically, the discretized physical control equations are converted into parallel computing tasks executable by the GPU, with each grid cell's computation process mapped to an independent computation thread. For example, in landslide stability calculations, each thread independently calculates the stability coefficient based on the soil parameters (density, internal friction angle) and hydrological parameters (groundwater level) of its corresponding grid cell, with all threads executing simultaneously to achieve large-scale parallel computation. To optimize computational performance, a data partitioning strategy is designed: the entire monitoring area is divided into several sub-regions, and the grid cell data for each sub-region is allocated to different computing cores of the GPU, reducing data access conflicts between threads. Simultaneously, shared memory technology is used to cache frequently accessed parameters (such as global gravitational acceleration values), further accelerating the computation process. Through this design, the physical state indicators of thousands of grid cells can be calculated within seconds, meeting real-time monitoring requirements.

[0071] Based on the results of parallel computing, a physical state index field is generated for the target geological hazard body. This physical state index field is a three-dimensional field (two-dimensional space + one-dimensional time) consistent with the original grid, where each grid cell corresponds to a physical state index value at each time step. For example, in the landslide stability index field, the value of each grid cell represents its stability coefficient at the current moment (between 0 and 1, with smaller values ​​indicating greater instability); in the debris flow dynamic index field, the value of each grid cell represents its momentum flux. After generation, the physical state index field is validated: on the one hand, by comparing it with field monitoring data (such as measured values ​​from surface displacement gauges and pore water pressure gauges), the accuracy of the calculation results is ensured; on the other hand, the rationality of the index field is checked using a geological expert knowledge base, for example, excluding abnormal grid cells with stability coefficients > 1. If the validation fails, the parameters of the physical control equations are adjusted or the parameter field data is re-extracted until the index field meets the accuracy requirements. The final output physical state index field will be used as input to a spatiotemporal graph neural network for subsequent risk evolution pattern mining.

[0072] To adapt to the dynamic evolution of geological hazards, a dynamic update mechanism for the physical state index field is designed. As low-orbit satellite IoT continuously collects new multi-source data, the multi-dimensional data cube is updated in real time, and the system triggers a recalculation process for the physical state index every 10 minutes. During recalculation, only the changed parts of the parameter field (such as the latest rainfall and groundwater level) need to be updated, while the stable parts from historical calculations (such as soil mechanical parameters) are reused, significantly reducing the computational load. Simultaneously, a real-time feedback channel is established: if the physical state index of a certain grid cell exceeds a preset threshold (such as a landslide stability coefficient <0.8), the system immediately marks the cell as a high-risk area and prioritizes edge nodes to perform intensive monitoring of that area (such as increasing the sampling frequency to once per minute). Through this dynamic update and feedback mechanism, the physical state index field ensures that it always reflects the latest state of the geological hazard potential body, providing a reliable basis for accurate early warning.

[0073] This embodiment achieves efficient and accurate calculation of physical state indicators of large-scale grid cells by deeply integrating the classic physical mechanism model of geological disasters with GPU parallel computing technology.

[0074] Furthermore, the mechanism-guided attention module performs multi-layer message passing and aggregation on the multi-dimensional temporal attributes of nodes to update the node's own state and obtain the updated node, specifically including: Based on the time-series data of the physical state indicators of each node, the spatiotemporal correlation between any two nodes is calculated, and the spatiotemporal correlation is normalized into physical guidance weights. In the mechanism-guided attention module, the physical guidance weight is used as a bias term in the attention score calculation to adjust the attention weight distribution of message passing between nodes and obtain the adjusted attention coefficient. The multi-dimensional temporal attribute features of nodes are weighted and aggregated using the adjusted attention coefficients, graph convolution feature updates are performed, and temporal convolution is connected after each graph convolution layer to capture the temporal evolution pattern of node features. Repeated multi-layer graph convolution and temporal convolution operations allow the node features to continuously integrate neighborhood information and temporal dynamics.

[0075] In this embodiment, the specific implementation steps of the mechanism-guided attention module's multi-dimensional temporal attribute message passing and aggregation method are as follows: 1. Calculation of physical guidance weights between nodes For any two nodes in the spacetime graph and The time series data of its physical state indicators are as follows: and Where T is the time step, As a physical state index dimension, For dimension The matrix is ​​then used. First, the spatiotemporal correlation between the physical state indices of two nodes is calculated. To do this, the concept of a spatiotemporal correlation function is introduced, which describes the correlation structure of the physical state indices in the spatiotemporal domain.

[0076] Spatiotemporal correlation function Defined as: in, This is a spatiotemporal correlation function used to describe the correlation structure of physical state indices in the spatiotemporal domain. It is a spatial distance vector. The time lag represents the offset of the physical state index in the time dimension, reflecting the changes of the physical state index over time and the temporal relationship of physical state changes between different nodes. To satisfy the spatial distance is Time lag is The number of sample pairs, For nodes In position The physical state index value at time t. For nodes In position ,time The physical state index values.

[0077] Regarding time lag Taking landslide monitoring as an example, assuming multiple monitoring nodes are set up on the landslide, and each node monitors the physical state index of displacement with a time lag. This represents the displacement index of a node relative to another node or itself at different times, in the time dimension. Assume the displacement of node A at time t is... In time The displacement is By comparing the displacement values ​​at these two moments, we can understand how the displacement of node A changes over time. If By setting the value to 1 day, we can observe the change in displacement of node A within a day, thereby determining the activity of the landslide body at that node.

[0078] Furthermore, assuming there are nodes A and B, with time lag... It can be used to analyze the order and correlation of displacement changes between node A and node B. For example, when When = 0, compare the displacements of node A and node B at the same time t. and This allows us to understand the displacement state of two nodes at the same moment; when When >0, for example =1 day, compare the displacement of node A at time t. and node B in time displacement This allows analysis of whether the displacement change at node B lags behind that at node A, and the length of that lag. If it is found that the displacement change at node B is always one day later than that at node A, and the trends are similar, it can be inferred that the displacement change of the landslide body may have a temporal relationship of propagation from node A to node B. This is of great significance for analyzing the movement mechanism of the landslide body and predicting its development trend. In geological disaster monitoring, time lag, by appropriately selecting units of time and distance, can effectively reflect the changes in physical state indicators over time and the temporal relationship of physical state changes between different nodes, providing important evidence for early warning and prevention of geological disasters.

[0079] Further assuming a landslide area, node A is located in the upper part of the landslide body near the potential initiation point of the sliding surface, and node B is located in the lower part of the landslide body. The monitored physical state index is the horizontal displacement of the nodes, with the time unit selected as days (d) and the displacement unit selected as meters (m). The displacement data of nodes A and B recorded during 10 consecutive days of monitoring are as follows: To find time lag Correlation analysis can be used to calculate different The displacement of node A at time t and the displacement of node B at time t are shown below. The correlation between displacements.

[0080] when When =0, compare the displacements of node A and node B at the same time and calculate their correlation coefficient.

[0081] when When =1, compare the displacement of node A at time t with the displacement of node B at time t+1. For example, compare the displacement of node A on day 1 (0.2m) with the displacement of node B on day 2 (0.1m), the displacement of node A on day 2 (0.5m) with the displacement of node B on day 3 (0.3m), and so on, and calculate the correlation coefficient at this time.

[0082] Change in sequence value( =2, =3, etc.), repeat the above comparison and calculation process, and finally obtain different results respectively. The displacement of node A at time t and the displacement of node B at time t are shown below. The correlation coefficient between displacements, and the value of the correlation coefficient at its maximum. The value represents the time lag of the displacement change of node B relative to node A.

[0083] The correlation coefficient can be calculated using the Pearson correlation coefficient formula. for: n is the number of data points, i.e., the number of days monitored. It is the displacement value of node A on day i. Is node B in the ? Displacement value of the day, and These are the average displacement values ​​of node A and node B, respectively.

[0084] when When =0, the correlation coefficient for: when When =1, the correlation coefficient for: when When =2, the correlation coefficient for: And so on, which will not be elaborated further here.

[0085] In practical calculations, the spatiotemporal correlation function fitted using an exponential model is: in, Let d be the spatiotemporal correlation function fitted using an exponential model, where d is the spatial Euclidean distance between the two nodes. For nugget constant, For sill values, and These are spatial range and temporal range, respectively.

[0086] Spatiotemporal correlation between two nodes The spatiotemporal correlation function fitted by the exponential model can be derived as follows: Will Normalized to the [0,1] interval, the physical guidance weights are obtained. : in, and These are the minimum and maximum values ​​of the correlation between all node pairs, respectively.

[0087] 2. Mechanism-guided attention coefficient calculation In each layer of the spatiotemporal graph attention network, for nodes and its neighboring nodes , Represents a node The attention coefficient is calculated for the set of neighboring nodes. First, the node features (including observed attributes and physical state indicators) are transformed through a shared linear transformation, and then a feature-based attention score is calculated. Specifically, for a node... and Their attention score The calculation is as follows: in, and They are nodes and nodes The node features are denoted by W, which is a learnable weight matrix. For attention weights, Attention weights The transpose of , and || for concatenation. for Activation function.

[0088] Then, physical guiding weights Introducing attention scores as a bias term: in, To incorporate physically guided attention scores, the effects of feature-based attention scores and physical guidance were considered comprehensively. These are learnable scaling parameters used to control the strength of the physics guidance.

[0089] Finally, the attention score is normalized using the softmax function to obtain the attention coefficient. : Indicates a node Attention score after physical guidance of all neighboring nodes The summation after exponentiation is used as the denominator for normalization.

[0090] 3. Message passing and feature aggregation After obtaining the attention coefficient, the node The updated features are obtained by weighted aggregation of the features of neighboring nodes: in, For nodes The updated feature vector is obtained by weighted aggregation of the features of neighboring nodes. It is a non-linear activation function. Indicates a node Feature vectors of all neighboring nodes Perform a weighted summation, with the weights being the attention coefficients. This enables communication from neighboring nodes to the central node. Message passing and feature aggregation.

[0091] To capture the temporal evolution of node features, a Temporal Convolutional Network (TCN) is introduced after each graph attention layer. Specifically, for each node, its features at consecutive time steps are treated as a time series, and convolutional operations are performed using the TCN to extract temporal features.

[0092] 4. Multi-layer stacking and high-level feature generation By stacking multiple layers of the aforementioned mechanism-guided graph attention layers and temporal convolutional layers, node features continuously fuse broader neighborhood information and more complex temporal patterns. Ultimately, each node obtains a high-level embedded feature that integrates multi-source observation data, physical state indicators, and spatiotemporal correlation information guided by physical mechanisms.

[0093] 5. Risk assessment output The final high-level embedding features are input into two independent multilayer perceptrons (MLPs), used for integrated risk level classification and hazard factor contribution regression, respectively. For risk level classification, a softmax activation function is used to output the probability of each risk category; for hazard factor contribution regression, a sigmoid or linear activation function is used to output the contribution score of each factor.

[0094] This embodiment integrates the physical mechanisms of geological disasters into an attention mechanism, thereby achieving efficient aggregation of multi-dimensional temporal attributes and in-depth mining of risk evolution patterns.

[0095] In some embodiments, in step S104 above, after performing lightweight anomaly detection at the edge node, dynamically selecting a collaborative sensing, local aggregation, or cloud-based deep analysis strategy based on the anomaly level and network status specifically includes: A lightweight anomaly detection model deployed at the edge nodes of geological disaster hazard points analyzes local sensor data in real time and outputs anomaly confidence levels. If the anomaly confidence level is low and the current network status is good, a collaborative awareness request is initiated to neighboring nodes to verify the anomaly. If the confidence level is medium, then multi-node data is aggregated at the cluster head gateway for correlation analysis; If the confidence level is high or the network is interrupted, the compressed and encrypted critical data will be requested for deep inference in the cloud via an emergency link.

[0096] In this embodiment, a lightweight anomaly detection model is pre-loaded into the edge nodes deployed at geological hazard sites. This model, trained on historical data, can quickly analyze the temporal characteristics of local sensor data (such as displacement, tilt angle, and soil moisture content). The edge nodes receive multi-source data in real time from ground sensor nodes, low-orbit remote sensing satellites, and satellite navigation systems. After data cleaning and standardization, the data is input into the lightweight model for anomaly analysis. The model output is an anomaly confidence level, with values ​​divided into three levels: low (0-30%), medium (30%-70%), and high (70%-100%), representing the probability that the current data deviates from the normal pattern.

[0097] Specifically, the training steps for the lightweight anomaly detection model include: collecting long-term historical data from potential geological hazard points, including data such as displacement, tilt angle, and soil moisture content collected by ground sensor nodes, information such as surface deformation and vegetation cover obtained by low-orbit remote sensing satellites, and precise location data provided by satellite navigation systems, and then processing the data through data cleaning, standardization, and spatiotemporal alignment, resulting in historical multi-source data.

[0098] Feature extraction from historical multi-source data: For time-series data such as displacement and tilt angle, statistical features such as mean, variance, maximum, and minimum values ​​are extracted to reflect the overall distribution and fluctuation of the data. Time-series variation characteristics, such as rate of change and acceleration, are calculated to capture dynamic trends. For example, the rate of change of displacement reflects the movement speed of geological bodies and is important for judging the occurrence of disasters such as landslides. Fourier transform or wavelet transform methods are used to transform time-series data from the time domain to the frequency domain to extract frequency features. Different geological disasters may be accompanied by vibrations at specific frequencies; analyzing frequency features can better identify early signs of disasters. Data from ground-based sensor nodes, low-orbit remote sensing satellites, and satellite navigation systems are fused to extract comprehensive features. For example, combining displacement data and surface deformation data can more accurately determine the deformation of geological bodies; combining soil moisture data with vegetation cover data helps analyze the impact of factors such as rainfall on geological disasters. Principal component analysis (PCA) and other methods are used to reduce the dimensionality of the fused features, remove redundant information, improve the training efficiency and generalization ability of the model, and thus form multi-source feature data. Considering the limited computing resources and storage capacity of edge nodes, lightweight anomaly detection models are chosen, such as Isolation Forest and One-Class Support Vector Machine (SVM). Isolation Forest, based on a binary tree structure, isolates outliers by randomly partitioning the feature space, offering advantages such as fast training speed and low memory consumption. One-Class SVM separates normal and outlier data by finding an optimal hyperplane, making it suitable for single-class classification problems and effectively handling anomalies in geological disaster data.

[0099] Multi-source feature data is divided into training and test sets. The selected model is trained using the training set, and its parameters are adjusted, such as the number of trees in an isolated forest, the kernel function parameters in a support vector machine, and the penalty coefficient, to achieve good performance on the training set. The trained model is evaluated using the test set, and its performance is measured by metrics such as accuracy, recall, and F1 score.

[0100] The output of the trained lightweight anomaly detection model is the anomaly confidence score, which is divided into three levels: low (0-30%), medium (30%-70%), and high (70%-100%), to represent the probability that the current data deviates from the normal pattern.

[0101] When the temporal characteristics of the input parameters change little and fluctuate within the normal range, the model will output a low anomaly confidence score. For example, if the rate of change of parameters such as displacement and tilt angle is small, and parameters such as soil moisture content and surface deformation are in a stable state and have a high similarity to historical normal data, it indicates that the current geological environment is relatively stable and the possibility of geological disasters is low.

[0102] If the input parameters exhibit some anomalous changes, but these changes are insufficient to constitute a clear indication of a disaster, the model will output a medium confidence level for the anomalies. Examples include a slight increasing trend in displacement, minor fluctuations in tilt angle, and a slight increase in soil moisture content. These changes may be influenced by short-term climatic factors or localized geological activity and require further monitoring and analysis.

[0103] When the input parameters exhibit significant abnormal changes that are consistent with the characteristics of geological disasters, the model will output a high anomaly confidence score. Examples include a rapid increase in displacement within a short period, a significant change in tilt angle, soil moisture content exceeding a critical value, and large-scale abnormal surface deformation. These situations indicate that a geological disaster may be imminent or has already occurred, requiring immediate early warning and response measures.

[0104] When the anomaly confidence level output by an edge node is low, and the network status monitoring module confirms that the current communication link is stable (e.g., signal strength, bandwidth utilization, etc. meet preset thresholds), the edge node automatically initiates a collaborative sensing request to 3-5 neighboring edge nodes. Upon receiving the request, the neighboring nodes synchronously collect sensor data for the corresponding area and transmit it back to the initiating node. The initiating node then performs spatiotemporal alignment of its own data with that of its neighboring nodes and verifies the existence of an anomaly through a weighted average or voting mechanism. If the collaborative sensing result still shows an anomaly confidence level below the threshold, it is determined to be a false alarm, and the edge node only uploads the verification result to the cloud for archiving; if the collaborative sensing result shows an anomaly confidence level increasing, it is upgraded to a medium-level anomaly and proceeds to the next processing step.

[0105] When the anomaly confidence level is medium, the edge node first sends an aggregation request to its cluster head gateway (elected by multiple edge nodes within the region, responsible for coordinating data aggregation). Upon receiving the request, the cluster head gateway broadcasts a data collection instruction to 5-10 surrounding edge nodes, requesting each node to upload sensor data for a specified time period. The cluster head gateway performs correlation analysis on the collected multi-node data, including spatiotemporal feature matching and multi-parameter cross-validation, to eliminate noise interference from single-point data. After analysis, the cluster head gateway generates an aggregation report containing the anomaly area range, correlation parameters, and preliminary risk level, and uploads it to the cloud for in-depth analysis via low-Earth orbit satellite IoT. Simultaneously, the cluster head gateway issues adjusted sampling frequency instructions (e.g., from minute-level to second-level) to edge nodes within the region to strengthen monitoring of key areas.

[0106] When the anomaly confidence level is high, or when an edge node detects a network outage (such as the loss of three consecutive heartbeat packets), the edge node immediately initiates an emergency response procedure. First, the edge node compresses and encrypts local critical data (such as high-frequency sampling data from the last 10 minutes, and complete data segments before and after the anomaly event), using a lightweight symmetric encryption algorithm (such as AES-128) to ensure data security. Then, the edge node uploads the encrypted data packets to the cloud via the emergency communication link of the low-Earth orbit satellite IoT (prioritizing high-priority channels). Upon receiving the data, the cloud invokes a pre-trained global early warning model for deep inference, combining geological disaster physical equations with spatiotemporal neural networks to generate a detailed report including the comprehensive risk level, the contribution of disaster-causing factors, and dynamic early warning thresholds. If the network outage continues for more than a preset time, the cloud will send simplified early warning information to relevant personnel via backup communication methods such as BeiDou short messages to ensure that critical information is not lost.

[0107] After completing in-depth analysis, the cloud-based system dynamically adjusts the sampling frequency and task priority of edge nodes based on the overall geological disaster situation. For example, it issues a high-frequency command of "sampling once every 5 seconds" to edge nodes in high-risk areas, while adjusting to a low-frequency mode of "sampling once every 30 seconds" for nodes in low-risk areas. Simultaneously, it reallocates communication resources based on network load, prioritizing data transmission from high-level anomaly nodes. These adjustment commands are transmitted to edge nodes via low-Earth orbit satellite IoT, forming a closed-loop optimization mechanism between the edge and the cloud, continuously improving the adaptability and robustness of the monitoring and early warning system.

[0108] This embodiment realizes dynamic strategy selection based on anomaly level and network status, which effectively reduces the pressure on data transmission and cloud computing while ensuring real-time monitoring, and provides reliable technical support for early warning of geological disasters.

[0109] In some embodiments, step S104 above, which involves dynamically adjusting the edge node sampling frequency and task priority based on the global situation via the cloud, specifically includes: By aggregating monitoring data reported by edge nodes in various regions and local analysis results in the cloud, and taking the spatial coordinates of potential hazards, the time of risk generation, and the type of main disaster-causing factors as inputs, a space-time-spectrum variogram model is constructed. Kriging interpolation is then performed by combining the comprehensive risk level and the contribution of disaster-causing factors to generate a risk intensity map and an uncertainty spatial distribution map covering the entire region. Based on the risk intensity map, the spatial distribution map of uncertainty, and the geographical location, remaining energy and communication load of edge nodes, a constrained optimization model is constructed with the goal of maximizing monitoring benefits and minimizing system energy consumption. Solve the constrained optimization model to calculate the optimal target sampling frequency and dynamic task priority for each edge node; The target sampling frequency and dynamic task priority are encapsulated into control commands and sent to the corresponding edge nodes.

[0110] In this embodiment, the cloud first aggregates monitoring data reported by each edge node and local analysis results, as well as the comprehensive risk level and contribution of disaster-causing factors calculated by the cloud. Using this data as input, the cloud constructs a space-time-spectrum variogram model. This model quantifies spatial correlation by analyzing the variation characteristics of risk indicators in different regions and time periods. Subsequently, using the Kriging interpolation method, the discrete monitoring data is spatially continuousized to generate a risk intensity map covering the entire region. The value of each grid cell in the map represents the potential geological hazard risk value of that region. Simultaneously, an uncertainty spatial distribution map is generated, reflecting the confidence level of risk estimates for each region (e.g., the variance distribution of interpolation errors calculated using the variogram). These two maps provide the basic input for the subsequent optimization of the model, representing the global risk situation.

[0111] Based on the generated risk intensity map and uncertainty spatial distribution map, the cloud further integrates the real-time status information of edge nodes, including geographical location (e.g., whether it is located in a high-risk area), remaining energy (estimated by battery voltage or power percentage), and communication load (e.g., current data transmission volume and channel occupancy). Using these parameters as constraints, a constrained optimization model is constructed, with the core objectives being: (1) Maximize monitoring benefits: Prioritize the monitoring accuracy of high-risk areas and areas with high data uncertainty, achieved by increasing the sampling frequency or increasing the weight of analysis tasks; (2) Minimize system energy consumption: Reduce the task load of edge nodes in low-risk areas or with low remaining energy, and avoid premature node failure due to oversampling.

[0112] The constrained optimization model balances the two objectives through weighted summation or analytic hierarchy process, with the weights dynamically adjusted according to actual disaster emergency needs (e.g., focusing on monitoring benefits during the flood season and energy consumption control during the dry season).

[0113] The cloud-based system employs heuristic algorithms (such as genetic algorithms or particle swarm optimization) to solve the constrained optimization model. The algorithm treats edge nodes as individuals and uses a weighted sum of global monitoring benefits and energy consumption as the fitness function, iteratively searching for the optimal solution set. The solution results for each edge node include a target sampling frequency (e.g., once per second to once every 10 minutes) and dynamic task priorities (divided into high, medium, and low levels). High-priority tasks include high-frequency data acquisition and emergency anomaly reporting; medium-priority tasks include routine data aggregation and partial model updates; and low-priority tasks include low-frequency data backup and system status self-checks.

[0114] The cloud encapsulates the obtained target sampling frequency and task priority into standardized control commands. The command format includes node ID, target parameter value, effective time, and validity period. These commands are then distributed to the corresponding edge nodes via a reliable transmission channel of the low-Earth orbit satellite IoT (e.g., prioritizing high-bandwidth channels). Upon receiving the commands, the edge nodes immediately adjust the operating mode of their local sampling modules (e.g., shortening / extending the sampling interval) and update the task scheduling queue to ensure high-priority tasks are executed first. Simultaneously, the nodes report the parameter adjustment results (e.g., actual sampling frequency, changes in remaining energy) back to the cloud for continuous optimization of the control strategy.

[0115] This embodiment enables refined control of edge nodes based on global situation, which effectively extends the overall operating cycle of the system while ensuring the accuracy of geological disaster monitoring, and provides reliable technical support for disaster prevention.

[0116] Furthermore, the process involves aggregating monitoring data reported by edge nodes in various regions via the cloud and combining it with local analysis results. Using the spatial coordinates of potential hazard points, the time of risk generation, and the type of the main disaster-causing factor as input, a space-time-spectrum variogram model is constructed. Kriging interpolation is then performed, combining the comprehensive risk level and the contribution of the disaster-causing factor, to generate a risk intensity map and an uncertainty spatial distribution map covering the entire region. Specifically, this includes: By aggregating monitoring data from various edge nodes in the cloud and local analysis results, the spatial coordinates of potential hazards and the time of risk generation are extracted, and the main disaster-causing factor type is determined by combining the contribution of disaster-causing factors calculated in the cloud. Based on the main disaster-causing factor types, a spectral distance metric is defined to characterize the differences in disaster driving mechanisms, and a space-time-spectral variogram model is constructed by combining the spatial coordinates of hazard points and the risk generation time. Using the comprehensive risk level calculated in the cloud as the sample attribute value, the Kriging equation system is established and solved using the space-time-spectrum variogram model to obtain the optimal weight coefficients at the grid points to be interpolated. Based on the optimal weighting coefficient and the comprehensive risk level, the risk intensity estimate of each grid point to be interpolated is calculated, and the Kriging variance of the risk intensity estimate is calculated simultaneously as a quantitative indicator of cognitive uncertainty. The risk intensity estimates and Kriging variances of all grid points to be interpolated are visualized and rendered to generate a risk intensity map and an uncertainty spatial distribution map.

[0117] In this embodiment, the cloud first receives real-time monitoring data and local analysis results reported by edge nodes in various regions. The data includes the spatial coordinates of geological hazard points (such as latitude, longitude, and elevation), risk generation timestamps (accurate to the second), real-time monitoring values ​​of various disaster-causing factors (such as rainfall, surface displacement rate, soil moisture content, etc.), and preliminary assessment results of the comprehensive risk level calculated by edge nodes based on preset rules. The cloud further cleans and verifies the data, removes outliers, and combines a historical geological hazard database and an expert knowledge base to determine the main disaster-causing factor type for each hazard point through weighted voting or machine learning classification models.

[0118] To address the differences in the types of primary disaster-causing factors, a cloud-based spectral distance metric is defined to quantify the spatial correlation of different disaster-driving mechanisms. For example, for "rainfall-dominated" hazard points, the spectral distance can be calculated by combining the distribution characteristics of rainfall isopleths with the topographic slope factor; for "earthquake-dominated" hazard points, the spectral distance is defined based on the relationship between fault distance and magnitude attenuation. Subsequently, a three-dimensional input space is constructed using the spatial coordinates of the hazard point (two-dimensional planar location), the risk generation time (one-dimensional time axis), and the spectral distance (one-dimensional driving mechanism axis) to build a space-time-spectral variogram model. This model analyzes the spatiotemporal evolution of risk levels of hazard points of the same type in historical disaster events, fits the variogram curve, and quantifies the influence weights of different dimensional parameters on risk correlation.

[0119] Using the comprehensive risk level calculated in the cloud as the sample attribute value, a space-time-spectral variogram model is employed to describe the spatial autocorrelation between sample points. For the grid points to be interpolated, a system of Kriging equations is established in the cloud, with the number of equations equal to the number of known sample points. Each equation includes a variogram term and a Lagrange multiplier term. The system of equations is solved by minimizing the estimation variance to obtain the optimal weight coefficient for each grid point to be interpolated. The weight coefficient reflects the contribution of known sample points to the risk estimation of the target grid point; for example, sample points that are close to each other and have similar spectral distances have higher weights.

[0120] Based on the optimal weighting coefficients obtained from the solution, the cloud calculates the estimated risk intensity for each grid point to be interpolated, specifically the weighted sum of the comprehensive risk level of each known sample point and its corresponding weighting coefficient. Simultaneously, the kriging variance of this estimate is calculated as a quantitative indicator of cognitive uncertainty. The kriging variance is derived through a variogram model and the weighting coefficients; a smaller value indicates a more reliable estimation result.

[0121] The cloud-based system maps the risk intensity estimates and Kriging variances of all grid points to Geographic Information System (GIS) layers. A risk intensity map is generated using color gradients and contour overlays (e.g., red indicates high-risk areas, blue indicates low-risk areas). An uncertainty spatial distribution map is generated using transparency overlays or heatmaps (e.g., high-uncertainty areas are displayed as semi-transparent patches). The two layers are overlaid with coordinate alignment, allowing users to zoom in and out to view local details via an interactive interface, or overlay basic geographic data such as terrain and roads to aid analysis. The final risk intensity map and uncertainty spatial distribution map are stored in a standardized format (e.g., GeoTIFF) and pushed in real-time to the edge node control module and early warning decision module, providing a basis for dynamically adjusting monitoring strategies.

[0122] This embodiment realizes a refined assessment and visualization of geological disaster risks across the entire region, overcoming the shortcomings of traditional methods such as isolated data and insufficient quantification of uncertainty, and providing key technical support for the efficient operation of the low-orbit satellite Internet of Things monitoring and early warning system.

[0123] Furthermore, based on the risk intensity map, the uncertainty spatial distribution map, and the geographical location, remaining energy, and communication load of edge nodes, a constrained optimization model is constructed with the objective of maximizing monitoring benefits and minimizing system energy consumption. Specifically, this includes: Collect the geographical location, remaining energy estimate, and real-time communication link load of each edge node; Based on the risk intensity map and the uncertainty spatial distribution map, the risk intensity value and uncertainty value corresponding to the geographical location of each edge node are obtained; Using the sampling frequency and task priority of each edge node as decision variables, and the risk intensity value and uncertainty value as key parameters, a benefit function is set to reflect the gain of monitoring information. Construct a cost function that reflects the total energy consumption and communication congestion of the system. The value of the cost function is positively correlated with the sampling frequency, task priority, and real-time communication link load of all edge nodes. With the optimization objective of maximizing the benefit function and minimizing the cost function, a constrained optimization model is established, taking the estimated residual energy of each edge node, the load of the real-time communication link, and the physical feasible range of the decision variables as constraints.

[0124] In this embodiment, the cloud uses the communication protocol of a low-Earth orbit satellite Internet of Things (LEO) to periodically collect status data from each edge node, including geographical location, estimated remaining energy, and real-time communication link load. The geographical location is obtained in real-time through a satellite navigation system, accurate to the meter level; the estimated remaining energy is calculated based on a model of node battery voltage and current consumption; and the real-time communication link load is calculated as bandwidth utilization by statistically analyzing the number and size of data packets transmitted per unit time. All parameters are encrypted and transmitted to the cloud database, and timestamped to ensure data timeliness.

[0125] The cloud loads pre-generated risk intensity maps and uncertainty spatial distribution maps. These two maps are based on a geographic coordinate system and store the risk intensity value (range 0-1, with higher values ​​indicating higher risk) and uncertainty value (range 0-1, with higher values ​​indicating lower estimation reliability) of each grid point in a rasterized format. For the geographical location of each edge node, the cloud extracts the risk intensity value and uncertainty value of the corresponding grid point from the map as key parameters. For example, edge nodes located in high-risk-high-uncertainty areas require priority attention.

[0126] Using the sampling frequency (times / minute) and task priority (dimensionless value, range 1-10, with higher values ​​indicating higher priority) of each edge node as decision variables, a benefit function is constructed to reflect the monitoring information gain. The benefit function consists of two parts: the first part is the risk coverage benefit, which is positively correlated with the risk intensity of the area where the edge node is located; the higher the risk, the more significant the benefit increase from increasing the sampling frequency. The second part is the uncertainty elimination benefit, which is positively correlated with the uncertainty value; increasing the sampling frequency can reduce uncertainty, thereby improving the benefit.

[0127] A cost function is constructed to quantify the total energy consumption and communication congestion level of the system. Its value is positively correlated with the sampling frequency, task priority, and real-time communication link load of all edge nodes. The cost function consists of two terms: the first term is the energy consumption cost, which is linearly related to the sampling frequency; the higher the sampling frequency, the greater the node energy consumption. The second term is the congestion penalty term; when the communication load exceeds a threshold (such as 80% bandwidth utilization), the cost function value increases exponentially to suppress excessive congestion.

[0128] The optimization objective is to maximize the weighted sum of the benefit function and minimize the cost function, with the weighting coefficients adjusted according to actual needs. Four types of constraints are set: the first is an energy constraint, requiring that the estimated remaining energy of each edge node, given the combination of sampling frequency and task priority, must be greater than a minimum safety threshold; the second is a communication constraint, requiring that the total communication load of all nodes must not exceed the system's maximum bandwidth capacity; the third is a physical feasibility constraint, requiring that the sampling frequency be within the hardware's supported range and that task priorities be integers between 1 and 10; and the fourth is a spatial balance constraint, requiring that the difference in task priorities between adjacent edge nodes not exceed three levels to avoid monitoring blind spots. By integrating the objective function and constraints, a complete constrained optimization model is established.

[0129] Taking a landslide disaster scenario as an example, the constrained optimization model achieves dynamic resource allocation through the following specific process: First, the sampling frequency (unit: times / minute) and task priority (integer level 1-10) of edge nodes are defined as decision variables. Considering the differences in geological conditions and deformation characteristics in different areas of the landslide, such as the need for more intensive monitoring in weak interlayer areas, the sampling frequency and task priority are used as adjustable parameters to flexibly allocate resources according to the actual situation.

[0130] Construct a benefit function that includes risk coverage benefits (positively correlated with the risk intensity value of the landslide area) and uncertainty elimination benefits (positively correlated with the risk uncertainty value), and a cost function that includes system energy consumption (positively correlated with sampling frequency) and communication congestion costs (positively correlated with task priority and real-time communication load); set four types of constraints: energy constraints require that the remaining energy of the nodes is not lower than the safety threshold (e.g., 30% remaining power), communication constraints ensure that the total load of all nodes does not exceed the maximum bandwidth capacity of the system (e.g., 1Gbps), physical feasibility constraints limit the sampling frequency to the range supported by the hardware (e.g., 4-10 times / minute) and the priority is an integer, and spatial balance constraints require that the priority difference between adjacent nodes does not exceed 3 levels to avoid monitoring blind spots.

[0131] When constructing the benefit function, regarding the risk coverage benefit, detailed geological surveys and stability analyses of the landslide body are conducted to divide areas into different risk levels. Higher risk intensity values ​​indicate a greater likelihood and severity of landslides in that area, thus requiring higher monitoring frequency and priority to ensure timely detection of disaster signs. Regarding the uncertainty reduction benefit, the uncertainty of landslide disasters stems from the complexity of geological conditions and the unpredictability of external triggering factors. Increasing monitoring of areas with higher uncertainty can obtain more data, reduce the uncertainty in understanding the landslide status, and improve the accuracy of early warnings.

[0132] A genetic algorithm is used to solve the model. Through steps such as population initialization, fitness calculation, selection of superior individuals, crossover and mutation, and iterative optimization, the optimal solution satisfying the constraints is found. When the landslide risk level increases, the sampling frequency of nodes in high-risk areas is automatically increased from 5 times / minute to 8 times / minute, and the task priority is adjusted from level 6 to level 9. Based on the landslide risk assessment results, the monitoring strategy is adjusted in a timely manner, focusing on high-risk areas and increasing the frequency and priority of data collection to more accurately grasp the dynamic changes of landslides. Simultaneously, a dynamic feedback mechanism adjusts the configuration in real time based on the remaining energy of the nodes (e.g., reducing the sampling frequency to 4 times / minute when it drops to 20%) and communication load. During actual operation, the power and communication status of the nodes are continuously monitored. When abnormal situations occur, the sampling frequency and task priority of the nodes are adjusted in a timely manner to ensure the stable operation of the system. For example, when the node power is low, its sampling frequency is reduced to reduce energy consumption; when the communication load is high, data transmission of high-priority tasks is prioritized, and the transmission frequency of low-priority tasks is appropriately reduced.

[0133] Ultimately, through the dynamic resource allocation process of the aforementioned constrained optimization model, the dual objectives of maximizing monitoring benefits and minimizing system energy consumption are achieved, thereby improving the timeliness and accuracy of landslide early warning and effectively reducing the losses caused by landslide disasters.

[0134] This embodiment achieves a dynamic balance between monitoring benefits and system energy consumption, overcoming the shortcomings of blind control strategies and uneven resource allocation in traditional methods, and providing quantitative decision support for the efficient operation of low-orbit satellite Internet of Things monitoring systems.

[0135] In one possible implementation, the benefit function aims to quantify the "information gain" achievable by configuring the sampling frequency and task priorities of edge nodes. Its core design principle guides the system to prioritize resource allocation to high-risk and high-uncertainty areas. The benefit function formula is as follows: Where U represents the total monitoring benefit of the system; the larger the value, the higher the information gain brought about by resource allocation. K represents the total number of edge nodes, and k represents the edge node index. This represents the sampling frequency of edge node k. This represents the task priority weight of edge node k. The risk intensity value at the location of edge node k is obtained from the risk intensity map. The risk uncertainty value at the location of edge node k is obtained from the uncertainty space distribution map. This represents the risk intensity benefit weighting coefficient. This represents the sampling frequency benefit adjustment coefficient. This represents the weighting coefficient for uncertainty benefits. This indicates the maximum risk intensity. Indicates the maximum sampling frequency supported by the hardware. The maximum value representing uncertainty. This represents the maximum value of the task priority weight.

[0136] This is used to drive the system to allocate higher sampling frequencies to high-risk edge nodes. This is used to drive the system to allocate higher communication and processing priorities to edge nodes with high cognitive uncertainty, and to prioritize the acquisition and processing of data from these edge nodes, which can most effectively reduce the global cognitive blind spot.

[0137] In one possible implementation, the cost function is used to quantify the system resources consumed and the negative effects of increasing the sampling frequency and task priority, mainly including energy consumption and communication congestion. The cost function formula is as follows: Where C represents the total system cost; the larger the value, the more resources are consumed in resource allocation, and the heavier the system burden. This represents the energy consumption weighting coefficient for sampling frequency. This represents the energy consumption weighting coefficient for task priority. This represents the communication congestion cost weighting coefficient. This represents the benchmark coefficient for unit energy consumption. This represents the real-time load of the communication link where edge node k is located. This indicates the maximum real-time load of the communication link.

[0138] Energy consumption cost is used to directly calculate the adjustment cost of all edge nodes. and The estimated energy consumption generated is the main physical cost of operating the system. As a soft penalty for communication congestion costs, if edge nodes are given high priority when their communication links are already busy, it will exacerbate channel contention and increase the risk of data collisions and delays.

[0139] In some embodiments, in steps S101 to S105 above, the method further includes: The edge nodes of geological disaster hazard points in various places are used as local clients. The monitoring data accumulated by the local clients and the corresponding historical disaster tags are used to independently train a lightweight early warning threshold generation model locally and generate local model parameters. The local model parameters are uploaded to the cloud, which serves as a federated server, to calculate the ability of the local model parameters to represent spatial correlation and generate local confidence weights. Based on the federated server, the local model parameters and corresponding confidence weights uploaded by each local client are encrypted and aggregated using a weighted average algorithm to generate a global early warning model. A virtual global risk characterization field is generated based on the basic geological and environmental parameters of each region stored in the cloud, and the global early warning model is subjected to knowledge distillation and regularization constraints. The global early warning model, after knowledge distillation and regularization constraints, is distributed to all clients, enabling local clients to adjust the global early warning model based on local geological environment parameters and generate localized dynamic early warning thresholds.

[0140] In this embodiment, edge nodes deployed at various geological disaster hazard points serve as local clients. Each client independently trains a lightweight early warning threshold generation model locally using its accumulated monitoring data (such as time-series data from ground sensors and feature values ​​from satellite remote sensing images) and historical disaster labels (such as binary labels indicating whether landslides or debris flows have occurred). This model employs a shallow neural network structure, with multi-source monitoring data features as input and disaster occurrence probability values ​​as output. During training, the client optimizes the model parameters using gradient descent until the model's accuracy on the local validation set reaches a preset threshold, ultimately generating a local model parameter file containing the weights and biases of each neuron.

[0141] Each local client encrypts its trained model parameters and uploads them to the cloud, which acts as a federated server. The cloud then calculates the spatial relevance representation of each client's model parameters: first, it extracts key dimensions related to geological features (such as weights corresponding to terrain slope); then, it combines spatial attribute data such as historical disaster frequency and geological structural complexity of the client's region; finally, it uses a logistic regression model to evaluate the explanatory power of the client's model parameters for regional risk patterns, generating local confidence weights (ranging from 0 to 1, with higher values ​​indicating higher parameter reliability). For example, clients located near fault zones and experiencing frequent historical disasters will have higher confidence weights assigned to their model parameters.

[0142] The cloud-based system acts as a federated server, collecting all encrypted local model parameters and corresponding confidence weights uploaded by clients. It then aggregates these parameters using a weighted average algorithm: for each parameter dimension (e.g., the first weight value of a neuron in the first layer), all client parameter values ​​for that dimension are multiplied by their confidence weights, summed, and then divided by the sum of all weights to obtain the global parameter value for that dimension. This process is repeated until all parameter dimensions are aggregated, generating a global early warning model. This model has the same structure as the local model, but its parameters incorporate geological features from multiple regions, enabling cross-regional risk prediction.

[0143] Based on stored fundamental geological parameters (such as lithology and soil moisture) and environmental parameters (such as rainfall and seismic intensity) for various regions, a virtual global risk characterization field is constructed in the cloud. The geographic space is divided into 1km × 1km grids, and each grid calculates a risk index (range 0-100) based on the fundamental parameters, forming a risk distribution map covering the entire region. This risk characterization field is used as a teacher model to perform knowledge distillation on the global early warning model: by minimizing the mean squared error between the global early warning model output and the risk index of the corresponding grid in the risk characterization field, the global model is forced to learn fundamental geological patterns. Simultaneously, an L2 regularization term is added to the loss function to constrain the variation of model parameters and prevent overfitting to local data. For example, if the weight of rainfall in the client data for a certain region is abnormally high, the regularization term will pull it back to a reasonable range.

[0144] The cloud distributes the global early warning model parameters, refined through knowledge distillation and regularization, to all local clients. Upon receiving these parameters, each client, combined with its stored local geological environment parameters (such as soil type and vegetation cover at specific locations), fine-tunes the global model: adding a local feature branch to the model input layer, inputting the concatenated local geological parameters and original monitoring data into the model, and performing incremental training with a small amount of local data to adjust the parameters of the last two layers, making the output more closely reflect the actual local risk situation. Ultimately, a localized dynamic early warning threshold is generated. This embodiment achieves collaborative optimization of "global model sharing knowledge - local model adapting to the environment", overcoming the problems of poor model adaptability and high risk of privacy leakage in traditional methods, and significantly improving the accuracy and reliability of geological disaster early warning.

[0145] Reference Figure 2 The present invention also provides a computer device 2, including: a memory 202 and a processor 201, and a computer program 203 stored in the memory 202. When the computer program 203 is executed on the processor 201, it implements the geological disaster monitoring and early warning method based on low-orbit satellite Internet of Things as described in any of the above methods.

[0146] The computer device 2 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 2 may include, but is not limited to, a processor 201 and a memory 202. Those skilled in the art will understand that... Figure 2 The computer device 2 is merely an example and does not constitute a limitation on the computer device 2. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0147] The processor 201 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0148] In some embodiments, the memory 202 may be an internal storage unit of the computer device 2, such as a hard disk or memory. In other embodiments, the memory 202 may be an external storage device of the computer device 2, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, the memory 202 may include both internal and external storage units of the computer device 2. The memory 202 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 202 can also be used to temporarily store data that has been output or will be output.

[0149] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

Claims

1. A geological disaster monitoring and early warning method based on a low-orbit satellite Internet of Things, characterized in that, The method specifically includes: Real-time collection of multi-source data on potential geological disaster sites is achieved through low-orbit remote sensing satellites, ground-based sensing nodes, and satellite navigation systems. Multi-source data is spatiotemporally aligned and interpolated, and mapped onto a three-dimensional spatiotemporal grid centered on geological hazard points to form a multi-dimensional data cube. Mechanism constraints are constructed using physical equations of geological hazards, and spatiotemporal graph neural networks are combined to extract spatiotemporal correlation features from multidimensional data cubes. Through mechanism-guided attention mechanisms, risk evolution patterns are deeply mined, and comprehensive risk levels and the contribution of disaster-causing factors are output. After performing lightweight anomaly detection at the edge nodes, the system dynamically selects collaborative perception, local aggregation, or cloud-based deep analysis strategies based on the anomaly level and network status. The system also dynamically adjusts the sampling frequency and task priority of the edge nodes based on the global situation through the cloud. Key information for early warning decisions is generated and stored on the blockchain. When the overall risk level is determined to exceed a preset risk threshold, a multimodal communication link is triggered simultaneously to adaptively select the optimal path to release standardized early warning information.

2. The method of claim 1, wherein, The process of performing spatiotemporal alignment and interpolation on multi-source data and mapping it onto a three-dimensional spatiotemporal raster centered on geological hazard hazard points to form a multidimensional data cube specifically includes: Based on a unified spatial coordinate system and time reference standard, multi-source data is transformed to a unified spatiotemporal benchmark through coordinate transformation and time synchronization protocols. For remote sensing images, ground sensor time-series data and satellite positioning data in multi-source data, spatiotemporal alignment processing is performed using geometric correction, resampling interpolation and positioning techniques, respectively. The three-dimensional spatial range and raster resolution are defined with geological hazard hazard points as the core, and the time slices are divided at equal intervals along the time dimension to construct a spatiotemporal raster sequence; By using spatial allocation algorithms and attribute interpolation methods, multi-source data that has undergone spatiotemporal alignment is mapped to the corresponding cells of a spatiotemporal raster sequence, generating a three-dimensional attribute field for each time slice. The three-dimensional attribute fields of each time slice are stacked to form a multidimensional data cube, and each raster cell of the multidimensional data cube is assigned a unique spatiotemporal identifier and associated with multidimensional attributes.

3. The method of claim 2, wherein, The process involves mapping the spatiotemporally aligned multi-source data to corresponding cells of a spatiotemporal raster sequence using spatial allocation algorithms and attribute interpolation methods, generating a three-dimensional attribute field for each time slice. Specifically, this includes: Determine the raster cell to be mapped in the spatiotemporal raster sequence and its corresponding time slice; Using a spatiotemporal kriging interpolation model, based on the spatiotemporal coordinates of the grid cell, the spatiotemporal variability function value between the grid cell and all spatiotemporally aligned point monitoring data is calculated. Based on the spatiotemporal variability function value, the Kriging equation system is constructed and solved to obtain the interpolation weight of each point monitoring data. The attribute values ​​of the point monitoring data are then interpolated and weighted and fused using the interpolation weight. For the spatiotemporally aligned area monitoring data, the spatial overlap between the area monitoring data and the grid cells is calculated, and the spatial overlap is used as a weight to perform spatial weighted fusion of the attribute values ​​of the area monitoring data. The interpolation weighted fusion results of point monitoring data and the spatial allocation weighted fusion results of area monitoring data are integrated to generate the fusion attribute values ​​of raster cells, and all raster cells are traversed to form the three-dimensional attribute field of the corresponding time slice.

4. The method according to claim 1, characterized in that, The method utilizes physical equations of geological hazards to construct mechanistic constraints, and combines spatiotemporal graph neural networks to extract spatiotemporal correlation features from multidimensional data cubes. Through a mechanism-guided attention mechanism, it achieves in-depth mining of risk evolution patterns and outputs a comprehensive risk level and the contribution of disaster-causing factors. Specifically, this includes: Extract all grid cells of the target geological hazard hazard body from the multidimensional data cube, and calculate the physical state index of each grid cell based on the geological hazard physical mechanism model; Using each grid cell as a node, a spatiotemporal graph neural network is constructed based on the correlation between its spatial adjacency and physical state indicators; The physical state index is transformed into prior constraints of nodes and edges. These prior constraints are then injected into the attention weight calculation mechanism of the spatiotemporal graph neural network through a learnable mapping function, forming a mechanism-guided attention module. The mechanism-guided attention module performs multi-layer message passing and aggregation on the multi-dimensional temporal attributes of nodes to update the node's own state and obtain the updated node. Based on the updated nodes, the comprehensive risk level of the target geological hazard body and the contribution of each disaster-causing factor are output through classification and regression heads, respectively.

5. The method according to claim 4, characterized in that, The calculation of the physical state indices of each grid cell based on the geological disaster physical mechanism model specifically includes: Based on the target geological hazard type, the corresponding physical control equations are matched from a pre-built physical equation library, and the physical control equations are spatially discretized to be suitable for grid cell calculation. The parameter fields of each grid cell are extracted from the multidimensional data cube, and the parameter fields include geomechanical parameters, hydrological parameters and load parameters; Based on the discretized physical control equations, the parameter fields of all grid cells are solved simultaneously using a parallel computing architecture of a graphics processor to obtain the physical state index field.

6. The method according to claim 4, characterized in that, The mechanism-guided attention module performs multi-layer message passing and aggregation on the multi-dimensional temporal attributes of nodes to update the node's own state and obtain the updated node, specifically including: Based on the time-series data of the physical state indicators of each node, the spatiotemporal correlation between any two nodes is calculated, and the spatiotemporal correlation is normalized into physical guidance weights. In the mechanism-guided attention module, the physical guidance weight is used as a bias term in the attention score calculation to adjust the attention weight distribution of message passing between nodes and obtain the adjusted attention coefficient. The multi-dimensional temporal attribute features of nodes are weighted and aggregated using the adjusted attention coefficients, graph convolution feature updates are performed, and temporal convolution is connected after each graph convolution layer to capture the temporal evolution pattern of node features. Repeated multi-layer graph convolution and temporal convolution operations allow node features to continuously integrate neighborhood information and temporal dynamics.

7. The method according to claim 1, characterized in that, The method of dynamically adjusting the sampling frequency and task priority of edge nodes based on the global situation via the cloud specifically includes: By aggregating monitoring data reported by edge nodes in various regions and local analysis results in the cloud, and taking the spatial coordinates of potential hazards, the time of risk generation, and the type of main disaster-causing factors as inputs, a space-time-spectrum variogram model is constructed. Kriging interpolation is then performed by combining the comprehensive risk level and the contribution of disaster-causing factors to generate a risk intensity map and an uncertainty spatial distribution map covering the entire region. Based on the risk intensity map, the spatial distribution map of uncertainty, and the geographical location, remaining energy and communication load of edge nodes, a constrained optimization model is constructed with the goal of maximizing monitoring benefits and minimizing system energy consumption. Solve the constrained optimization model to calculate the optimal target sampling frequency and dynamic task priority for each edge node; The target sampling frequency and dynamic task priority are encapsulated into control commands and sent to the corresponding edge nodes.

8. The method according to claim 7, characterized in that, The process involves aggregating monitoring data reported by edge nodes in various regions via the cloud and combining it with local analysis results. Using the spatial coordinates of potential hazard points, the time of risk generation, and the type of the main disaster-causing factor as input, a space-time-spectrum variogram model is constructed. Kriging interpolation is then performed, combining the comprehensive risk level and the contribution of disaster-causing factors, to generate a risk intensity map and an uncertainty spatial distribution map covering the entire region. Specifically, this includes: By aggregating monitoring data from various edge nodes in the cloud and local analysis results, the spatial coordinates of potential hazards and the time of risk generation are extracted, and the main disaster-causing factor type is determined by combining the contribution of disaster-causing factors calculated in the cloud. Based on the main disaster-causing factor types, a spectral distance metric is defined to characterize the differences in disaster driving mechanisms, and a space-time-spectral variogram model is constructed by combining the spatial coordinates of hazard points and the risk generation time. Using the comprehensive risk level calculated in the cloud as the sample attribute value, the Kriging equation system is established and solved using the space-time-spectrum variogram model to obtain the optimal weight coefficients at the grid points to be interpolated. Based on the optimal weighting coefficient and the comprehensive risk level, the risk intensity estimate of each grid point to be interpolated is calculated, and the Kriging variance of the risk intensity estimate is calculated simultaneously as a quantitative indicator of cognitive uncertainty. The risk intensity estimates and Kriging variances of all grid points to be interpolated are visualized and rendered to generate a risk intensity map and an uncertainty spatial distribution map.

9. The method according to claim 7, characterized in that, The constrained optimization model, based on the risk intensity map, uncertainty spatial distribution map, and the geographical location, remaining energy, and communication load of edge nodes, aims to maximize monitoring benefits and minimize system energy consumption. Specifically, it includes: Collect the geographical location, remaining energy estimate, and real-time communication link load of each edge node; Based on the risk intensity map and the uncertainty spatial distribution map, the risk intensity value and uncertainty value corresponding to the geographical location of each edge node are obtained; Using the sampling frequency and task priority of each edge node as decision variables, and the risk intensity value and uncertainty value as key parameters, a benefit function is set to reflect the gain of monitoring information. Construct a cost function that reflects the total energy consumption and communication congestion of the system. The value of the cost function is positively correlated with the sampling frequency, task priority, and real-time communication link load of all edge nodes. With the optimization objective of maximizing the benefit function and minimizing the cost function, a constrained optimization model is established, taking the estimated residual energy of each edge node, the load of the real-time communication link, and the physical feasible range of the decision variables as constraints.

10. The method according to any one of claims 1 to 9, characterized in that, The method further includes: The edge nodes of geological disaster hazard points in various places are used as local clients. The monitoring data accumulated by the local clients and the corresponding historical disaster tags are used to independently train a lightweight early warning threshold generation model locally and generate local model parameters. The local model parameters are uploaded to the cloud, which serves as a federated server, to calculate the ability of the local model parameters to represent spatial correlation and generate local confidence weights. Based on the federated server, the local model parameters and corresponding confidence weights uploaded by each local client are encrypted and aggregated using a weighted average algorithm to generate a global early warning model. A virtual global risk characterization field is generated based on the basic geological and environmental parameters of each region stored in the cloud, and the global early warning model is subjected to knowledge distillation and regularization constraints. The global early warning model, after knowledge distillation and regularization constraints, is distributed to all clients, enabling local clients to adjust the global early warning model based on local geological environment parameters and generate localized dynamic early warning thresholds.

Images