An automatic geological hazard identification system and method based on multi-source remote sensing data

Abstract

This invention discloses an automatic geological hazard identification system and method based on multi-source remote sensing data, specifically relating to the field of geological hazard identification. The system includes a multi-modal remote sensing data acquisition module, a cross-domain physical fusion module, a spatiotemporal evolution decision module, a multi-cascade early warning decision module, an optimization control module, and a visualization module. The system and method utilize a domain adversarial network to generate virtual features through the cross-domain physical fusion module, improving the model generalization ability for cross-regional applications. Simultaneously, it deeply explores the physical correlations between multi-source remote sensing data, eliminating reliance on manually designed rules. Through a three-layer processing chain technology consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer, and a critical identification layer, it effectively improves the ability to capture hazard features in complex geological environments, particularly enhancing the accuracy of identifying precursory minor deformations and reducing the risk of missed detections.

Description

Technical Field

[0001] This invention relates to the field of geological hazard identification technology, and more specifically, to an automatic geological hazard identification system and method based on multi-source remote sensing data. Background Technology

[0002] With the rapid development of remote sensing technology and the increasing demand for geological disaster monitoring, the need for geological disaster identification using remote sensing data is becoming increasingly prominent. Traditional geological disaster identification methods mainly rely on manual interpretation of single-source remote sensing images, which consist of satellite remote sensing images, UAV aerial data, and ground sensors. By manually interpreting the texture features of visible light images, signs of geological disasters are identified, and finally, a comprehensive risk assessment is conducted. However, traditional methods heavily depend on the experience of experts, are inefficient, and are easily affected by subjective factors, making it difficult to achieve large-scale and rapid geological disaster identification, especially in complex terrain and vegetated areas, where the identification accuracy will significantly decrease.

[0003] To address the shortcomings of traditional methods in terms of efficiency and accuracy, existing technologies utilize end-to-end convolutional neural network models, image processing algorithms, and machine learning to automate the extraction of geological hazard features, while employing parallel computing architectures to improve processing speed.

[0004] However, in practical use, it still has some shortcomings. For example, the generalization ability of existing models is limited when applied across regions. When there are significant differences between the geological structure and environmental background of the target area and the training dataset, the model's recognition accuracy will drop significantly. At the same time, most current multimodal data fusion strategies remain at the feature level and fail to deeply explore the physical correlation between different remote sensing data sources. Moreover, the feature extraction process often relies on manually designed rules, making it difficult to comprehensively capture the disaster characteristics in complex geological environments. In particular, the recognition accuracy is not high for precursory small deformations, and there is a risk of missed reports. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an automatic geological disaster identification system and method based on multi-source remote sensing data, which solves the problems mentioned in the background art through the following scheme.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An automatic geological hazard identification system based on multi-source remote sensing data includes a system operation database, a system central processing module, and a user information terminal, and also includes:

[0008] Multimodal remote sensing data acquisition module: used to receive and perform spatial registration and radiometric correction on the first remote sensing data to generate the first remote sensing features;

[0009] Cross-domain physical fusion module: used to process the first remote sensing feature through a cross-domain physical fusion mechanism and output the second remote sensing feature. The cross-domain physical fusion mechanism includes a domain adversarial network mechanism and a physical constraint mechanism.

[0010] Spatiotemporal evolution decision module: used to obtain a third remote sensing feature from the second remote sensing feature through an automatic recognition model, wherein the automatic recognition model is a three-layer processing chain consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer and a critical recognition layer;

[0011] Multi-level early warning decision module: Based on the aforementioned third remote sensing features, a geological disaster assessment matrix is ​​constructed, and graded early warning signals are generated;

[0012] Optimization control module: used to receive the first remote sensing features of the target area in real time and perform optimization control operations, the optimization control operations being used to optimize and adjust the cross-domain physical fusion module and the spatiotemporal evolution decision module;

[0013] Visualization module: used to map the graded early warning signals into a three-dimensional geological model map, displaying the explanation and analysis of the causes of geological disasters;

[0014] The system operation database includes all data text of the system operation and collects information text output by each module in real time. The system central processing module is used to control the information text instructions output by each module in the central control system. The user information terminal is a device for receiving information output from the system.

[0015] Preferably, in the multimodal remote sensing data acquisition module, the first remote sensing data includes satellite multispectral images, InSAR radar interferometry data, and lidar point cloud data. The satellite multispectral images include multispectral data in the visible light band, near-infrared band, and short-wave infrared band. The InSAR radar interferometry data consists of master-slave image pairs. The lidar point cloud data includes three-dimensional coordinates, laser echo intensity, and the time difference between the first and last echoes.

[0016] Preferably, the cross-domain physical fusion module, by using the cross-domain physical fusion mechanism to process the first remote sensing feature, specifically includes:

[0017] A deep domain adversarial network consisting of a feature extractor, a label classifier, and a domain classifier is used to generate virtual features corresponding to the first remote sensing features.

[0018] A physical correlation map is constructed between the first remote sensing features and the corresponding geological parameters of the target area, and constitutive relationships are mined through graph neural networks;

[0019] The physical correlation diagram is based on prior geological knowledge and uses a Bayesian network to model the conditional dependencies between the first remote sensing features.

[0020] The second remote sensing feature includes the target area enhanced domain-adaptive feature vector and the physical correlation graph structure.

[0021] Preferably, the spatiotemporal evolution decision module processes the second remote sensing feature through the automatic identification model, wherein the three-layer processing chain of the automatic identification model satisfies:

[0022] The size of the 3D dilated convolution kernel in the spatiotemporal feature extraction layer is positively correlated with the node degree of the physical correlation graph;

[0023] The dimension of the node state vector of the disaster body spatiotemporal graph corresponding to the dynamic graph evolution layer is equal to the dimension of the hidden layer of the LSTM unit in the critical recognition layer;

[0024] The output gradient of the critical recognition layer is backpropagated to the dilation rate adjustment coefficient of the spatiotemporal feature extraction layer.

[0025] Preferably, in the spatiotemporal evolution decision module, the weight matrix WM corresponding to the three-dimensional dilated convolution kernel... ck The degree D of the nodes in the physical association graph is dynamically calculated, specifically as follows:

[0026]

[0027] Among them, D i Let ∑D represent the degree of node i in the physical association graph. j It is represented as the sum of the degrees of all nodes in the physical association graph. This is represented by the normalization of node degree using the Softmax function, and WM0 represents the weight matrix corresponding to the pre-set 3D dilated convolution kernel.

[0028] Preferably, the spatiotemporal evolution decision module calculates the priority of deformation propagation paths in the bimodal graph attention network using a structural attention head, specifically as follows:

[0029]

[0030] Where, p (ti,tj) Let represent the priority of the deformation propagation path from node ti to node tj, and let p represent the learnable weight vector that measures the similarity of node attributes corresponding to the spatiotemporal graph structure of the disaster body. T This is represented as the transpose operation on p, Wh ti and Wh tj Represented as the embedding representation of nodes ti and tj, [Wh i ‖Wh j ] represents the concatenated vector of the embedding representations of nodes ti and tj, LeakyReLU (p T [Wh i ‖Wh j]) represents p T [Wh i ‖Wh j Apply the LeakyReLU activation function, exp(LeakyReLU(p T [Wh ti ‖Wh tj ])) represents LeakyReLU(p T [Wh i ‖Wh j The activation result is applied using an exponential function, where tk represents all neighboring nodes of node ti.

[0031] Preferably, the operation method of the spatiotemporal evolution decision module and the critical identification layer specifically includes:

[0032] The cumulative effect characteristics of the deformation rate of the disaster body are extracted from the temporal sequence of the node states in the spatiotemporal diagram of the disaster body;

[0033] Mutual change collaboration features are captured by flattening the adjacency matrix and node features of the spatiotemporal graph of the disaster body;

[0034] The fusion weights are automatically adjusted based on the warning level at time t-1.

[0035] Preferably, the multi-level early warning decision module uses a disaster risk level instruction generated based on the quantitative analysis results of a probability risk assessment matrix as the graded early warning signal. The disaster risk level instruction includes four early warning levels: Level I, Level II, Level III, and Level IV. Level I indicates that a geological disaster is about to occur or has already shown initial destructive characteristics. Level II indicates that the disaster body has entered a rapid deformation stage and there is a possibility of short-term instability. Level III indicates that the geological disaster risk level has increased significantly and a local area has entered a deformation accumulation stage. Level IV indicates that there is a potential risk in the target area.

[0036] To achieve the above objectives, the present invention provides the following technical solution: an automatic geological hazard identification method based on multi-source remote sensing data, comprising implementing the aforementioned automatic geological hazard identification system based on multi-source remote sensing data, including:

[0037] S1: Receive and perform spatial registration and radiometric correction on the first remote sensing data to generate the first remote sensing feature;

[0038] S2: The first remote sensing feature is processed through a cross-domain physical fusion mechanism, and a second remote sensing feature is output. The cross-domain physical fusion mechanism includes a domain adversarial network mechanism and a physical constraint mechanism.

[0039] S201: A deep domain adversarial network consisting of a feature extractor, a label classifier, and a domain classifier generates virtual features corresponding to the first remote sensing feature and outputs a domain-adaptive feature vector that enhances the target region.

[0040] S202: By constructing a physical correlation map between the first remote sensing feature and the geological parameters corresponding to the target area, the physical correlation map is based on geological prior knowledge, uses a Bayesian network to model the conditional dependencies between multi-source data, mines the constitutive relationships between modes through a graph neural network, and outputs the structure of the physical correlation map representing the modes.

[0041] S3: The second remote sensing feature is processed by an automatic recognition model to obtain the third remote sensing feature. The automatic recognition model is a three-layer processing chain consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer, and a critical recognition layer.

[0042] S301: Three-dimensional dilated convolution is used to capture multi-scale spatiotemporal features in the second remote sensing features, and the multi-scale spatiotemporal features are fused according to the dilation coefficient;

[0043] S302: Construct a spatiotemporal graph structure of the disaster body corresponding to multi-scale spatiotemporal features, identify the deformation propagation path through a dual-modal graph attention network, and output the spatiotemporal evolution features of the disaster body;

[0044] S303: Construct the dynamic process of disaster body based on LSTM-Transformer hybrid framework to output disaster body evolution labels;

[0045] S4: Based on the aforementioned third remote sensing features, construct a geological disaster assessment matrix and generate graded early warning signals;

[0046] S5: Receive the first remote sensing features of the target area in real time and perform optimization control operations, wherein the optimization control operations are used to optimize and adjust the cross-domain physical fusion mechanism and the automatic identification model;

[0047] S6: Map the graded early warning signals into a three-dimensional geological model map to display the explanation and analysis of the causes of geological disasters.

[0048] Preferably, step S5, performing optimization control operations, specifically includes:

[0049] When the KL divergence between the first remote sensing feature of the target area and historical data exceeds a preset threshold, the domain adaptation network parameters of the domain adversarial network mechanism are updated through online knowledge distillation.

[0050] If the domain-adaptive network parameter update is triggered by the domain adversarial network mechanism, the weight distribution of the three-dimensional dilated convolution kernel is adjusted synchronously.

[0051] The technical effects and advantages of this invention are as follows:

[0052] 1. This invention solves the problem of registration error accumulation caused by inconsistent spatial references of multi-source data by using a synergistic optimization technique of spatial registration and radiometric correction, and significantly reduces the risk of reduced generalization ability due to data heterogeneity when applied across regions.

[0053] 2. This invention utilizes a domain adversarial network to generate virtual features through a cross-domain physical fusion module, thereby improving the model's generalization ability when applied across regions. It solves the problem of decreased recognition accuracy caused by differences in the geological environment between the target area and the training data in existing models. At the same time, it deeply explores the physical correlation between multi-source remote sensing data, thus eliminating the reliance on manually designed rules.

[0054] 3. This invention effectively improves the ability to capture disaster features in complex geological environments, especially the accuracy of identifying precursory minor deformations, by using a three-layer processing chain technology consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer, and a critical identification layer. This reduces the risk of missed reports and breaks through the limitations of traditional feature-level fusion strategies in physical mechanism mining and feature expression. Attached Figure Description

[0055] Figure 1 This is a module structure diagram of an automatic geological disaster identification system based on multi-source remote sensing data provided in an embodiment of this application.

[0056] Figure 2 This is a flowchart illustrating an automatic geological hazard identification method based on multi-source remote sensing data, according to an embodiment of this application. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0059] Hereinafter, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0060] As attached Figure 1 The system shown is an automatic geological disaster identification system based on multi-source remote sensing data, including a system operation database, a system central processing module, and a user information terminal. The system operation database includes all data texts of the system operation and collects information texts output by each module in real time. The system central processing module is used to control the information text instructions output by each module in the central control system. The user information terminal is a device for receiving information output from the system.

[0061] It should be noted that the system's central processing unit (CPU) is the core computing and control unit of the entire automatic geological hazard identification system based on multi-source remote sensing data. It includes one or more processing cores, connected to various modules within the system via a PCIe bus and a gigabit Ethernet interface. By running or executing instructions, programs, code sets, or instruction sets stored in the system's operating database, and by calling data stored therein, it performs various functions of the automatic geological hazard identification system based on multi-source remote sensing data, including spatial registration and radiometric correction of multimodal remote sensing data, feature parameter extraction, and hazard risk assessment, to ensure the accuracy and efficiency of identification. The system operates... The database stores a large amount of data related to automatic geological hazard identification, including a mathematical association rule base established based on geomechanical principles and remote sensing inversion models, and a large amount of historical monitoring data, including time-series records of changes in parameters such as deformation rate, lithological permeability, and rainfall threshold, as well as past geological hazard types and corresponding remote sensing image features. When executing various functions, the system's central processing unit frequently retrieves this data from the system's operating database to perform operations such as data correction, model parameter adjustment, and risk threshold setting, thereby achieving real-time monitoring and early warning of automatic geological hazard identification based on multi-source remote sensing data. The user information terminal provides an interface for users to interact with the system by connecting to a 4K geological 3D display screen and a multispectral imager via an HDMI 2.1 interface and USB4 protocol.

[0062] In this embodiment, an automatic geological disaster identification system based on multi-source remote sensing data further includes a multimodal remote sensing data acquisition module, a cross-domain physical fusion module, a spatiotemporal evolution decision module, a multi-cascade early warning decision module, an optimization control module, and a visualization module;

[0063] Among them, the multimodal remote sensing data acquisition module is used to receive and perform spatial registration and radiometric correction on the first remote sensing data to generate the first remote sensing features;

[0064] The cross-domain physical fusion module is used to process the first remote sensing feature through the cross-domain physical fusion mechanism and output the second remote sensing feature. The cross-domain physical fusion mechanism includes a domain adversarial network mechanism and a physical constraint mechanism.

[0065] The spatiotemporal evolution decision module is used to obtain the third remote sensing feature from the second remote sensing feature through an automatic recognition model. The automatic recognition model is a three-layer processing chain consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer, and a critical recognition layer.

[0066] The multi-level early warning decision module constructs a geological disaster assessment matrix and generates graded early warning signals based on the third remote sensing features.

[0067] The optimization control module is used to receive the first remote sensing features of the target area in real time and perform optimization control operations, which are used to optimize and adjust the cross-domain physical fusion module and the spatiotemporal evolution decision module.

[0068] The visualization module is used to map the graded early warning signals into a three-dimensional geological model map, displaying the explanation and analysis of the causes of geological disasters.

[0069] Specifically, in the multimodal remote sensing data acquisition module, first remote sensing data is collected through collaborative observation using multiple platform remote sensing sensors. The first remote sensing data includes satellite multispectral imagery, InSAR radar interferometry data, and lidar point cloud data. The first remote sensing data is preprocessed to perform spatial registration and radiometric correction on the first remote sensing data to obtain the first remote sensing features corresponding to the first remote sensing data. The first remote sensing features include multispectral features, InSAR features, and lidar features.

[0070] It should be noted that the satellite multispectral imagery in the first remote sensing data includes multispectral data in the visible light band, near-infrared band, and short-wave infrared band; the InSAR radar interferometry data consists of master-slave image pairs; and the lidar point cloud data includes three-dimensional coordinates, laser echo intensity, and the time difference between the first and last echoes. The multispectral features in the first remote sensing features include apparent reflectance, normalized vegetation index, and normalized water index for each band; the InSAR features include annual average deformation rate, coherence coefficient matrix, and elevation change; and the lidar features include a digital elevation model and initial classification labels based on the number of echoes and laser echo intensity. The initial classification labels include, but are not limited to, ground points, vegetation points, and building points.

[0071] This embodiment utilizes multi-platform remote sensing sensors to acquire the first remote sensing data, including but not limited to using multispectral imagery from Landsat-8OLI and Sentinel-2MSI satellites, covering the 450-680nm visible light band, the 700-900nm near-infrared band, and the 1550-2300nm short-wave infrared band, with a spatial resolution of 10-30 meters and a temporal resolution of 5-10 days. Simultaneously, it acquires C-band radar interferometric data with a wavelength of 5.6cm from the Sentinel-1SAR satellite, generating differential interferograms through ascent and descent orbit observations, achieving a spatial resolution of 5×20 meters and a deformation measurement accuracy of ±3mm / year to accurately monitor minute surface deformations. Finally, it acquires high-density lidar point cloud data using an airborne RIEGL VQ-1560i system, with a density of ≥20 points / m². 2 The wavelength is 1550nm and the elevation accuracy is ±5cm.

[0072] Furthermore, the raw data transmitted from remote sensing sensors across multiple platforms undergoes format standardization processing, including: converting satellite multispectral imagery into radiance values ​​and embedding them into the WGS84 geographic coordinate system; reducing speckle noise and generating phase and coherence coefficient matrices for InSAR radar interferometric data through multi-view processing; and denoising lidar point cloud data using a statistical outlier removal algorithm and classifying ground points and non-ground points to generate a digital elevation model.

[0073] Furthermore, the spatial registration of the first remote sensing data after format standardization includes: using a digital elevation model generated from lidar point cloud data as a spatial reference, resampling the digital elevation model to the UTM projection coordinate system using a bilinear interpolation algorithm; extracting scale-invariant feature points from the differential interferograms of satellite multispectral imagery and InSAR radar interferometry data, with parameters set to contrast > 0.03 and curvature ratio < 10, and using a Brute-Force matcher for bidirectional matching, filtering matching points with a matching score > 0.7; applying the Levenberg-Marquardt algorithm to correct the position of the initial matching points, and using the RANSAC algorithm to remove mismatches; performing an affine transformation based on the matching points, and solving the transformation parameters using the least squares method, with the residual controlled within 0.3 pixels; and resampling the low-resolution data to the digital elevation model using bicubic convolution interpolation.

[0074] Furthermore, the radiometric correction of the first remote sensing data after format standardization includes: atmospheric correction of the apparent reflectance of the land surface using the 6S radiative transfer model for satellite multispectral imagery; elimination of terrain shading effects using the SCS+C model based on the digital elevation model; processing of the entangled phase using the minimum cost flow (MCF) algorithm for InSAR radar interferometric data; and distance attenuation correction of the laser echo intensity for lidar point cloud data, specifically expressed as follows:

[0075]

[0076] Where I1 represents the laser echo intensity after distance attenuation correction, I0 represents the original laser echo intensity, R represents the laser ranging value, and R0 represents the preset distance.

[0077] Specifically, in the cross-domain physical fusion module, the first remote sensing feature is processed through a cross-domain physical fusion mechanism, which consists of a domain adversarial network mechanism and a physical constraint mechanism. The domain adversarial network mechanism uses a deep domain adversarial network composed of a feature extractor, a label classifier, and a domain classifier to generate virtual features corresponding to the first remote sensing feature and outputs a domain-adaptive feature vector enhanced for the target region. The physical constraint mechanism constructs a physical correlation graph between the first remote sensing feature and the geological parameters corresponding to the target region. This physical correlation graph is based on prior geological knowledge, utilizes a Bayesian network to model the conditional dependencies between multi-source data, mines constitutive relationships between modes through a graph neural network, and outputs a physical correlation graph structure representing the modes. The output second remote sensing feature includes a domain-adaptive feature vector enhanced for the target region and a physical correlation graph structure.

[0078] In this embodiment, the feature extractor uses a ResNet-50 backbone network, removing the last fully connected layer to output a 1024-dimensional feature vector; the label classifier uses a bilinear interpolation layer to output the probability distribution of disaster types; the domain classifier consists of a gradient inversion layer and a fully connected layer, outputting binary classification results for the source and target domains; a three-layer Bayesian network is used to construct a physical correlation graph between the first remote sensing features and the geological parameters corresponding to the target area. In the three-layer Bayesian network, the observation layer nodes include the apparent reflectance, InSAR deformation rate, and initial classification labels based on the number of echoes and the intensity of laser echoes for each band; the hidden layer nodes include the geological parameters corresponding to the target area; and the decision layer nodes include the risk probabilities of geological disasters stored in the system's operating database. The physical correlation graph between the representation modes consists of nodes composed of multispectral features, InSAR features, and geological parameters. The edge types include physical mechanism edges encoding geomechanical equations, statistically correlated edges connected by feature pairs with mutual information greater than 0.3, and spatiotemporal co-occurrence edges connecting spatially adjacent and temporally synchronized nodes.

[0079] Specifically, in the spatiotemporal evolution decision module, the second remote sensing feature is processed by an automatic identification model, which consists of a three-layer processing chain consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer, and a critical identification layer; the third remote sensing feature includes multi-scale spatiotemporal features, spatiotemporal evolution features of the disaster body, and disaster body evolution labels.

[0080] Furthermore, the automatic identification model is a pre-constructed three-layer processing chain. By inputting the second remote sensing feature into the automatic identification model, the automatic identification model obtains the third remote sensing feature based on the second remote sensing feature. The three-layer processing chain of the automatic identification model includes: a spatiotemporal feature extraction layer that uses three-dimensional dilated convolution to capture multi-scale spatiotemporal features in the second remote sensing feature, and fuses the multi-scale spatiotemporal features according to the dilation coefficient, wherein the size of the three-dimensional dilated convolution kernel is positively correlated with the node degree of the physical correlation graph; a dynamic graph evolution layer that constructs the spatiotemporal graph structure of the disaster body corresponding to the multi-scale spatiotemporal features, identifies the deformation propagation path through a dual-modal graph attention network, and outputs the spatiotemporal evolution features of the disaster body, wherein the node state vector dimension of the spatiotemporal graph of the disaster body corresponding to the dynamic graph evolution layer is equal to the hidden layer dimension of the LSTM unit in the critical identification layer; and a critical identification layer that constructs the dynamic process of the disaster body based on the LSTM-Transformer hybrid framework to output the disaster body evolution label, wherein the output gradient of the critical identification layer is backpropagated to the dilation rate adjustment coefficient of the spatiotemporal feature extraction layer.

[0081] Furthermore, the implementation steps of the spatiotemporal feature extraction layer include: employing a three-dimensional dilated convolution architecture, wherein the weight matrix WM corresponding to the three-dimensional dilated convolution kernel... ck Based on the node degree D in the physical correlation diagram i Dynamic calculation is performed, specifically as follows:

[0082]

[0083] Among them, D i Let ∑D represent the degree of node i in the physical association graph. j It is represented as the sum of the degrees of all nodes in the physical association graph. This is represented by the normalization of node degree using the Softmax function, where WM0 represents the weight matrix corresponding to the pre-set 3D dilated convolution kernel. The multi-scale spatiotemporal features include spatial scale features and temporal scale features. In this embodiment, spatial scale features capture micro-topographic changes on the land surface using a 3D convolution kernel with a small dilation rate; a convolution kernel with a medium dilation rate captures mesoscale spatial correlations such as topographic gradient and slope aspect consistency; a convolution kernel with a large dilation rate extracts macro-topographic trends; and temporal scale features capture the time window for sudden events (less than 7 days), the time window for seasonal changes (set to 1-6 months), and the time window for the cumulative effect of human engineering impacts (set to more than 1 year).

[0084] Furthermore, the construction process of the dynamic graph evolution layer includes: converting multi-scale spatiotemporal features into a disaster body spatiotemporal graph structure, wherein the node attributes corresponding to the disaster body spatiotemporal graph structure include multispectral features, InSAR features, and lidar features; and calculating the priority of deformation propagation paths in the dual-modal graph attention network through a structural attention head, specifically expressed as:

[0085]

[0086] Where, p (ti,tj) Let represent the priority of the deformation propagation path from node ti to node tj, and let p represent the learnable weight vector that measures the similarity of node attributes corresponding to the spatiotemporal graph structure of the disaster body. T This is represented as the transpose operation on p, Wh ti and Wh tj Represented as the embedding representation of nodes ti and tj, [Wh i ‖Wh j ] represents the concatenated vector of the embedding representations of nodes ti and tj, LeakyReLU (p T [Wh i ‖Wh j ]) represents p T [Wh i ‖Wh j Apply the LeakyReLU activation function, exp(LeakyReLU(p T [Wh ti ‖Wh tj ])) represents LeakyReLU(p T [Wh i ‖Wh j The activation result is applied using an exponential function, where tk represents all neighboring nodes of node ti.

[0087] In this embodiment, the nodes of the spatiotemporal graph structure of the disaster body correspond to 50×50 meter grid cells. Spatial adjacent edges are automatically connected within a 3×3 neighborhood. Deformation propagation edges are established as directed edges based on the displacement gradient direction, and are activated when the gradient is greater than 5mm / 100m. Physical constraint edges are selected for effective connections using the Mohr-Coulomb criterion. The graph structure is refreshed every 6 hours. Regions with abrupt changes in deformation rate greater than 10mm / day are automatically divided into sub-grids. Regions with incoherence coefficients less than 0.25 are suspended from monitoring. The propagation weight of relevant edges is strengthened by real-time rainfall data greater than 50mm / h.

[0088] Furthermore, the operation method of the critical identification layer includes: the LSTM branch extracts the cumulative effect features of the deformation rate of the disaster body through the temporal sequence of the node states in the disaster body spatiotemporal graph; the Transformer branch captures the mutation synergy features through the flattened sequence of the adjacency matrix and node features of the disaster body spatiotemporal graph; and the fusion weights of the LSTM and Transformer outputs are automatically adjusted according to the warning level at time t-1, where time t-1 represents the time before the current time t.

[0089] Specifically, in the multi-level early warning decision module, a probabilistic risk assessment matrix is ​​constructed based on the spatiotemporal evolution features and disaster body evolution labels in the third remote sensing features. In this embodiment, the spatial location dimension of the probabilistic risk assessment matrix divides the target area into 1m×1m grid cells using the UTM coordinate system, with each cell corresponding to a unique coordinate code. The disaster type dimension is preset to include three types of geological disaster labels: landslide, debris flow, and ground subsidence, with probability values ​​ranging from [0, 1]. The evolution stage dimension is divided into four stages: stable period, incubation period, acceleration period, and critical instability period. The stage labels are output by the critical identification layer. The input features for risk probability calculation include the spatiotemporal evolution features of the disaster body from the node embedding vector of the disaster body spatiotemporal graph structure. A graded early warning signal is generated based on the probabilistic risk assessment matrix, and the graded early warning signal is backpropagated to the dynamic graph evolution layer to optimize the learnable weight vector in the dual-modal graph attention network that measures the similarity of node attributes corresponding to the spatiotemporal graph structure of the disaster body.

[0090] It should be noted that the tiered early warning signal is a disaster risk level instruction generated based on the quantitative analysis results of the probability risk assessment matrix. The disaster risk level instruction includes four warning levels: Level I, Level II, Level III, and Level IV. The trigger condition for a Level I warning is that the disaster risk probability enters an extremely high probability range or the evolution stage is determined to be in a critical instability period. This indicates that a geological disaster is about to occur or has already shown initial destructive characteristics, requiring the activation of the highest-level emergency response plan, closure of surrounding traffic routes, and real-time push of warning information to public terminals such as SMS and mobile apps through multiple channels. The trigger condition for a Level II warning is that the disaster risk probability reaches a medium-to-high probability range or the evolution stage is determined to be in an accelerated deformation period. This indicates that the disaster body has entered a rapid deformation stage, with a possibility of short-term instability. This requires delineating a danger zone containing the potential disaster impact area, initiating resident evacuation procedures, and implementing targeted engineering reinforcement measures. Level III... Level 1 warning is triggered when the probability of disaster occurrence enters the medium probability range or the evolution stage is identified as the gestation period, indicating a significant increase in the geological disaster risk level and the local area entering the deformation accumulation stage. It is necessary to activate the high-resolution UAV inspection program, initiate emergency plan rehearsals, and restrict non-essential personnel from entering potentially risky areas. Level 4 warning is triggered when the probability of disaster occurrence is in the low probability range and the evolution stage is determined to be the stable period, indicating that there is potential risk in the target area. It is necessary to activate the basic monitoring mechanism, increase the InSAR deformation data update frequency to 1.5-2 times the regular monitoring cycle, send phased monitoring reports to relevant departments, and recommend conducting regular manual inspections. In this embodiment, the extremely high probability range of disaster risk is set to [0.9, 1), the medium-high probability range is set to [0.7, 0.9), the medium probability range is set to [0.5, 0.7), and the low probability range is set to [0.3, 0.5].

[0091] Specifically, in the optimization control module, when the KL divergence between the first remote sensing feature of the target area and historical data exceeds a preset threshold, the domain adaptation network parameters of the domain adversarial network mechanism of the cross-domain physical fusion module are updated through online knowledge distillation to minimize the early warning error of the target area and improve the cross-regional generalization ability of the domain adaptation feature vector; if the domain adaptation network parameter update of the domain adversarial network mechanism is triggered, the three-dimensional dilated convolution kernel weight distribution of the spatiotemporal evolution decision module is adjusted synchronously; and the connection edges of unrelated modes in the physical association graph are automatically closed according to the geological structure type of the target area.

[0092] Specifically, in the visualization module, the hierarchical early warning signal is mapped to the physical correlation diagram output by the cross-domain physical fusion module and the spatiotemporal evolution characteristics of the disaster body output by the spatiotemporal evolution decision module into a geomechanical interpretation model, and an interpretability analysis of the disaster causes is provided.

[0093] As attached Figure 2The method for automatic identification of geological hazards based on multi-source remote sensing data, as shown, includes:

[0094] S1: Receive and perform spatial registration and radiometric correction on the first remote sensing data to generate the first remote sensing feature;

[0095] S2: The first remote sensing feature is processed through a cross-domain physical fusion mechanism, and a second remote sensing feature is output. The cross-domain physical fusion mechanism includes a domain adversarial network mechanism and a physical constraint mechanism.

[0096] S201: A deep domain adversarial network consisting of a feature extractor, a label classifier, and a domain classifier generates virtual features corresponding to the first remote sensing feature and outputs a domain-adaptive feature vector that enhances the target region.

[0097] S202: By constructing a physical correlation map between the first remote sensing feature and the geological parameters corresponding to the target area, the physical correlation map is based on geological prior knowledge, uses a Bayesian network to model the conditional dependencies between multi-source data, mines the constitutive relationships between modes through a graph neural network, and outputs the structure of the physical correlation map representing the modes.

[0098] S3: The second remote sensing feature is processed by an automatic recognition model to obtain the third remote sensing feature. The automatic recognition model is a three-layer processing chain consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer, and a critical recognition layer.

[0099] S301: Three-dimensional dilated convolution is used to capture multi-scale spatiotemporal features in the second remote sensing features, and the multi-scale spatiotemporal features are fused according to the dilation coefficient;

[0100] S302: Construct a spatiotemporal graph structure of the disaster body corresponding to multi-scale spatiotemporal features, identify the deformation propagation path through a dual-modal graph attention network, and output the spatiotemporal evolution features of the disaster body;

[0101] S303: Construct the dynamic process of disaster body based on LSTM-Transformer hybrid framework to output disaster body evolution labels;

[0102] S4: Based on the aforementioned third remote sensing features, construct a geological disaster assessment matrix and generate graded early warning signals;

[0103] S5: Receive the first remote sensing features of the target area in real time and perform optimization control operations, wherein the optimization control operations are used to optimize and adjust the cross-domain physical fusion mechanism and the automatic identification model;

[0104] S6: Map the graded early warning signals into a three-dimensional geological model map to display the explanation and analysis of the causes of geological disasters.

[0105] In this embodiment, the optimization control operation in S5 includes: when the KL divergence between the first remote sensing feature of the target area and the historical data exceeds a preset threshold, updating the domain adaptation network parameters of the domain adversarial network mechanism through online knowledge distillation to minimize the early warning error of the target area and improve the cross-regional generalization ability of the domain adaptation feature vector; if the domain adaptation network parameter update of the domain adversarial network mechanism is triggered, the weight distribution of the three-dimensional dilated convolution kernel is adjusted synchronously.

[0106] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0107] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic geological hazard identification system based on multi-source remote sensing data, comprising a system operation database, a system central processing module, and a user information terminal, characterized in that, Also includes: Multimodal remote sensing data acquisition module: used to receive and perform spatial registration and radiometric correction on the first remote sensing data to generate the first remote sensing features; Cross-domain physical fusion module: used to process the first remote sensing feature through a cross-domain physical fusion mechanism and output the second remote sensing feature. The cross-domain physical fusion mechanism includes a domain adversarial network mechanism and a physical constraint mechanism. Spatiotemporal evolution decision module: used to obtain a third remote sensing feature from the second remote sensing feature through an automatic recognition model, wherein the automatic recognition model is a three-layer processing chain consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer and a critical recognition layer; The spatiotemporal evolution decision module processes the second remote sensing feature through the automatic identification model. The three-layer processing chain of the automatic identification model satisfies the following: the size of the three-dimensional dilated convolution kernel of the spatiotemporal feature extraction layer is positively correlated with the node degree of the physical correlation graph; the dimension of the node state vector of the spatiotemporal graph of the disaster body corresponding to the dynamic graph evolution layer is equal to the hidden layer dimension of the LSTM unit in the critical identification layer; the output gradient of the critical identification layer is backpropagated to the dilation rate adjustment coefficient of the spatiotemporal feature extraction layer. Multi-level early warning decision module: Based on the aforementioned third remote sensing features, a geological disaster assessment matrix is ​​constructed, and graded early warning signals are generated; Optimization control module: used to receive the first remote sensing features of the target area in real time and perform optimization control operations, the optimization control operations being used to optimize and adjust the cross-domain physical fusion module and the spatiotemporal evolution decision module; Visualization module: used to map the graded early warning signals into a three-dimensional geological model map, displaying the explanation and analysis of the causes of geological disasters; The system operation database includes all data text of the system operation and collects information text output by each module in real time. The system central processing module is used to handle information text instructions output by each module in the central control system. The user information terminal is a device for receiving information output from the system.

2. The automatic geological hazard identification system based on multi-source remote sensing data according to claim 1, characterized in that: The multimodal remote sensing data acquisition module includes satellite multispectral images, InSAR radar interferometry data, and lidar point cloud data. The satellite multispectral images include multispectral data in the visible light band, near-infrared band, and short-wave infrared band. The InSAR radar interferometry data consists of master-slave image pairs. The lidar point cloud data includes three-dimensional coordinates, laser echo intensity, and the time difference between the first and last echoes.

3. The automatic geological hazard identification system based on multi-source remote sensing data according to claim 1, characterized in that: The cross-domain physical fusion module, through the cross-domain physical fusion mechanism, specifically includes: A deep domain adversarial network consisting of a feature extractor, a label classifier, and a domain classifier is used to generate virtual features corresponding to the first remote sensing features. A physical correlation map is constructed between the first remote sensing features and the corresponding geological parameters of the target area, and constitutive relationships are mined through graph neural networks; The physical correlation diagram is based on prior geological knowledge and uses a Bayesian network to model the conditional dependencies between the first remote sensing features. The second remote sensing feature includes the target area enhanced domain-adaptive feature vector and the physical correlation graph structure.

4. The automatic geological hazard identification system based on multi-source remote sensing data according to claim 1, characterized in that: The spatiotemporal evolution decision module contains a weight matrix corresponding to the three-dimensional dilated convolution kernel. Based on the node degree in the physical correlation diagram Dynamic calculation is performed, specifically as follows: in, Represented as nodes in the physical association graph The degree, It is represented as the sum of the degrees of all nodes in the physical association graph. This is represented by the normalization of node degree using the Softmax function. This represents the weight matrix corresponding to a pre-set 3D dilated convolution kernel.

5. The automatic geological hazard identification system based on multi-source remote sensing data according to claim 1, characterized in that: The spatiotemporal evolution decision module calculates the priority of deformation propagation paths in the bimodal graph attention network using a structural attention head, specifically as follows: in, Represented as nodes To the node The priority of deformation propagation paths, This is represented as a learnable weight vector that measures the similarity of node attributes corresponding to the spatiotemporal graph structure of a disaster body. Indicated as to The transpose operation, and Represented as nodes and nodes Embedded representation, Represented as nodes and nodes The embedding represents the concatenated vector. Indicated as to Apply the LeakyReLU activation function. Indicated as to The activation result applies an exponential function. Represented as nodes All neighboring nodes.

6. The automatic geological hazard identification system based on multi-source remote sensing data according to claim 1, characterized in that: The spatiotemporal evolution decision module, specifically the operation method of the critical identification layer, includes: The cumulative effect characteristics of the deformation rate of the disaster body are extracted from the temporal sequence of the node states in the spatiotemporal diagram of the disaster body. Mutual change synergy features are captured by flattening the adjacency matrix and node features of the spatiotemporal graph of the disaster body; The fusion weights are automatically adjusted based on the warning level at time t-1.

7. The automatic geological hazard identification system based on multi-source remote sensing data according to claim 1, characterized in that: The multi-level early warning decision module generates a disaster risk level instruction based on the quantitative analysis results of the probability risk assessment matrix. The disaster risk level instruction includes four levels of early warning: Level I, Level II, Level III, and Level IV. Level I early warning indicates that a geological disaster is about to occur or has already shown initial signs of damage. A Level II warning indicates that the disaster body has entered a rapid deformation stage and there is a possibility of short-term instability; a Level III warning indicates that the geological disaster risk level has increased significantly and that some areas have entered a stage of deformation accumulation; a Level IV warning indicates that there is potential risk in the target area.

8. A method for automatic identification of geological hazards based on multi-source remote sensing data, employing an automatic geological hazard identification system based on multi-source remote sensing data as described in any one of claims 1-7, characterized in that, include: S1: Receive and perform spatial registration and radiometric correction on the first remote sensing data to generate the first remote sensing feature; S2: The first remote sensing feature is processed through a cross-domain physical fusion mechanism, and a second remote sensing feature is output. The cross-domain physical fusion mechanism includes a domain adversarial network mechanism and a physical constraint mechanism. S201: A deep domain adversarial network consisting of a feature extractor, a label classifier, and a domain classifier generates virtual features corresponding to the first remote sensing feature and outputs a domain-adaptive feature vector that enhances the target region. S202: By constructing a physical correlation map between the first remote sensing feature and the geological parameters corresponding to the target area, the physical correlation map is based on geological prior knowledge, uses a Bayesian network to model the conditional dependencies between multi-source data, mines the constitutive relationships between modes through a graph neural network, and outputs the structure of the physical correlation map representing the modes. S3: The second remote sensing feature is processed by an automatic recognition model to obtain the third remote sensing feature. The automatic recognition model is a three-layer processing chain consisting of a spatiotemporal feature extraction layer, a dynamic graph evolution layer, and a critical recognition layer. S301: Three-dimensional dilated convolution is used to capture multi-scale spatiotemporal features in the second remote sensing features, and the multi-scale spatiotemporal features are fused according to the dilation coefficient; S302: Construct a spatiotemporal graph structure of the disaster body corresponding to multi-scale spatiotemporal features, identify the deformation propagation path through a dual-modal graph attention network, and output the spatiotemporal evolution features of the disaster body; S303: Construct the dynamic process of disaster body based on LSTM-Transformer hybrid framework to output disaster body evolution labels; S4: Based on the aforementioned third remote sensing features, construct a geological disaster assessment matrix and generate graded early warning signals; S5: Receive the first remote sensing features of the target area in real time and perform optimization control operations, wherein the optimization control operations are used to optimize and adjust the cross-domain physical fusion mechanism and the automatic identification model; S6: Map the graded early warning signals into a three-dimensional geological model map to display the explanation and analysis of the causes of geological disasters.

9. The method for automatic identification of geological hazards based on multi-source remote sensing data according to claim 8, characterized in that: S5 executes an optimization control operation, specifically including: When the KL divergence between the first remote sensing feature of the target area and historical data exceeds a preset threshold, the domain adaptation network parameters of the domain adversarial network mechanism are updated through online knowledge distillation. If the domain-adaptive network parameter update is triggered by the domain adversarial network mechanism, the weight distribution of the three-dimensional dilated convolution kernel is adjusted synchronously.

Images