A GIS-based multi-source data fusion geological disaster early identification system

Abstract

This invention discloses a GIS-based multi-source data fusion system for early identification of geological hazards, belonging to the field of geological hazard monitoring and prevention. It includes a multi-source heterogeneous data management and preprocessing module, a hybrid feature extraction and knowledge graph construction module, a three-driven fusion identification module, a multi-scale spatiotemporal evolution law identification module, and an interpretable risk assessment and dynamic update module. This invention achieves a hierarchical intelligent identification strategy of "full-domain screening - regional precise judgment - individual location" by integrating the efficient automation capabilities of data-driven approaches, the mechanical precision of physical mechanisms, and the scenario adaptability of expert knowledge. This adapts to the hazard identification needs of regions with different scales and geological conditions. It transforms the traditional black-box model into a white-box model, fully disclosing the algorithm variable definitions, calculation logic, parameter strategies, and boundary processing methods, ensuring that the identification results are reliable and filling the technical gap of insufficient credibility in the field of geological hazard identification.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and prevention and mitigation technology, and in particular to a GIS-based multi-source data fusion system for early identification of geological disasters. Background Technology

[0003] Currently, the technical methods for early identification of potential geological disaster sites, both domestically and internationally, are mainly divided into two categories.

[0004] The first category focuses on methods centered on physical mechanisms and geological exploration. For example, publication number CN113378396B discloses a method for early identification of geological hazard hazard points in small watersheds, proposing a technical solution that integrates composite high-definition spatial identification technologies (InSAR, LiDAR, and UAV optical remote sensing) with detailed ground surveys. It constructs a three-dimensional high-precision digital scene, combining indoor and outdoor experiments (geomechanical tests, model tests) and numerical simulation methods to build a geological hazard initiation identification model and a critical discrimination index system. This method starts from geotechnical mechanics, and its conclusions have clear physical significance. However, its limitations are: it highly relies on ground surveys and geomechanical tests, resulting in high data acquisition costs and long cycles, making it difficult to promote on a large scale; and the physical model is highly regional, with parameters difficult to transfer to areas with significant differences in geological conditions.

[0005] The second category is data-driven methods centered on multi-source remote sensing data and deep learning. For example, CN119247203A discloses an automatic geological hazard identification method and system based on multi-source remote sensing data. This method proposes acquiring optical imagery, synthetic aperture radar (SAR) data, and digital elevation model (DEM) data, and using techniques such as cross-modal registration, dynamic weight allocation based on channel attention mechanisms, and convolutional neural network prediction to achieve automatic geological hazard identification. This method leverages the wide coverage advantage of remote sensing data and combines it with the powerful feature extraction capabilities of deep learning, significantly improving the efficiency of large-scale identification. However, its inherent drawbacks are: the purely data-driven "black box" model lacks geotechnical support, has poor interpretability, and therefore lacks credibility in major engineering decisions; furthermore, it has weak identification capabilities for new hazard patterns under extreme conditions, relying solely on statistical data features and failing to incorporate prior geological knowledge for adaptive weight optimization, resulting in significant technical limitations.

[0006] In summary, existing technologies exhibit a clear "mechanism-data" disconnect. The former has a "clear mechanism but low efficiency," while the latter is "highly efficient but has a vague mechanism, lacks prior knowledge, and has poor adaptability."

[0007] Therefore, there is a need for a multi-source data fusion early identification system for geological hazards that can balance physical constraints and data-driven efficiency, and has good interpretability and generalization ability. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a GIS-based multi-source data fusion system for early identification of geological disasters, which solves the problems mentioned in the background section.

[0009] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a GIS-based multi-source data fusion system for early identification of geological hazards, deployed on a cloud-based GIS platform, comprising:

[0010] Module 1: Multi-source Heterogeneous Data Management and Preprocessing Module

[0011] This system is used to automatically access and store multi-source heterogeneous data from the area under test, including optical imagery, SAR data, DEM data, ground survey data, physical and mechanical parameters of soil and rock, hydrogeological data, and meteorological and precipitation data. Utilizing GIS spatial analysis capabilities, the system performs unified coordinate system transformation, cross-modal spatial registration (error <1.5 pixels), unification of temporal and spatial resolution scales, data cleaning, and normalization on the above data to generate a standardized multi-source raster dataset.

[0012] Module 2: Hybrid Feature Extraction and Knowledge Graph Construction Module, including:

[0013] (1) Data-driven feature extraction unit: Extracts surface environment and deformation feature maps such as NDVI, surface deformation rate, and slope based on multi-source raster dataset.

[0014] (2) Mechanism-driven feature extraction unit: Based on the GIS platform, the preset physical identification model is called to spatially calculate the geotechnical mechanical parameters (cohesion, internal friction angle, permeability coefficient) and generate mechanism-driven feature maps such as slope stability coefficient and critical rainfall threshold.

[0015] (3) Disaster Knowledge Graph Construction Unit: A semi-automatic method combining top-down ontology definition and bottom-up data extraction is adopted to construct a geological disaster-specific knowledge graph. The pre-set knowledge graph ontology model includes four core nodes: disaster-inducing environment, disaster-causing factors, disaster type, and evolution stage. Node attributes cover key indicators such as geological age, lithological combination, landform type, average annual rainfall, seismic intensity, and historical disaster frequency. The relationships between nodes include semantic relationships such as "inducing," "exacerbating," "inhibiting," and "matching." The knowledge graph is deployed on the Neo4j graph database and can accurately output expert prior weight vectors adapted to the area to be tested through a hybrid retrieval strategy of GIS spatial range retrieval and geological attribute matching, providing knowledge support for feature fusion.

[0016] Module 3: The "Mechanism-Data-Expert Knowledge" three-driven fusion recognition module, specifically including:

[0017] (1) Physically guided feature enhancement subunit:

[0018] Mechanism-driven feature maps are used as physical prior constraints and embedded in the training process of deep learning networks. This invention designs a physical consistency regularization loss term. Let the disaster probability map predicted by the data-driven network be... The normalized value of the stability coefficient in the mechanism-driven feature map is Both are two-dimensional raster spatial distribution maps. To balance global correlation and local spatial structure information, the covariance calculation adopts a local sliding window mean calculation method: a fixed-size sliding window traverses the entire feature map, calculating the covariance of pixels within each window, and finally taking the average of all window calculation results as the global covariance value. This effectively preserves spatial neighborhood features and avoids the defect of global flattening vectors destroying spatial structure. A physical guidance loss term is defined. As follows (Formula 1):

[0019]

[0020] Where Cov is the local sliding window covariance operation, and ReLU is the linear rectified function. The mathematical logic of this formula is equivalent to... A positive penalty term is output to constrain network parameters only when the predicted probability and the physical stability coefficient are negatively correlated (violating the laws of geomechanics); when they are positively correlated, the penalty value is 0, and it does not interfere with normal training. Total loss function. Defined as data-driven supervised loss The weighted sum of (e.g., cross-entropy loss) and physical guidance loss (Formula 2):

[0021]

[0022] in, To dynamically and adaptively balance the hyperparameters, a dual-loss-level adaptive adjustment strategy is adopted, and the calculation formula is as follows: , It is a very small constant (with values) The constant 10 is used to prevent the denominator from being zero. This minimum constant is chosen based on the precision characteristics of computer floating-point operations; 10 can completely avoid the abnormal value of dividing the denominator by zero without changing the magnitude of the original loss function and the gradient update pattern, thus adapting to the accuracy requirements of geological hazard model training. This strategy can dynamically balance the weights according to the real-time magnitude of the two types of losses during training, balancing data fitting and physical constraints in the early stages of training, and gradually strengthening physical mechanism constraints in the later stages of training, solving the defect that fixed parameters cannot adapt to the training iteration process. This design forces the network not only to fit the labels during training but also to ensure that the predicted probability and the slope stability coefficient show a positive correlation trend, fundamentally suppressing the physical deviation prediction of the pure data model in the sparse sample area.

[0023] (2) Dynamic weight allocation sub-unit:

[0024] Feature recalibration is based on an improved channel attention mechanism. Unlike existing technologies that rely solely on fixed weight updates based on data statistics, this invention innovatively introduces prior knowledge constraints from knowledge graph experts to achieve adaptive fusion of data features and geological prior knowledge. This overcomes the technical shortcomings of simple linear superposition and significantly improves creativity and scenario adaptability. Let the total number of feature channels be... , The original channel weight vector is generated by the deep learning attention mechanism. This refers to the expert prior weight vector obtained from knowledge graph retrieval. Both are uniformly mapped to the logarithmic domain to ensure dimensional matching and unit consistency in scalar multiplication operations. The data-driven weight vector is defined as follows: The expert prior weight vector obtained from the knowledge graph retrieval is This scheme designs a dynamic fusion weight generation formula (Formula 3):

[0025]

[0026] in, To dynamically integrate scalar coefficients, the data reliability of the region is dynamically determined, satisfying the following requirements. ,and ( For the Sigmoid function, (Based on a comprehensive evaluation of remote sensing image quality, data completeness, and coherence). By weighted fusion across multiple data domains followed by Softmax normalization, intelligent sharpening and reasonable allocation of feature weights are achieved, avoiding distribution distortion caused by a second Softmax step. The generated fusion weights... Compared with the original multi-channel feature map Perform channel-by-channel weighting to generate a weighted fusion feature map. .

[0027] (3) Multi-model ensemble learning sub-unit:

[0028] The weighted fused feature maps are input in parallel into the data-driven branch (CNN) and the mechanism-driven branch (infinite slope model). The data-driven branch outputs the disaster probability. Mechanism branch outputs disaster probability To address the shortcomings of traditional fusion models, such as numerical instability and complete loss of extreme divergence information, this paper proposes a Bayesian fusion formula with boundary constraints and a divergence penalty factor. Furthermore, it applies boundary pruning and lower bound constraints to the probability values, thereby improving the algorithm's stability and engineering applicability (Formula 4).

[0029]

[0030] Among them, to avoid boundary value anomalies, limits are set. , Pick Minimal positive values ​​are used to avoid numerical invalidation caused by extreme probabilities of 0 and 1. The lower limit threshold for fusion probability is determined through comparative experiments of multiple sets of extreme divergence samples. This can effectively avoid the problem of probability dropping to zero and complete loss of effective information in extreme divergence scenarios, while also avoiding the introduction of invalid low-probability noise to interfere with the recognition results. As the divergence penalty coefficient, based on historical disaster samples from 12 typical small watersheds in the earthquake-stricken areas of Wenchuan, Lushan, and Jiuzhaigou in Sichuan Province, as well as in the mountainous areas of Southwest China, a grid search method was used to analyze the data. Parameter sensitivity analysis was performed within the interval [0,1] with a step size of 0.05. The interval with the optimal F1 score on the validation set and balanced accuracy in identifying multiple disasters such as landslides, debris flows, and collapses was selected, and the optimal range of values ​​was finally determined. This interval balances the effectiveness of the divergence penalty with the preservation of valid information, avoiding excessive or ineffective penalties. When the two branches predict in agreement, the fusion probability approaches a very high value; when the two branches diverge, the fusion probability is adaptively reduced while preserving the basic recognition reliability, ensuring the robustness, stability, and engineering rationality of the output results.

[0031] Module 4: Multi-scale Spatiotemporal Evolution Law Identification Module

[0032] It supports three-level spatial scale drilling analysis, from large regions (provincial level) to key areas (county level) and then to individual hazard points (single units), and establishes a multi-scale resolution matching mechanism: 20m resolution data is used for full-area screening of provincial-level large areas, 10m resolution data is used for detailed analysis of key county-level areas, and 5m high-resolution data is used for precise location of individual hazard points. For the joint disaster probability map output by module three, a statistical quantile adaptive threshold method is used to delineate high-probability risk areas. The top 5% of grids with the highest probability values ​​are selected by default as high-risk areas to be assessed. This threshold is based on statistical calibration of disaster samples across the entire southwestern mountainous region, which can balance the overall underreporting rate and computational redundancy. At the same time, the quantile threshold can be dynamically adjusted according to the regional disaster development characteristics to adapt to the identification needs of different risk levels. Time-series InSAR technology is used to extract multi-year surface deformation sequences, and an adaptive piecewise fitting method is used to monitor the complete evolution stage of "uniform deformation - accelerated deformation - impending landslide and instability" in hazard areas in real time, achieving a deep integration of static probability identification and dynamic evolution monitoring.

[0033] Module 5: Explainability Risk Assessment and Dynamic Update Module

[0034] Attention map backprojection technology is used to extract key evidence features that contribute most to the judgment, and these features are associated with the disaster-inducing and disaster-causing attributes of the knowledge graph to generate a natural language interpretable report with a complete chain of evidence. Simultaneously, the Elastic Weight Consolidation (EWC) incremental learning method is employed to achieve closed-loop evolution of the system. The core principle is to impose weight constraints on network parameters that have contributed significantly in historical training tasks, suppressing large-scale parameter updates and addressing the catastrophic forgetting problem caused by training with new samples. Specifically, new samples are used for incremental iterative training of the deep learning network in the field, while Bayesian inversion correction of the mechanistic model parameters is performed based on measured soil and rock data. This continuously optimizes the recognition accuracy and adaptability of the dual-drive model, forming a closed-loop intelligent evolutionary system of data iteration, model optimization, and accuracy improvement.

[0035] Furthermore, the overall workflow

[0036] S1: Collect and preprocess multi-source data, unify resolution, coordinate system and dimensions, and generate standardized raster datasets;

[0037] S2: Extract data-driven and mechanism-driven feature maps respectively, and construct and initialize a disaster knowledge graph based on the geological ontology model;

[0038] S3: Input the mechanism features as physical constraints into the deep learning network, complete the model training through the loss function with dynamic adaptive weights, generate fusion features by combining the prior weights of knowledge graph experts, and output the global joint disaster probability map through the Bayesian ensemble algorithm with boundary constraints.

[0039] S4: High-risk areas are delineated by adaptive statistical thresholds, focusing step by step according to three spatial scales, and the stage of disaster evolution is determined by combining time-series deformation data;

[0040] S5: Output an interpretable risk assessment report, consolidate incremental learning with elastic weights to complete model iterative optimization, and achieve dynamic system updates.

[0041] The beneficial effects of the GIS-based multi-source data fusion early identification system for geological hazards of this invention are as follows:

[0042] (1) This invention integrates the efficient automation capabilities driven by data, the mechanical precision of physical mechanisms and the scene adaptability of expert knowledge to realize a hierarchical intelligent identification strategy of "full-domain screening - regional precision judgment - single-unit positioning", which can be adapted to the disaster identification needs of different scales and different geological conditions.

[0043] (2) The physical consistency loss function, knowledge graph fusion weight algorithm, and constrained Bayesian fusion formula of the present invention turn the traditional black box model into a white box, fully disclose the algorithm variable definition, calculation logic, parameter strategy and boundary processing method, the technical solution is clear and implementable, the identification results are reliable, and fill the gap of low credibility of the existing technology.

[0044] (3) This invention solves the technical defects of traditional algorithms such as spatial information loss, parameter solidification, numerical anomalies, and failure in extreme scenarios by calculating local covariance, dynamic adaptive hyperparameters, probability boundary clipping, lower limit threshold constraints, and reasonable penalty coefficient values. The generalization ability of the model is significantly improved, and the recognition accuracy of zero-sample and few-sample regions is greatly optimized.

[0045] (4) The present invention adopts an elastic weight consolidation incremental learning strategy to avoid the catastrophic forgetting problem of the model. Combined with Bayesian inversion correction of mechanism parameters, the system can continuously iterate and evolve with the accumulation of data, providing a stable, accurate, implementable and interpretable platform tool for geological disaster prevention and mitigation decision-making.

[0046] (5) This invention constructs a new architecture with deep coupling of three drivers: data features, physical mechanisms and expert knowledge. Through dynamic prior constraints of knowledge graph, guidance of physical loss regularization, and fusion of divergence adaptive penalty, it solves the problems of single driver, missing dimensions and rigid logic in existing technologies, and has significant substantial improvement and creativity. Attached Figure Description

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0048] Figure 1 This is a schematic diagram of the framework of the present invention. Detailed Implementation

[0049] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0050] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1: System Architecture and Data Preprocessing

[0052] This embodiment uses a county in a mountainous area of ​​southwestern China as the area to be measured. The system adopts a B / S architecture, deployed on a cloud server, and configured with a GPU for model acceleration. Sentinel-1A SAR data (2018-2025), Sentinel-2 optical imagery, and ALOS 30m DEM data were collected, supplemented by typical regional ground survey geotechnical parameters (cohesion c=25kPa, internal friction angle φ=30°) obtained from field surveys. All data were unified to the WGS84 coordinate system, resampled to 5m, 10m, and 20m resolutions according to multi-scale requirements, and Z-score normalization was performed. Simultaneously, boundary preprocessing was performed on the original probabilistic features, and the data was cropped to […]. ,1- [ ] range, to avoid numerical anomalies in subsequent algorithms.

[0053] Example 2: Physically Guided Feature Enhancement and Training Implementation

[0054] We construct a U-Net++ deep learning network, taking multi-channel feature maps as input. During the training phase, traditional methods only compute... The dynamic weighted total loss strictly implemented in this embodiment The covariance is calculated using a 3×3 sliding window, calculating the mean value for each region to preserve local spatial correlation features. The ReLU(-Cov) operation logic in this invention is equivalent to max(0,-Cov), penalizing only negatively correlated samples that violate physical laws. Suppose a batch contains a landslide sample; the network predicts the probability of that landslide. The normalized value of the mechanism stability coefficient calculated based on the limit equilibrium principle... The local covariance is positive. The network updates normally. If in another unknown stable region, the network predicts the probability... However, physical mechanisms indicate that the bedrock at this location is stable. The local covariance is negative. A positive penalty value is generated, forcing the network to adjust its weights during backpropagation, thus causing subsequent predictions to converge towards physical laws. During training... The model dynamically and adaptively adjusts to the magnitude of the two types of losses, focusing on data fitting in the early stage of training and physical constraints in the later stage of training to improve the stability of model training.

[0055] Example 3: Implementation of Dynamic Weight Allocation and Feature Fusion

[0056] In this embodiment, the input feature map includes NDVI (channel 1), deformation rate (channel 2), and slope (channel 3), with a total of 3 feature channels C. The original weight vector is extracted automatically by the data-driven network. It emphasizes deformation rate characteristics. The area to be measured is a mountainous region with high earthquake intensity. Through a hybrid retrieval of knowledge graph space and attributes, the returned expert prior weight vector is analyzed. The weight of the "slope" channel has been significantly increased. Current InSAR data exhibits high coherence and high confidence. Calculated by Sigmoid , Both types of weight vectors are pre-mapped to the logarithmic domain to ensure computational consistency. Weighted fusion and Softmax normalization are completed according to Formula 3. Finally, the fused weights amplify the slope characteristic response, enabling the network to focus on the mechanical mechanism of earthquake-induced landslides. This achieves a deep integration of data features and geological prior knowledge, which is different from the simple data statistical weighting mode of existing technologies.

[0057] Example 4: Implementation of Multi-Model Integration and Joint Probability Calculation

[0058] The penalty coefficient in this embodiment Taking the optimal median value of 0.5 as the lower bound for fusion probability. Selecting key potential hazard point A, the data-driven CNN model outputs the probability. Mechanism drives the output probability of the infinite slope model The two branches predict highly similar results. Substituting this into Formula 4, the penalty term shows a very small decay, ultimately... The identification results are accurate and reliable. Compared to extreme divergence risk point B, , The two branches diverged significantly, and the penalty term effectively reduced the fusion probability. Simultaneously, by applying a lower bound threshold constraint, the final fusion probability stabilized above 0.05. This approach achieves both divergence penalty and preservation of valid analytical information, avoiding the extreme drawback of traditional algorithms where the probability could drop to zero. High-risk areas across the entire domain were dynamically defined using the top 5% probability quantiles, balancing comprehensive screening with computational efficiency.

[0059] Example 5: Closed-loop update implementation

[0060] After validating the identification results in the field, new samples are added to the incremental training pool. The Elastic Weight Consolidation (EWC) incremental learning method is employed to constrain the core parameters of the historical model, preventing new sample training from overwriting old features and avoiding catastrophic forgetting. During weekly offline iterative training, existing data and new samples are combined to minimize the dynamic weights. Total loss The network parameters are updated in real time. At the same time, the latest soil and rock parameters measured in the field are used to continuously optimize the model recognition accuracy and scene adaptability through a Bayesian inversion correction mechanism model database, forming a complete closed-loop intelligent evolution system.

[0061] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A GIS-based multi-source data fusion system for early identification of geological hazards, comprising a multi-source heterogeneous data management and preprocessing module, a hybrid feature extraction and knowledge graph construction module, a three-driven fusion identification module, a multi-scale spatiotemporal evolution law identification module, and an interpretable risk assessment and dynamic update module, characterized in that: The multi-source heterogeneous data management and preprocessing module is used to collect optical images, SAR data, DEM data, ground survey data, physical and mechanical parameters of soil and rock, hydrogeological data and meteorological and rainfall data of the area to be measured, and to complete unified coordinate system transformation, cross-modal spatial registration, unified spatiotemporal resolution, data cleaning and normalization processing to generate a standardized multi-source raster dataset. The hybrid feature extraction and knowledge graph construction module includes a data-driven feature extraction unit, a mechanism-driven feature extraction unit, and a disaster knowledge graph construction unit. These units extract surface environmental deformation feature maps and soil and rock mechanics mechanism feature maps, respectively, and construct a geological disaster-specific knowledge graph to achieve structured storage and retrieval of expert prior knowledge. The three-drive fusion recognition module is a "mechanism-data-expert knowledge" three-drive fusion recognition module, which is connected to the hybrid feature extraction and knowledge graph construction module. It includes a physically guided feature enhancement subunit, a dynamic weight allocation subunit, and a multi-model integrated learning subunit. It completes the three-drive feature fusion and disaster probability prediction through physical constraint loss function, knowledge-driven dynamic weight algorithm, and constrained Bayesian fusion algorithm. The multi-scale spatiotemporal evolution law identification module is connected to the three-drive fusion identification module. It is used to realize three-level spatial scale drilling analysis at the provincial, county, and individual hidden danger point levels, match multi-resolution data, delineate high-risk areas through adaptive thresholds, and monitor the disaster evolution stage in combination with time-series deformation data. The interpretability risk assessment and dynamic update module is connected to the multi-scale spatiotemporal evolution law identification module. It is used to generate an interpretable risk report with a chain of evidence and to achieve closed-loop iterative optimization of the model through incremental learning and parameter inversion.

2. The GIS-based multi-source data fusion geological disaster early identification system according to claim 1, characterized in that, The physically guided feature enhancement subunit is configured with a physically consistent regularization loss function, the specific formula of which is: In the formula, A two-dimensional raster map for predicting disaster probabilities using a network. This is a two-dimensional raster map of normalized slope stability coefficients. Cov is a local sliding window covariance operation, which calculates the covariance region by region using a fixed-size sliding window and then takes the mean, preserving spatial neighborhood structure information; the total loss function is... , To supervise loss using data-driven methods, For dynamically adaptive balancing hyperparameters.

3. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 2, characterized in that, The dynamic adaptive equilibrium hyperparameter The calculation method is as follows: , For taking values The smallest constant is chosen to adapt to the precision of computer floating-point operations, avoid division by zero in the denominator, and not interfere with the model gradient update, thus preventing the denominator from being zero; the ReLU(-Cov) operation logic is equivalent to A penalty constraint is applied only when the predicted probability is negatively correlated with the physical stability coefficient; the penalty value is 0 when they are positively correlated.

4. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 1, characterized in that, The disaster knowledge graph construction unit adopts a semi-automatic construction method of top-down ontology definition and bottom-up data extraction. The core nodes include four categories: disaster-inducing environment, disaster-causing factors, disaster type, and evolution stage. Node attributes include geological age, lithological combination, landform type, average annual rainfall, seismic intensity, and historical disaster frequency. Semantic relationships of inducing, aggravating, inhibiting, and matching are configured between nodes. The knowledge graph is deployed on the Neo4j graph database and outputs an expert prior weight vector adapted to the area to be tested through a hybrid retrieval strategy of GIS spatial range retrieval and geological attribute matching.

5. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 1, characterized in that, The fusion weight formula for the dynamic weight allocation subunit is: In the formula, C represents the total number of feature channels. Data-driven weight vectors generated for the attention mechanism. , where is the expert prior weight vector for knowledge graph retrieval, and both are logarithmic field vectors; and , The confidence level of remote sensing data is calculated by the Sigmoid function, and the optimized feature map is finally generated by weighting the original feature map channel by channel by fusing the weights.

6. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 1, characterized in that, The constrained Bayesian fusion formula for the multi-model ensemble learning subunit is as follows: In the formula, The output probabilities are for the data-driven branch and the mechanism-driven branch, respectively, and the boundaries are pruned to... interval; The divergence penalty coefficient was optimized based on samples from 12 typical disaster-prone small watersheds in the southwestern mountainous region and a sensitivity experiment of grid search parameters. The threshold for the lower limit of fusion probability was determined through experiments with multiple sets of extreme divergent samples, taking into account both information preservation and noise suppression effects.

7. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 1, characterized in that, The multi-scale spatiotemporal evolution law identification module is equipped with a multi-scale resolution adaptation mechanism. The provincial-level large area uses 20m resolution data for full-area screening, the county-level key area uses 10m resolution data for fine analysis, and the individual hidden danger point uses 5m resolution data for precise location. The statistical quantile adaptive threshold method is used to delineate high-probability risk areas. By default, the grids with the top 5% probability values ​​are selected as high-risk areas. This threshold is statistically calibrated by the disaster sample of the entire southwestern mountainous area, which can balance the missed report rate and the amount of calculation, and supports dynamic adjustment of the quantile threshold.

8. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 1, characterized in that, The multi-scale spatiotemporal evolution law identification module extracts surface deformation sequences based on time-series InSAR technology. Through an adaptive piecewise fitting method, it monitors the complete evolution stages of uniform deformation, accelerated deformation, and near-slip instability in the potential hazard area in real time, realizing the fusion of static probability identification and dynamic evolution monitoring.

9. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 1, characterized in that, The interpretability risk assessment and dynamic update module adopts elastic weight consolidation as an incremental learning strategy, which applies weight constraints to the core network parameters of historical training tasks to suppress large parameter updates and avoid catastrophic forgetting of the model; it completes incremental iterative training of the network by combining newly added field validation samples, and at the same time corrects the geotechnical parameter database of the mechanism model through Bayesian inversion.

10. The GIS-based multi-source data fusion early identification system for geological hazards according to claim 1, characterized in that, The interpretability risk assessment and dynamic update module extracts key evidence features for disaster identification using attention map back projection technology, associates them with disaster-causing attributes in the knowledge graph, and automatically generates a natural language interpretability risk assessment report with a complete chain of evidence.

Images