A rapid identification method for wide-area landslides based on interpretable intelligent algorithm

Abstract

This invention discloses a method for rapid identification of wide-area landslides based on an interpretable intelligent algorithm, which relates to the field of remote sensing science and technology. By integrating multi-source spatiotemporal data with dynamic knowledge graph weighting, a dual-channel feature extraction architecture is constructed: a lightweight CNN captures local image texture, and a graph convolutional network models geological spatial correlations. SHAP values ​​and causal reasoning are combined to generate interpretable contribution heat maps and rule chains. A knowledge graph bidirectional verification system is introduced, using a priori rules to verify spatial logical contradictions. Misidentified samples trigger the co-evolution of the model and rule base. A multi-dimensional credibility report is output, using Monte Carlo Dropout to quantify uncertainty and customize the interpretation granularity by role. A terrain-adaptive block-streaming processing architecture is employed, combined with lightweight edge deployment and federated learning, to achieve wide-area real-time early warning and dynamic model updates. This solution breaks through the limitations of traditional black-box models and forms a closed-loop disaster prevention system with physical drive, transparent decision-making, and second-level response.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing science and technology, and in particular to a method for rapid identification of wide-area landslides based on an interpretable intelligent algorithm. Background Art

[0002] Landslides are one of the world's major geological hazards. Traditional manual inspections and single-point monitoring are inefficient and have limited coverage. Wide-area rapid identification technology enables large-scale hazard screening, supporting emergency response. Advances in satellite constellations and drone technology have made it possible to monitor surface deformation with high temporal and spatial resolution. For example, the short-baseline DInSAR method can shorten the monitoring cycle to just six days, enabling real-time tracking of landslide risks caused by reservoir water level fluctuations.

[0003] Deep learning demonstrates powerful feature extraction capabilities in remote sensing image processing. For example, the U-Net++ model enables semantic segmentation of landslide boundaries, and the YOLO family of algorithms is used for real-time image detection. However, the "black box" nature of traditional models limits their trustworthiness in disaster decision-making, necessitating the introduction of interpretability techniques. A single data source is susceptible to cloud obstruction or interference from texture similarity. Fusion of multi-dimensional information, such as InSAR deformation, terrain slope, and rainfall data, can improve recognition robustness.

[0004] However, existing technologies still have problems that need to be solved urgently: the main one is the credibility crisis caused by the "black box" problem. The internal decision-making process of the current mainstream deep learning model is highly opaque, and the reasons why the model identifies an area as a landslide, the characteristics based on which it is based, and its reliability cannot be accurately answered. This leads to a lack of trust in the model results among end users such as geological experts and emergency management personnel, making it difficult to effectively integrate the automated identification results into the actual decision-making process. Summary of the Invention

[0005] In order to solve the above technical problems, a method for rapid identification of wide-area landslides based on an interpretable intelligent algorithm is provided. This technical solution solves the above-mentioned problems of insufficient implicit feature extraction capabilities and difficulty in multi-source data fusion.

[0006] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0007] A wide-area landslide rapid identification method based on an interpretable intelligent algorithm includes:

[0008] S1. Construct a multi-source spatiotemporal dataset based on satellite remote sensing images, high-precision terrain data, geological structure maps, historical landslide databases, and real-time meteorological data. Dynamically generate feature weights through the geoscience knowledge graph to normalize and enhance the input data.

[0009] S2 uses a dual-channel feature extractor. The first channel uses a lightweight convolutional neural network to extract local texture features of remote sensing images, and the second channel uses a graph convolutional network to model the spatial correlation between terrain and geological elements. It integrates SHAP values ​​and causal inference algorithms to generate a contribution heat map and logical rule chain for each prediction result.

[0010] S3. Establish a bidirectional verification system between landslide identification results and the geoscience knowledge graph. Use the prior constraints in the knowledge graph to perform spatial logic verification on the suspected landslide areas output by the model. When the threshold is triggered by the misjudgment sample marked by the user, the model fine-tuning module is automatically activated and the dynamic rule base in the knowledge graph is synchronously updated.

[0011] S4. Generate a visual report containing multi-dimensional credibility indicators, calculate the standard deviation of landslide probability prediction through Monte Carlo Dropout, and mark low-confidence areas; customize the interpretation granularity according to user role, and output feature contribution ranking, rule chain, and historical similar case comparison;

[0012] S5, wide-area real-time computing, uses a block-streaming processing architecture to achieve rapid identification of large-scale areas; dynamically adjusts the image block size based on GPU memory capacity and terrain complexity, giving priority to high-risk areas; deploys lightweight models on drones and ground monitoring stations, and uses federated learning to achieve real-time edge inference and synchronized updates of cloud models.

[0013] Preferably, the S1 specifically includes:

[0014] Satellite remote sensing images include optical images that provide RGB-NIR spectral features to capture vegetation anomalies and surface cracks. SAR data uses InSAR technology to extract millimeter-level surface deformation and monitor landslide creep through clouds and fog. Historical pre- and post-disaster images are time-series registered to construct a deformation rate field.

[0015] High-precision terrain data, generating high-resolution DEM through airborne LiDAR point cloud, extracting micro-topographic features; DSM integrates building and vegetation height;

[0016] Geological structural maps, including digital regional geological maps, are used to extract fault line buffer zones, the occurrence of weak interlayers, and the lithologic erosion resistance coefficient; active fault influence zones are annotated in conjunction with seismic network data;

[0017] A historical landslide database structures the location, scale, and triggering factors of landslides, and establishes a sample-environment correlation matrix. It also labels "pseudo-negative samples," i.e., landslides that were not recorded but actually occurred, to correct for machine learning biases.

[0018] Real-time meteorological data, through access to regional weather stations and GPM satellite rainfall data, calculates the effective rainfall index and soil saturation threshold; dynamically associates the spatial and temporal distribution of rainfall with the slope seepage model;

[0019] The time and space benchmarks are unified. At the spatial level, the WGS84UTM projection coordinate system is used. The geometric deviations of multi-source data are corrected through thin plate spline functions, and the geological map is topologically reconstructed. At the temporal level, a dynamic time window based on rainfall events and earthquake events is established. Meteorological data are interpolated to the time when remote sensing images are acquired at a regular granularity.

[0020] Preferably, the S1 specifically includes:

[0021] Dynamic feature weighting uses a knowledge graph weighting algorithm to calculate the semantic similarity between features and the landslide entity. The landslide entity is embedded with a graph neural network to quantify feature importance and generate feature weights based on the knowledge graph.

[0022] Adaptive scene adjustment to increase the SAR feature weight in vegetation-covered areas;

[0023] Enhance the features of physical constraints and use geoscientific logic filtering to eliminate feature combinations that violate physical laws; strengthen the key disaster factor chain; enhance adversarial resistance by using WassersteinGAN to generate difficult examples that conform to geological laws, and remove highly confusing features through feature confusion detection;

[0024] Domain adaptive normalization was used to divide the study area into geological units and independently calculate the Z-score value of each unit. The meteorological data were transformed using the cumulative distribution function. The InSAR deformation series were subjected to phase unwrapping error correction to constrain the cumulative deformation value to conform to the soil creep equation.

[0025] A multi-source conflict resolution strategy and evidence theory fusion are adopted. When there is a contradiction between optical imagery and SAR deformation data, the DS evidence theory is used to calculate the confidence interval. Uncertainty propagation modeling is used to quantify the impact of DEM errors on slope calculation, and an error band indicator is added to the eigenvector.

[0026] Preferably, the S2 specifically includes:

[0027] The first channel uses a lightweight CNN to extract local texture features. It uses MobileNetV3-Small as the backbone network and compresses the number of parameters through depthwise separable convolution, meeting the deployment requirements of edge devices while maintaining accuracy.

[0028] A multi-scale dilated convolution module was added to capture the cross-scale correlation between the cracking zone at the rear edge of the landslide and the swelling zone at the front edge. A spectral attention mechanism was introduced to weight the near-infrared and short-wave infrared bands to enhance the response to soil moisture anomalies in vegetation-covered areas. The high-dimensional feature tensor was output to preserve texture details.

[0029] Implementing adversarial training of the gradient reversal layer, injecting cloud occlusion and shadow noise samples simultaneously with the optical image input, forcing the network to learn the illumination-invariant landslide spectral fingerprint features;

[0030] The second channel uses a graph convolutional network to model spatial associations and construct a geological feature map. Using a grid of a given size as the basic unit, each node contains terrain factors, geological attributes, and environmental variables. A directed graph is constructed based on hydrological principles, connecting along the slope, calculating the flow direction based on the DEM, and establishing a dependency chain for downstream units. Furthermore, a structural association connection is constructed, forcing bidirectional links between units within a given range on both sides of the fault line to simulate stress transfer effects.

[0031] GCN message passing uses the GraphSAGE sampling and aggregation algorithm to iteratively update node representations:

[0032] Where, is the feature vector of node v at the kth layer, which is the node feature after being updated by the graph convolutional network; is the activation function, used to introduce nonlinearity; is the weight matrix, which is used to perform linear transformation on the concatenated features and learn the weights of the features; For the concatenation operation, the node’s own features and the aggregated neighbor features are combined into one vector; is the feature vector of node v in the k-1th layer, representing the node features of the previous layer; is an aggregation function used to aggregate the features of neighboring nodes. It is a terrain-constrained aggregation function. is the feature vector of neighbor node u in the k−1 layer; is the set of neighbor nodes of node v, representing other nodes connected to node v; the terrain potential energy constraint is embedded in the aggregation function AGG to make the spatial propagation conform to the gravity mechanism;

[0033] Dual-channel feature fusion, dynamic gated fusion module, embedding CNN output feature map and GCN output map, aligning spatial scales through coordinate transformation; strengthening CNN texture features in bedrock exposed areas, and focusing on GCN topological associations in hilly accumulation layer areas.

[0034] Preferably, the S2 specifically includes:

[0035] An explainable decision engine, SHAP value-driven contribution quantification, based on the multimodal SHAP adaptation algorithm, uses DeepSHAP for CNN channels to approximate Shapley values ​​through backpropagation and generate pixel-level heat maps; GraphSHAP is developed for GCN channels to calculate topological importance through node masking experiments;

[0036] Causal reasoning rule chain generation and a two-level rule extraction architecture are implemented. The first stage involves extracting rules from proxy models, predicting probabilities through DNN, fitting a decision tree proxy model, and pruning to obtain a primary rule set. The second stage involves geo-causal enhancement, which involves verifying the causal relationship based on the primary rule set. If the rules match, they are incorporated into the knowledge graph. If there is a conflict, counterfactual corrections are performed.

[0037] The causal verification method, based on the Pearl causal graph model, verifies the independence of the do-operator of the rule antecedent; and verifies the physical rationality of the rule by calling the geomechanical simulator;

[0038] Uncertainty transfer visualization, contribution-confidence coupling, superimposing the probability variance generated by Monte Carlo Dropout on the heat map, and outputting logical rules with support and confidence.

[0039] Preferably, the S3 specifically includes:

[0040] Knowledge graph-driven spatial logic verification includes topological verification of the prior rule base, terrain conflict resolution, and spatial overlay analysis of the landslide boundary output by the model. If the slope of the landslide's rear edge is less than a predetermined angle, or the front edge is outside the river network buffer zone, a terrain conflict alert is automatically triggered, and correction suggestions are generated. Geological attribute matching uses association rules from the knowledge graph to verify the lithology coding of the suspected landslide area. If there is a conflict with the rule base, a manual review process is initiated.

[0041] Verify the association of spatiotemporal events and analyze the coupling of inducing factors. Retrieve meteorological and seismic events before and after the landslide timestamp. If the model prediction results lack the support of inducing factors, they will be marked and a causal chain missing report will be output. Compare the similarity of historical cases. Graph embedding technology is used to calculate the similarity between the current landslide and historical cases in the knowledge graph. If a significant deviation is found, it indicates the possibility of a new landslide mechanism that requires expert intervention.

[0042] Physical process simulation verification, rapid mechanical simulation, integrated simplified limit equilibrium model, calculate the potential sliding surface safety factor based on DEM data. If the safety factor of the high-risk area predicted by the model is greater than the established threshold, it is judged as a physical contradiction and the prediction result is forced to be corrected; hydrological process verification, through the distributed hydrological model, simulates the surface runoff and underground seepage in the landslide area, and detects whether the landslide range predicted by the model is within the seepage saturation zone to eliminate the misjudgment of unsaturated soil.

[0043] Preferably, the S3 specifically includes:

[0044] Multi-dimensional attribution analysis for intelligent diagnosis of misjudgment samples, feature contribution inversion, reverse analysis of misjudgment samples using SHAP values, and identification of the core features that cause errors; rule chain conflict detection, comparing the matching paths of misjudgment samples with knowledge graph rules, and locating invalid rule nodes;

[0045] Counterfactual sample generation: Generative Adversarial Networks (GANs) generate virtual samples that are similar in features to misclassified samples but correctly classified for model comparison training.

[0046] The dynamic rule base is updated online, and the rule credibility decay mechanism adds a dynamic weight to each rule. When a rule fails to be verified multiple times in a row, the weight automatically decays until it is confirmed invalid by manual review. Causal rule mining uses the FCI algorithm to extract new causal relationships from misjudged samples, and after expert review, it is injected into the knowledge graph in the form of RDF triples.

[0047] Incremental transfer learning uses ElasticWeightConsolidation technology to constrain the fluctuation range of important weights when fine-tuning model parameters; knowledge-model bidirectional mapping establishes a linkage mechanism between rule confidence and model layer weights. When the weight of a rule increases, the learning rate of the CNN / GCN layer of the corresponding feature channel increases, accelerating the model's absorption of new knowledge.

[0048] Preferably, the S4 specifically includes:

[0049] Monte Carlo Dropout-driven confidence mapping, reliability calibration of probability prediction, randomly enabling a predetermined proportion of convolution kernel Dropout during the model inference phase, calculating the standard deviation of the landslide probability for each grid cell, and performing threshold grading; uncertainty tracing analysis, and automatically associating data quality reports for low-confidence areas.

[0050] Preferably, the S5 specifically includes:

[0051] The block-streaming processing architecture uses a pyramid tile segmentation strategy, spatially partitioning, and dynamically generating an adaptive tile pyramid based on satellite image resolution and GPU memory capacity. It also uses time-series streaming access to build a multi-level pipeline cache queue, prioritizing real-time SAR data streams and triggering local recalculation when new data arrives.

[0052] Dynamic scheduling driven by terrain complexity defines a terrain complexity index based on a complexity quantification model. A greedy scheduling algorithm dynamically selects a processing strategy based on the remaining video memory. When the video memory exceeds a predetermined ratio, high-risk complex terrain is processed in descending order of the terrain complexity index. When the video memory is less than the predetermined ratio, secondary segmentation is enabled to subdivide areas with high terrain complexity indices.

[0053] Blockchain fingerprint marking technology is used to generate a unique hash value for each block. When the network is interrupted, only the unfinished blocks need to be retransmitted; after an abnormal crash on the edge, a model state snapshot is used.

[0054] Preferably, the S5 specifically includes:

[0055] Lightweight deployment at the edge: the backbone network uses MobileNetV3-Lite, which, after INT8 quantization, can achieve 25fps inference on Jetson Xavier NX. The GCN channel removes low-weight edges through graph pruning technology.

[0056] The federated learning-driven model features secure gradient aggregation. After local training, each edge node only uploads incremental model parameters. The cloud aggregates and updates the global model through Paillier homomorphic encryption. A differentiated learning mechanism focuses on fine-tuning deformation characteristics at nodes in earthquake-prone areas, while nodes in heavy rainfall areas strengthen hydrological coupling training. Incremental update hot deployment allows seamless switching using a blue-green release strategy after the new model passes AB testing. Risk scoring is achieved through the fusion of multi-source data, a dynamic risk index is constructed, and graded warning triggers are implemented.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] This paper proposes a dual-channel architecture, embedding geological prior knowledge into the bottom layer of the model. The lightweight CNN channel uses the spectral attention mechanism to enhance the abnormal response of moisture content and break through the interference of vegetation cover. The GCN channel constructs the spatial correlation of fault zones based on the Coulomb stress transfer formula and accurately captures the hidden structural slip control mechanism. The interpretable output of causal rules integrates multimodal SHAP values ​​and two-level rule refinement: pixel-level thermal maps are generated to locate key precursors such as microcracks, and logical rule chains verified by geomechanical simulation are output, making the decision basis traceable and questionable.

[0059] Through dynamic correction of knowledge graphs, terrain contradictions and rock type conflicts are automatically detected based on the prior rule base to block empirical misjudgments; strong constraints are verified by simulation, and the limit equilibrium model and hydrological-mechanical coupling simulation are introduced to calculate 100,000-level potential sliding surfaces in parallel, mark the areas with conflicting safety factors, verify the spatial consistency of the landslide boundary and the seepage saturated zone, and eliminate false alarms of unsaturated soil.

[0060] Terrain-adaptive edge computing uses dynamic block streaming processing, intelligently segments images according to the terrain complexity index, and uses a greedy scheduling algorithm to prioritize high-risk areas, improving computing resource utilization. The lightweight model on the edge supports 25FPS real-time scoring and achieves secure cross-regional knowledge sharing through homomorphic encryption gradient aggregation. New rules are synchronized across the entire region within 3 hours. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Flowchart of the method for rapid identification of wide-area landslides based on interpretable intelligent algorithm. DETAILED DESCRIPTION

[0062] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0063] Reference Figure 1 As shown in the figure, the wide-area landslide rapid identification method based on interpretable intelligent algorithm includes:

[0064] A wide-area landslide rapid identification method based on an interpretable intelligent algorithm includes:

[0065] S1. Construct a multi-source spatiotemporal dataset based on satellite remote sensing images, high-precision terrain data, geological structure maps, historical landslide databases, and real-time meteorological data. Dynamically generate feature weights through the geoscience knowledge graph to normalize and enhance the input data.

[0066] S2 uses a dual-channel feature extractor. The first channel uses a lightweight convolutional neural network to extract local texture features of remote sensing images, and the second channel uses a graph convolutional network to model the spatial correlation between terrain and geological elements. It integrates SHAP values ​​and causal inference algorithms to generate a contribution heat map and logical rule chain for each prediction result.

[0067] S3. Establish a bidirectional verification system between landslide identification results and the geoscience knowledge graph. Use the prior constraints in the knowledge graph to perform spatial logic verification on the suspected landslide areas output by the model. When the threshold is triggered by the misjudgment sample marked by the user, the model fine-tuning module is automatically activated and the dynamic rule base in the knowledge graph is synchronously updated.

[0068] S4. Generate a visual report containing multi-dimensional credibility indicators, calculate the standard deviation of landslide probability prediction through Monte Carlo Dropout, and mark low-confidence areas; customize the interpretation granularity according to user role, and output feature contribution ranking, rule chain, and historical similar case comparison;

[0069] S5, wide-area real-time computing, uses a block-streaming processing architecture to achieve rapid identification of large-scale areas; dynamically adjusts the image block size based on GPU memory capacity and terrain complexity, giving priority to high-risk areas; deploys lightweight models on drones and ground monitoring stations, and uses federated learning to achieve real-time edge inference and synchronized updates of cloud models.

[0070] It should be noted that the data layer (S1) constructs a "geological digital twin" through multi-source spatiotemporal data to solve the problem of feature fragmentation in traditional methods:

[0071] Satellite remote sensing images capture transient surface signals, such as crack expansion after rain; geological structure maps provide billion-year-scale evolutionary context, such as fault activity; and real-time meteorological data inject dynamic trigger variables, such as critical rainfall thresholds.

[0072] The knowledge layer (S3) uses the geoscience knowledge graph as a physical law verifier to transform expert experience into computable prior rules, such as "bedding slope + heavy rainfall → prone to landslides", constraining the model output to conform to geological logic;

[0073] In the model layer (S2), the dual-channel feature extractor realizes the micro-texture perception (CNN channel recognizes millimeter-level deformation) and the macro-system association (GCN channel models hydrogeological topology), avoiding the cognitive blind spots of a single perspective.

[0074] Depth of interpretability (S2 / S4), SHAP values ​​quantify the contribution of SAR imagery, such as deformation features accounting for 58% of the decision weight; causal rule chains generate actionable instructions, such as "if monthly rainfall > 300mm, initiate inspection";

[0075] A breakthrough in computing efficiency (S5) was achieved, with block-streaming processing reducing the assessment time for an area of ​​300,000 square kilometers to 28 minutes; and federated learning achieved edge (drone) inference latency of less than 2 seconds.

[0076] Closed-loop process: real-time satellite monitoring → S5 priority processing of high-risk areas → S2 generation of rule chains and heat maps → S4 customized reports for emergency commanders → evacuation instructions / engineering disposal → on-site misjudgment marking → S3 knowledge graph and model update → real-time satellite monitoring.

[0077] Said S1 specifically includes:

[0078] Satellite remote sensing images include optical images that provide RGB-NIR spectral features to capture vegetation anomalies and surface cracks. SAR data uses InSAR technology to extract millimeter-level surface deformation and monitor landslide creep through clouds and fog. Historical pre- and post-disaster images are time-series registered to construct a deformation rate field.

[0079] High-precision terrain data, generating high-resolution DEM through airborne LiDAR point cloud, extracting micro-topographic features; DSM integrates building and vegetation height;

[0080] Geological structural maps, including digital regional geological maps, are used to extract fault line buffer zones, the occurrence of weak interlayers, and the lithologic erosion resistance coefficient; active fault influence zones are annotated in conjunction with seismic network data;

[0081] A historical landslide database structures the location, scale, and triggering factors of landslides, and establishes a sample-environment correlation matrix. It also labels "pseudo-negative samples," i.e., landslides that were not recorded but actually occurred, to correct for machine learning biases.

[0082] Real-time meteorological data, through access to regional weather stations and GPM satellite rainfall data, calculates the effective rainfall index and soil saturation threshold; dynamically associates the spatial and temporal distribution of rainfall with the slope seepage model;

[0083] The time and space benchmarks are unified. At the spatial level, the WGS84UTM projection coordinate system is used. The geometric deviations of multi-source data are corrected through thin plate spline functions, and the geological map is topologically reconstructed. At the temporal level, a dynamic time window based on rainfall events and earthquake events is established. Meteorological data are interpolated to the time when remote sensing images are acquired at a regular granularity.

[0084] Said S1 specifically includes:

[0085] Dynamic feature weighting uses a knowledge graph weighting algorithm to calculate the semantic similarity between features and the landslide entity. The landslide entity is embedded with a graph neural network to quantify feature importance and generate feature weights based on the knowledge graph.

[0086] Adaptive scene adjustment to increase the SAR feature weight in vegetation-covered areas;

[0087] Enhance the features of physical constraints and use geoscientific logic filtering to eliminate feature combinations that violate physical laws; strengthen the key disaster factor chain; enhance adversarial resistance by using WassersteinGAN to generate difficult examples that conform to geological laws, and remove highly confusing features through feature confusion detection;

[0088] Domain adaptive normalization was used to divide the study area into geological units and independently calculate the Z-score value of each unit. The meteorological data were transformed using the cumulative distribution function. The InSAR deformation series were subjected to phase unwrapping error correction to constrain the cumulative deformation value to conform to the soil creep equation.

[0089] A multi-source conflict resolution strategy and evidence theory fusion are adopted. When there is a contradiction between optical imagery and SAR deformation data, the DS evidence theory is used to calculate the confidence interval. Uncertainty propagation modeling is used to quantify the impact of DEM errors on slope calculation, and an error band indicator is added to the eigenvector.

[0090] It should be noted that the time and space benchmarks are unified:

[0091] In the spatial dimension, thin plate spline functions were used to correct topological errors in geological maps, such as fault line offsets, so that the matching accuracy between a 1:50,000 scale geological map and 0.5-meter resolution LiDAR data reached sub-pixel level (RMSE < 0.3 meters).

[0092] In the temporal dimension, an event-driven dynamic window is constructed. Using a 72-hour rainstorm event as a benchmark, the Sentinel-1 transit time is automatically aligned with the rainfall peak moment to capture the delayed response of "rainfall intensity-deformation";

[0093] Data enhancement embedded in physical laws and scientific constraints for adversarial generation require WassersteinGAN-generated samples to pass dual-criteria verification:

[0094] Mechanical rationality: the deformation sequence conforms to the Burgers creep model of soil; ground formation cause association: the angle between the crack direction and the slope extracted by DEM is ≤15°;

[0095] Explicit control of error propagation, quantifying the impact of LiDAR point cloud density (>8 points / ㎡) on micro-topography parameters, and appending credible interval markers to the feature vector to avoid model learning distorted signals.

[0096] geological scene Feature weight adaptation rule Scientific basis Bedrock exposed areas Rock erosion resistance coefficient ↑ (0.63) DEM microtopography ↑ (0.58) Hard rock landslides are controlled by the combination of structural surfaces and terrain Deep accumulation area InSAR deformation↑(0.71) Soil saturation↑(0.65) Pore ​​water pressure dominates shear strength decay Urban artificial slopes DSM building load ↑ (0.69) vibration monitoring ↑ (0.52) Human activities change the original stress field

[0097] Intelligent arbitration of multi-source conflicts, three-step decision-making based on DS evidence theory:

[0098] Basic probability distribution, optical crack confidence (m1=0.32) vs SAR deformation confidence (m2=0.81);

[0099] Conflict coefficient calculation: (high conflict);

[0100] Weighted decision-making, calling the "vegetation coverage area deformation priority" rule in the knowledge graph, adopting SAR results and marking optical interference sources.

[0101] The S2 specifically includes:

[0102] The first channel uses a lightweight CNN to extract local texture features. It uses MobileNetV3-Small as the backbone network and compresses the number of parameters through depthwise separable convolution, meeting the deployment requirements of edge devices while maintaining accuracy.

[0103] A multi-scale dilated convolution module was added to capture the cross-scale correlation between the cracking zone at the rear edge of the landslide and the swelling zone at the front edge. A spectral attention mechanism was introduced to weight the near-infrared and short-wave infrared bands to enhance the response to soil moisture anomalies in vegetation-covered areas. The high-dimensional feature tensor was output to preserve texture details.

[0104] Implementing adversarial training of the gradient reversal layer, injecting cloud occlusion and shadow noise samples simultaneously with the optical image input, forcing the network to learn the illumination-invariant landslide spectral fingerprint features;

[0105] The second channel uses a graph convolutional network to model spatial associations and construct a geological feature map. Using a grid of a given size as the basic unit, each node contains terrain factors, geological attributes, and environmental variables. A directed graph is constructed based on hydrological principles, connecting along the slope, calculating the flow direction based on the DEM, and establishing a dependency chain for downstream units. Furthermore, a structural association connection is constructed, forcing bidirectional links between units within a given range on both sides of the fault line to simulate stress transfer effects.

[0106] GCN message passing uses the GraphSAGE sampling and aggregation algorithm to iteratively update node representations: Where, is the feature vector of node v at the kth layer, which is the node feature after being updated by the graph convolutional network; is the activation function, used to introduce nonlinearity; is the weight matrix, which is used to perform linear transformation on the concatenated features and learn the weights of the features; For the concatenation operation, the node’s own features and the aggregated neighbor features are combined into one vector; is the feature vector of node v in the k-1th layer, representing the node features of the previous layer; is an aggregation function used to aggregate the features of neighboring nodes. It is a terrain-constrained aggregation function. is the feature vector of neighbor node u in the k−1 layer; is the set of neighbor nodes of node v, representing other nodes connected to node v; the terrain potential energy constraint is embedded in the aggregation function AGG to make the spatial propagation conform to the gravity mechanism;

[0107] Dual-channel feature fusion, dynamic gated fusion module, embedding CNN output feature map and GCN output map, aligning spatial scales through coordinate transformation; strengthening CNN texture features in bedrock exposed areas, and focusing on GCN topological associations in hilly accumulation layer areas.

[0108] The S2 specifically includes:

[0109] An explainable decision engine, SHAP value-driven contribution quantification, based on the multimodal SHAP adaptation algorithm, uses DeepSHAP for CNN channels to approximate Shapley values ​​through backpropagation and generate pixel-level heat maps; GraphSHAP is developed for GCN channels to calculate topological importance through node masking experiments;

[0110] Causal reasoning rule chain generation and a two-level rule extraction architecture are implemented. The first stage involves extracting rules from proxy models, predicting probabilities through DNN, fitting a decision tree proxy model, and pruning to obtain a primary rule set. The second stage involves geo-causal enhancement, which involves verifying the causal relationship based on the primary rule set. If the rules match, they are incorporated into the knowledge graph. If there is a conflict, counterfactual corrections are performed.

[0111] The causal verification method, based on the Pearl causal graph model, verifies the independence of the do-operator of the rule antecedent; and verifies the physical rationality of the rule by calling the geomechanical simulator;

[0112] Uncertainty transfer visualization, contribution-confidence coupling, superimposing the probability variance generated by Monte Carlo Dropout on the heat map, and outputting logical rules with support and confidence.

[0113] It should be noted that the dual-channel design:

[0114] Lightweight CNN channel:

[0115] The spectral attention mechanism targets soil moisture anomalies in vegetated areas by weighting the near-infrared (NIR) and short-wave infrared (SWIR) bands (with a weight ratio of 3:1), enabling the model to capture moisture content changes of more than 0.5% even in dense vegetation.

[0116] Multi-scale dilated convolution captures the tensile cracks (<1 cm width) at the trailing edge of the landslide with a 32×32 pixel window, while correlating the leading edge bulge area with a large 128×128 receptive field to establish cross-scale deformation correlation.

[0117] Adversarial robust training, with a gradient reversal layer forcing the network to maintain spectral feature invariance (within-class variance < 0.03) even under 70% cloud cover, eliminating false positives caused by hillshades.

[0118] GCN spatial modeling with geoscientific constraints:

[0119] Hydrological topology construction, generating a directed acyclic graph (DAG) based on the DEM water flow direction, so that node message transmission strictly follows the direction of gravity (the aggregation weight of downstream nodes is 3.2 times that of upstream nodes);

[0120] Fault stress transfer simulation, forced bidirectional connection within a 2 km buffer zone on both sides of the active fault, through the Coulomb stress transfer formula Corrected edge weights, is the change of normal stress, μ is the rock friction coefficient, is the change of tangential stress;

[0121] Terrain constrained aggregation function: Where, is the AGG aggregation function, which is used to aggregate the features of the nodes to generate new node features. In the k-1th layer, the feature vector of node u, k-1 represents the number of layers of the neural network, and u is a node in the graph; is the feature vector of neighbor node u in the k−1 layer; For node u, which belongs to the neighborhood node set of node v, traverse all nodes u in the neighborhood of node v; The denominator is a function of Δz, which is used to calculate the weight. Its value changes with the change of Δz, thus affecting the contribution of each neighborhood node during aggregation. is the potential energy attenuation coefficient, which is set to 0.05 / m and is used to control the influence of the height difference Δz on the weight; is the height difference between adjacent nodes, that is, the height difference between two adjacent nodes in space, which reflects the undulation of the terrain;

[0122] Landform type CNN weights GCN weights Scientific basis Bedrock exposed areas 0.78 0.22 Surface texture is sensitive to rock mass structure Sedimentary hills 0.35 0.65 Hydrogeological correlation dominates stability Urban artificial slopes 0.60 0.40 Building loads change local stress fields

[0123] Coordinate transformation alignment, using thin plate spline transformation to map GCN graph nodes to CNN feature map coordinates, with a spatial registration error of <0.5 pixels.

[0124] Explainability Engine:

[0125] SHAP value-driven decision-making perspective, pixel-level attribution (DeepSHAP), reverse tracing landslide prediction results to specific pixels in Sentinel-2 imagery to locate potential tear points;

[0126] Topological importance (GraphSHAP), by masking fault zone nodes, was found to contribute 68% to the prediction of a landslide in Ganzi (far exceeding the slope factor of 21%), revealing the hidden structural control mechanism;

[0127] Closed-loop verification of the causal rule chain and two-level rule extraction architecture:

[0128] Surrogate model pruning, using deep forest to fit the DNN decision boundary and extract IF-THEN rules, such as "IF slope > 35° AND moon deformation > 15mm THEN risk > 0.9";

[0129] Enhanced geo-causality and physical rationality testing: FLAC3D mechanical simulator is used to verify whether the rule antecedent results in a safety factor < 1.0; counterfactual correction: when the rule "heavy rainfall induces landslides" fails in arid areas, the "IF irrigation leakage = TRUE" condition is added;

[0130] Uncertainty visualization: Overlay probability variance contours (generated by Monte Carlo Dropout) on the contribution heatmap to mark low-confidence decision areas.

[0131] Rule-graph bidirectional drive:

[0132] New rules are injected. When a causal rule passes the Pearldo-operator test, such as P(landslide|do(rainfall=300mm))>0.7, it is automatically converted into an RDF triple and stored in the graph;

[0133] Conflict resolution mechanism: if the rule conflicts with the "rock landslide requires the development of structural planes" in the atlas, the counterfactual generator is activated to create virtual samples (adjusting the rock layer inclination to a reasonable range), triggering model fine-tuning;

[0134] Edge-Cloud Interpretation Synchronization:

[0135] Lightweight interpretation generation, with the mobile app outputting simplified rule chains, such as steep slope + crack → high risk; the command center platform displays dynamic mechanical simulation graphics, such as the visualization of stress propagation to weak interlayers;

[0136] Federated learning is supported, where edge devices only upload rule confidence (not raw data), and the cloud aggregates and updates the global explanation model.

[0137] The S3 specifically includes:

[0138] Knowledge graph-driven spatial logic verification includes topological verification of the prior rule base, terrain conflict resolution, and spatial overlay analysis of the landslide boundary output by the model. If the slope of the landslide's rear edge is less than a predetermined angle, or the front edge is outside the river network buffer zone, a terrain conflict alert is automatically triggered, and correction suggestions are generated. Geological attribute matching uses association rules from the knowledge graph to verify the lithology coding of the suspected landslide area. If there is a conflict with the rule base, a manual review process is initiated.

[0139] Verify the association of spatiotemporal events and analyze the coupling of inducing factors. Retrieve meteorological and seismic events before and after the landslide timestamp. If the model prediction results lack the support of inducing factors, they will be marked and a causal chain missing report will be output. Compare the similarity of historical cases. Graph embedding technology is used to calculate the similarity between the current landslide and historical cases in the knowledge graph. If a significant deviation is found, it indicates the possibility of a new landslide mechanism that requires expert intervention.

[0140] Physical process simulation verification, rapid mechanical simulation, integrated simplified limit equilibrium model, calculate the potential sliding surface safety factor based on DEM data. If the safety factor of the high-risk area predicted by the model is greater than the established threshold, it is judged as a physical contradiction and the prediction result is forced to be corrected; hydrological process verification, through the distributed hydrological model, simulates the surface runoff and underground seepage in the landslide area, and detects whether the landslide range predicted by the model is within the seepage saturation zone to eliminate the misjudgment of unsaturated soil.

[0141] The S3 specifically includes:

[0142] Multi-dimensional attribution analysis for intelligent diagnosis of misjudgment samples, feature contribution inversion, reverse analysis of misjudgment samples using SHAP values, and identification of the core features that cause errors; rule chain conflict detection, comparing the matching paths of misjudgment samples with knowledge graph rules, and locating invalid rule nodes;

[0143] Counterfactual sample generation: Generative Adversarial Networks (GANs) generate virtual samples that are similar in features to misclassified samples but correctly classified for model comparison training.

[0144] The dynamic rule base is updated online, and the rule credibility decay mechanism adds a dynamic weight to each rule. When a rule fails to be verified multiple times in a row, the weight automatically decays until it is confirmed invalid by manual review. Causal rule mining uses the FCI algorithm to extract new causal relationships from misjudged samples, and after expert review, it is injected into the knowledge graph in the form of RDF triples.

[0145] Incremental transfer learning uses ElasticWeightConsolidation technology to constrain the fluctuation range of important weights when fine-tuning model parameters; knowledge-model bidirectional mapping establishes a linkage mechanism between rule confidence and model layer weights. When the weight of a rule increases, the learning rate of the CNN / GCN layer of the corresponding feature channel increases, accelerating the model's absorption of new knowledge.

[0146] It should be noted that the spatial logic check:

[0147] Topographic contradictions are resolved. When the model outputs a landslide trailing edge slope of <15° (empirical threshold), a 3D slope structure inversion is triggered. This involves reconstructing the sliding bed geometry based on the LiDAR point cloud. If a potential curved sliding surface (radius of curvature > 50m) is detected, it is corrected to a deep landslide and the boundary is adjusted.

[0148] Lithology rule verification: When the "sandstone-shale interbeds → prone to bedding sliding" rule in the knowledge graph conflicts with the model prediction, AI recognition of thin-section microscopic images is initiated, and mineral composition analysis (quartz content <35%) confirms the lithology coding error, and dynamically updates the regional geological map database;

[0149] Physical simulation verification:

[0150] Fast limit equilibrium calculations, integrating the Bishop simplified algorithm, parallelize 100,000 potential slip surfaces on a GPU. If the safety factor Fs in the high-risk area of ​​the model is greater than 1.3, a counterfactual seepage field generator (with a 2m increase in pressure head) is called for reassessment to capture hidden hydrological factors.

[0151] For hydrological-mechanical coupling verification, TOUGH-FLAC co-simulation is used. When the prediction range deviates from the seepage saturation zone (saturation Sr < 85), the InSAR deformation interpretation parameters are adjusted inversely, such as tightening the atmospheric delay error tolerance from 3mm to 1mm to eliminate false alarms in dry areas.

[0152] Multi-dimensional attribution diagnosis engine:

[0153] SHAP reverse tracing, focusing on a misjudgment case at a tunnel entrance on the Sichuan-Tibet Railway in 2025, found that the shortwave infrared band weight was abnormally high (0.78). This was traced back to spectral interference from construction dust, triggering adversarial training to add a new "dust-reflectivity" interference sample library;

[0154] A rule chain conflict map is constructed to build a failed rule propagation network. When the "fault impact zone → high risk" rule fails three times in a row, a newly discovered strike-slip fault (not included in the knowledge map) is located and a drone lidar scan is initiated to confirm the fault extension.

[0155] Counterfactual Causal Mining:

[0156] The FCI algorithm-driven causal discovery was used to extract the groundwater level hysteresis response rules (do(daily seepage > 100m 3 )→P(landslide|30-day lag)=0.67), expanding the traditional rainfall-dominated causal chain;

[0157] GAN virtual samples were generated to create a counterfactual scenario of "faults but no landslides", and the model was trained to distinguish necessary causality (structural surface penetration >80%) from relevant noise (topography undulation), increasing the accuracy of causal rules from 72% to 91%.

[0158] Flexible rule weight mechanism:

[0159] Credibility decay curve, rule weight , is the initial rule weight, that is, the weight value of the rule at the starting time t=0; is the attenuation coefficient, which is used to control the speed at which the weight decays over time; t is the time variable, which represents the time elapsed; Positively correlated with the number of verification failures, such as the "slope > 35° → high risk" rule due to the special case of moraine soil, Adjust from 0.01 to 0.05;

[0160] Cross-platform rule synchronization: when a provincial platform discovers a new rule of "mine blasting vibration → joint activation", it is injected into the national map through federal knowledge distillation, and the simultaneous update of nodes in 28 provinces is completed within 24 hours.

[0161] Model-knowledge bidirectional mapping:

[0162] Parameter elasticity constraint, using ElasticWeightConsolidation technology, locks the spectral attention layer parameters (Fisher information > 0.8) when fine-tuning CNN channels to prevent catastrophic forgetting of historical landslide features.

[0163] The learning rate is dynamically scaled. When the weight of the "strong earthquake induced" rule increases, the learning rate of the corresponding GCN fault channel increases from 0.001 to 0.005, and the model absorbs new knowledge 3 times faster.

[0164] The S4 specifically includes:

[0165] Monte Carlo Dropout-driven confidence mapping, reliability calibration of probability prediction, randomly enabling a predetermined proportion of convolution kernel Dropout during the model inference phase, calculating the standard deviation of the landslide probability for each grid cell, and performing threshold grading; uncertainty tracing analysis, and automatically associating data quality reports for low-confidence areas.

[0166] It should be noted that Monte Carlo Dropout:

[0167] To scientifically calibrate the reliability of probability, we use a random perturbation inference mechanism. During the model inference phase, 50 random dropout samples (with a convolution kernel dropout rate of 15%) are used to generate the landslide probability distribution P∼N(μ,σ2). The standard deviation σ becomes the core indicator for quantifying uncertainty:

[0168] σ<0.05→high confidence zone (dark green, decision weight >90%);

[0169] 0.05≤σ≤0.15→medium confidence zone (yellow, additional investigation required);

[0170] σ>0.15→low confidence zone (red, mandatory manual review); terrain adaptive threshold classification, setting dynamic confidence threshold for complex mountainous areas:

[0171] Landform type High confidence threshold Scientific basis Bedrock steep slope σ<0.08 Surface features are clear and easy to identify Gentle slope of accumulation layer σ<0.12 Hydrological processes increase the difficulty of interpretation Urban artificial slopes σ<0.10 Building occlusion leads to incomplete data

[0172] Multi-role explanation:

[0173] The provincial command center receives the rule chain priority ranking and supports resource scheduling decisions;

[0174] County and city emergency response teams obtain historical similar cases for comparison and reference historical response plans;

[0175] Geological experts review characteristic contribution heat maps and overlay InSAR deformation fields to locate precursors to rupture in locked sections.

[0176] Visualization of causal rules:

[0177] Dynamic logic chain generation, converting the node importance output by GraphSHAP into an executable instruction chain:

[0178] If the fault is within the affected zone (≤2 km) and the monthly cumulative deformation is >30 mm, then start borehole tiltmeter monitoring (92% confidence level).

[0179] The counterfactual warning module generates a virtual scenario when the rule antecedent is missing but high risk is predicted: "If there is an unexplored weak interlayer here, the risk probability will increase from 0.35 to 0.91."

[0180] The S5 specifically includes:

[0181] The block-streaming processing architecture uses a pyramid tile segmentation strategy, spatially partitioning, and dynamically generating an adaptive tile pyramid based on satellite image resolution and GPU memory capacity. It also uses time-series streaming access to build a multi-level pipeline cache queue, prioritizing real-time SAR data streams and triggering local recalculation when new data arrives.

[0182] Dynamic scheduling driven by terrain complexity defines a terrain complexity index based on a complexity quantification model. A greedy scheduling algorithm dynamically selects a processing strategy based on the remaining video memory. When the video memory exceeds a predetermined ratio, high-risk complex terrain is processed in descending order of the terrain complexity index. When the video memory is less than the predetermined ratio, secondary segmentation is enabled to subdivide areas with high terrain complexity indices.

[0183] Blockchain fingerprint marking technology is used to generate a unique hash value for each block. When the network is interrupted, only the unfinished blocks need to be retransmitted; after an abnormal crash on the edge, a model state snapshot is used.

[0184] The S5 specifically includes:

[0185] Lightweight deployment at the edge: the backbone network uses MobileNetV3-Lite, which, after INT8 quantization, can achieve 25fps inference on Jetson Xavier NX. The GCN channel removes low-weight edges through graph pruning technology.

[0186] The federated learning-driven model features secure gradient aggregation. After local training, each edge node only uploads incremental model parameters. The cloud aggregates and updates the global model through Paillier homomorphic encryption. A differentiated learning mechanism focuses on fine-tuning deformation characteristics at nodes in earthquake-prone areas, while nodes in heavy rainfall areas strengthen hydrological coupling training. Incremental update hot deployment allows seamless switching using a blue-green release strategy after the new model passes AB testing. Risk scoring is achieved through the fusion of multi-source data, a dynamic risk index is constructed, and graded warning triggers are implemented.

[0187] It should be noted that the block-streaming processing architecture:

[0188] Terrain-aware pyramid tile strategy:

[0189] Intelligent spatial partitioning based on terrain complexity index (α=0.7,β=0.3α), dynamically generate non-uniform tiles:

[0190] Alpine valley area (C>0.8): tile size 256×256 pixels (sub-meter analysis required);

[0191] Plain and hilly areas (C < 0.3): Tile size 1024 × 1024 pixels (taking efficiency into consideration); Greedy scheduling algorithm:

[0192] In the sufficient memory mode (>80%), high-risk areas with C>0.6 (such as fault zones and steep slopes) are prioritized.

[0193] In memory-intensive mode (<40%), the high-C area is split into 64×64 pixels to ensure continuous operation of Jetson Xavier NX.

[0194] Blockchain fingerprint tracing generates a SHA-3-256 hash fingerprint for each tile, and only retransmits unconfirmed blocks after a network interruption;

[0195] Model snapshot disaster recovery saves gradient compression snapshots every 5 minutes (only retains weights with Fisher information > 0.5) and restores to the most recent state within 10 seconds after an edge device crashes.

[0196] The lightweight model and extremely compressed backbone network make the MobileNetV3-Lite model only 3.7MB after INT8 quantization, and it maintains an inference speed of 25FPS even in a -20°C plateau environment.

[0197] Graph convolution pruning technology removes edges with a contribution of less than 0.05, reducing the number of GCN parameters from 2.1 million to 370,000 and optimizing inference latency from 350ms to 45ms.

[0198] Multi-sensor hard fusion directly fuses SAR deformation sequence + rain gauge data (not raw data uploaded) at the edge to generate a real-time risk index, reducing bandwidth usage by 90%.

[0199] Federated learning-driven models, local training emphasis in different regions and cloud aggregation strategies:

[0200] Region Type Local training focuses on Cloud aggregation strategy seismically active zone InSAR deformation feature fine-tuning Elastic Weighted Average (EWA) Typhoon-affected areas Short-duration rainfall intensity-seepage coupling modeling Gradient Clipping + Differential Privacy permafrost degradation zone Thermal melt slip time series pattern recognition Knowledge graph rule injection

[0201] The secure aggregation architecture uses Paillier homomorphic encryption to upload parameter increments and the global model update equations after cloud aggregation.

[0202] Implementation steps:

[0203] Step 1: Data fusion and dynamic weighting:

[0204] Access real-time data streams from satellite remote sensing (optical / SAR), airborne LiDAR, geological maps, meteorological stations, etc., unify the spatial reference through WGS84UTM projection, call the geoscience knowledge map to calculate feature weights, and use Z-score partition normalization to eliminate dimensional differences;

[0205] Enable physical constraint enhancement: Use WassersteinGAN to generate deformation sequences that conform to the Burgers creep model, and eliminate invalid combinations with DEM slopes > 75° (which violate the rock mass repose angle law);

[0206] Step 2: Lightweight CNN channels:

[0207] A MobileNetV3-Small backbone network was deployed, loaded with INT8 quantized weights, and multi-scale dilated convolution was used to capture the association between cracks (local 32×32 window) and bulges (global 128×128 window). A spectral attention mechanism was also implemented, and the near-infrared band (875nm) was weighted 3x to enhance the moisture content response.

[0208] Step 3: Geo-constrained GCN channels:

[0209] A hydrological topology map was constructed, using 30m grids as nodes. Directed edges were generated based on the DEM flow direction, and fault stress transfer rules were embedded. Nodes within 2km on both sides of active faults were forced to be bidirectionally connected. Edge weights were corrected according to Coulomb's formula. GraphSAGE aggregation was run, and node features were iteratively updated. The aggregation function embedded the height difference potential energy constraint.

[0210] Step 4: Dynamic feature fusion:

[0211] Thin plate spline transformation is used to align the CNN feature map with the GCN node spatial coordinates. The weight is adjusted according to the landform type. The CNN weight is 0.78 in the bedrock area and the GCN weight is 0.65 in the accumulation layer area.

[0212] Step 5: Space logic verification:

[0213] Terrain conflict detection automatically rejects predictions where the trailing edge slope is less than 15° and the leading edge exceeds the river network buffer zone. Lithology rule matching verifies the consistency between the lithology coding of the predicted area and the "sandstone → bedding sliding" rule in the knowledge graph.

[0214] Step 6: Physical simulation verification:

[0215] The Bishop limit equilibrium model was used to parallelly calculate 100,000 potential sliding surfaces, marking the conflicting areas with Fs>1.3. The TOUGH-FLAC hydrological coupling simulation was then launched to verify the spatial coincidence between the landslide boundary and the seepage saturation zone (Sr>85%).

[0216] Step 7: Feedback loop for misjudged samples:

[0217] SHAP reverse attribution locates the core features that lead to misjudgment, generates counterfactual samples, and creates a virtual scenario of "weak interlayer" to correct model deviations. Dynamic rule base updates use the FCI algorithm to discover new rules, which are then injected into the graph after expert review.

[0218] Step 8: Uncertainty Quantification:

[0219] Run 50 Monte Carlo Dropout samplings to generate a landslide probability standard deviation σ layer:

[0220] σ<0.05: dark green high confidence zone;

[0221] σ>0.15: red low confidence zone (automatically associated DEM error>2m and other traceability reports);

[0222] Step 9: Character Customization Explanation:

[0223] The provincial command center pushes the rule chain priority; the county and city emergency teams output historical similar cases; and geological experts overlay SHAP thermal maps and InSAR deformation fields to locate the rupture point of the locked section.

[0224] Step 10: Counterfactual decision making:

[0225] Generate virtual instructions. If there are undiscovered weak interlayers, the risk probability will increase from 0.35 to 0.91, guiding additional exploration.

[0226] Step 11: Chunking - Streaming:

[0227] Dynamically divide the terrain into blocks according to the terrain complexity index:

[0228] C>0.8: 256×256 pixels (mountain valleys); C<0.3: 1024×1024 pixels (plains and hills);

[0229] Greedy scheduling: When the video memory usage is greater than 80%, the high-risk area of ​​C>0.6 is prioritized;

[0230] Step 12: Edge Lightweight Deployment:

[0231] Jetson Xavier NX loaded a 3.7MB quantized model, achieving 25FPS real-time inference, and graph pruning to remove edges with a contribution of <0.05 (the number of GCN parameters was reduced to 370,000);

[0232] Step 13: Federated Learning Evolution:

[0233] After local training at the edge node, parameter increments are uploaded using Paillier homomorphic encryption. The cloud is aggregated according to regional characteristics, with fine-tuning of deformation characteristics focused on earthquake zones and enhanced hydrological coupling modeling in rainstorm areas. After the new model undergoes A / B testing, a blue-green release is implemented to achieve hot updates within 4 minutes.

[0234] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A wide-area landslide rapid identification method based on an interpretable intelligent algorithm is characterized by: include: S1. Construct a multi-source spatiotemporal dataset based on satellite remote sensing images, high-precision terrain data, geological structure maps, historical landslide databases, and real-time meteorological data. Dynamically generate feature weights through the geoscience knowledge graph to normalize and enhance the input data. S2 uses a dual-channel feature extractor. The first channel uses a lightweight convolutional neural network to extract local texture features of remote sensing images, and the second channel uses a graph convolutional network to model the spatial correlation between terrain and geological elements. It integrates SHAP values ​​and causal inference algorithms to generate a contribution heat map and logical rule chain for each prediction result. S3. Establish a bidirectional verification system between landslide identification results and the geoscience knowledge graph. Use the prior constraints in the knowledge graph to perform spatial logic verification on the suspected landslide areas output by the model. When the threshold is triggered by the misjudgment sample marked by the user, the model fine-tuning module is automatically activated and the dynamic rule base in the knowledge graph is synchronously updated. S4. Generate a visual report containing multi-dimensional credibility indicators, calculate the standard deviation of landslide probability prediction through Monte Carlo Dropout, and mark low confidence areas; Customize explanation granularity based on user roles, output feature contribution ranking, rule chains, and historical similar case comparisons; S5, wide-area real-time computing, uses a block-streaming processing architecture to achieve rapid identification of large-scale areas; dynamically adjusts the image block size based on GPU memory capacity and terrain complexity, giving priority to high-risk areas; deploys lightweight models on drones and ground monitoring stations, and uses federated learning to achieve real-time edge inference and synchronized updates of cloud models.

2. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 1 is characterized in that: The S2 specifically includes: The first channel uses a lightweight CNN to extract local texture features. It uses MobileNetV3-Small as the backbone network and compresses the number of parameters through depthwise separable convolution, meeting the deployment requirements of edge devices while maintaining accuracy. A multi-scale dilated convolution module was added to capture the cross-scale correlation between the cracking zone at the rear edge of the landslide and the swelling zone at the front edge. A spectral attention mechanism was introduced to weight the near-infrared and short-wave infrared bands to enhance the response to soil moisture anomalies in vegetation-covered areas. The high-dimensional feature tensor was output to preserve texture details. Implementing adversarial training of the gradient reversal layer, injecting cloud occlusion and shadow noise samples simultaneously with the optical image input, forcing the network to learn the illumination-invariant landslide spectral fingerprint features; The second channel uses a graph convolutional network to model spatial associations and construct a geological feature map. Using a grid of a given size as the basic unit, each node contains terrain factors, geological attributes, and environmental variables. A directed graph is constructed based on hydrological principles, connecting along the slope, calculating the flow direction based on the DEM, and establishing a dependency chain for downstream units. Furthermore, a structural association connection is constructed, forcing bidirectional links between units within a given range on both sides of the fault line to simulate stress transfer effects. GCN message passing uses the GraphSAGE sampling and aggregation algorithm to iteratively update node representations; Dual-channel feature fusion, dynamic gated fusion module, embedding CNN output feature map and GCN output map, aligning spatial scales through coordinate transformation; strengthening CNN texture features in bedrock exposed areas, and focusing on GCN topological associations in hilly accumulation layer areas.

3. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 2 is characterized in that: The S2 specifically includes: An explainable decision engine, SHAP value-driven contribution quantification, based on the multimodal SHAP adaptation algorithm, uses DeepSHAP for CNN channels to approximate Shapley values ​​through backpropagation and generate pixel-level heat maps; GraphSHAP is developed for GCN channels to calculate topological importance through node masking experiments; Causal reasoning rule chain generation and a two-level rule extraction architecture are implemented. The first stage involves extracting rules from proxy models, predicting probabilities through DNN, fitting a decision tree proxy model, and pruning to obtain a primary rule set. The second stage involves geo-causal enhancement, which involves verifying the causal relationship based on the primary rule set. If the rules match, they are incorporated into the knowledge graph. If there is a conflict, counterfactual corrections are performed. The causal verification method, based on the Pearl causal graph model, verifies the independence of the do-operator of the rule antecedent; and verifies the physical rationality of the rule by calling the geomechanical simulator; Uncertainty transfer visualization, contribution-confidence coupling, superimposing the probability variance generated by Monte Carlo Dropout on the heat map, and outputting logical rules with support and confidence.

4. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 3 is characterized in that: Said S1 specifically includes: Satellite remote sensing images include optical images that provide RGB-NIR spectral features to capture vegetation anomalies and surface cracks. SAR data uses InSAR technology to extract millimeter-level surface deformation and monitor landslide creep through clouds and fog. Historical pre- and post-disaster images are time-series registered to construct a deformation rate field. High-precision terrain data, generating high-resolution DEM through airborne LiDAR point cloud, extracting micro-topographic features; DSM integrates building and vegetation height; Geological structural maps, including digital regional geological maps, are used to extract fault line buffer zones, the occurrence of weak interlayers, and the lithologic erosion resistance coefficient; active fault influence zones are annotated in conjunction with seismic network data; A historical landslide database structures the location, scale, and triggering factors of landslides, establishing a sample-environment correlation matrix. This database also annotates "pseudo-negative samples," i.e., landslides that were not recorded but actually occurred, to correct for machine learning biases. Real-time meteorological data, through access to regional weather stations and GPM satellite rainfall data, calculates the effective rainfall index and soil saturation threshold; dynamically associates the spatial and temporal distribution of rainfall with the slope seepage model; The time and space benchmarks are unified. At the spatial level, the WGS84UTM projection coordinate system is used. The geometric deviations of multi-source data are corrected through thin plate spline functions, and the geological map is topologically reconstructed. At the temporal level, a dynamic time window based on rainfall events and earthquake events is established. Meteorological data are interpolated to the time when remote sensing images are acquired at a regular granularity.

5. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 4 is characterized in that: Said S1 specifically includes: Dynamic feature weighting uses a knowledge graph weighting algorithm to calculate the semantic similarity between features and the landslide entity. The landslide entity is embedded with a graph neural network to quantify feature importance and generate feature weights based on the knowledge graph. Adaptive scene adjustment to increase the SAR feature weight in vegetation-covered areas; Enhance the features of physical constraints and use geoscientific logic filtering to eliminate feature combinations that violate physical laws; strengthen the key disaster factor chain; enhance adversarial resistance by using WassersteinGAN to generate difficult examples that conform to geological laws, and remove highly confusing features through feature confusion detection; Domain adaptive normalization was used to divide the study area into geological units and independently calculate the Z-score value of each unit. The meteorological data were transformed using the cumulative distribution function. The InSAR deformation series were subjected to phase unwrapping error correction to constrain the cumulative deformation value to conform to the soil creep equation. A multi-source conflict resolution strategy and evidence theory fusion are adopted. When there is a contradiction between optical imagery and SAR deformation data, the DS evidence theory is used to calculate the confidence interval. Uncertainty propagation modeling is used to quantify the impact of DEM errors on slope calculation, and an error band indicator is added to the eigenvector.

6. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 5 is characterized in that: The S3 specifically includes: Knowledge graph-driven spatial logic verification includes topological verification of the prior rule base, terrain conflict resolution, and spatial overlay analysis of the landslide boundary output by the model. If the slope of the landslide's rear edge is less than a predetermined angle, or the front edge is outside the river network buffer zone, a terrain conflict alert is automatically triggered, and correction suggestions are generated. Geological attribute matching uses association rules from the knowledge graph to verify the lithology coding of the suspected landslide area. If there is a conflict with the rule base, a manual review process is initiated. Verify the association of spatiotemporal events and analyze the coupling of inducing factors. Retrieve meteorological and seismic events before and after the landslide timestamp. If the model prediction results lack the support of inducing factors, they will be marked and a causal chain missing report will be output. Compare the similarity of historical cases. Graph embedding technology is used to calculate the similarity between the current landslide and historical cases in the knowledge graph. If a significant deviation is found, it indicates the possibility of a new landslide mechanism that requires expert intervention. Physical process simulation verification, rapid mechanical simulation, integrated simplified limit equilibrium model, calculate the potential sliding surface safety factor based on DEM data. If the safety factor of the high-risk area predicted by the model is greater than the established threshold, it is judged as a physical contradiction and the prediction result is forced to be corrected; hydrological process verification, through the distributed hydrological model, simulates the surface runoff and underground seepage in the landslide area, and detects whether the landslide range predicted by the model is within the seepage saturation zone to eliminate the misjudgment of unsaturated soil.

7. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 6 is characterized in that: The S3 specifically includes: Multi-dimensional attribution analysis for intelligent diagnosis of misjudgment samples, feature contribution inversion, reverse analysis of misjudgment samples using SHAP values, and identification of the core features that cause errors; rule chain conflict detection, comparing the matching paths of misjudgment samples with knowledge graph rules, and locating invalid rule nodes; Counterfactual sample generation: Generative Adversarial Networks (GANs) generate virtual samples that are similar in features to misclassified samples but correctly classified for model comparison training. The dynamic rule base is updated online, and the rule credibility decay mechanism adds a dynamic weight to each rule. When a rule fails to be verified multiple times in a row, the weight automatically decays until it is confirmed invalid by manual review. Causal rule mining uses the FCI algorithm to extract new causal relationships from misjudged samples, and after expert review, it is injected into the knowledge graph in the form of RDF triples. Incremental transfer learning uses ElasticWeightConsolidation technology to constrain the fluctuation range of important weights when fine-tuning model parameters; knowledge-model bidirectional mapping establishes a linkage mechanism between rule confidence and model layer weights. When the weight of a rule increases, the learning rate of the CNN / GCN layer of the corresponding feature channel increases, accelerating the model's absorption of new knowledge.

8. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 7 is characterized in that: The S5 specifically includes: The block-streaming processing architecture uses a pyramid tile segmentation strategy, spatially partitioning, and dynamically generating an adaptive tile pyramid based on satellite image resolution and GPU memory capacity. It also uses time-series streaming access to build a multi-level pipeline cache queue, prioritizing real-time SAR data streams and triggering local recalculation when new data arrives. Dynamic scheduling driven by terrain complexity defines a terrain complexity index based on a complexity quantification model. A greedy scheduling algorithm dynamically selects a processing strategy based on the remaining video memory. When the video memory exceeds a predetermined ratio, high-risk complex terrain is processed in descending order of the terrain complexity index. When the video memory is less than the predetermined ratio, secondary segmentation is enabled to subdivide areas with high terrain complexity indices. Blockchain fingerprint marking technology is used to generate a unique hash value for each block. When the network is interrupted, only the unfinished blocks need to be retransmitted; after an abnormal crash on the edge, a model state snapshot is used.

9. The wide-area landslide rapid identification method based on an interpretable intelligent algorithm according to claim 8 is characterized in that: The S5 specifically includes: Lightweight deployment at the edge: the backbone network uses MobileNetV3-Lite, which, after INT8 quantization, can achieve 25fps inference on Jetson Xavier NX. The GCN channel removes low-weight edges through graph pruning technology. The federated learning-driven model features secure gradient aggregation. After local training, each edge node only uploads incremental model parameters. The cloud aggregates and updates the global model through Paillier homomorphic encryption. A differentiated learning mechanism focuses on fine-tuning deformation characteristics at nodes in earthquake-prone areas, while nodes in heavy rainfall areas strengthen hydrological coupling training. Incremental update hot deployment allows seamless switching using a blue-green release strategy after the new model passes AB testing. Risk scoring is achieved through the fusion of multi-source data, a dynamic risk index is constructed, and graded warning triggers are implemented.

10. The wide-area landslide rapid identification method based on an explainable intelligent algorithm according to claim 9 is characterized in that: The S4 specifically includes: Monte Carlo Dropout-driven confidence mapping, reliability calibration of probability prediction, randomly enabling a predetermined proportion of convolution kernel Dropout during the model inference phase, calculating the standard deviation of the landslide probability for each grid cell, and performing threshold grading; uncertainty tracing analysis, and automatically associating data quality reports for low-confidence areas.

Images