Geological disaster dynamic early warning method driven by multi-source data and intelligent model

Abstract

The invention discloses a geological disaster dynamic early warning method driven by multi-source data and an intelligent model, and belongs to the technical field of geological disaster early warning. The method comprises the following steps: acquiring multi-source heterogeneous geological monitoring data; performing space-time alignment and semantic standardization processing, and constructing a standardized data set; constructing a multi-modal deep fusion neural network model, extracting features through a parallel feature extraction sub-network, and carrying out feature fusion by adopting a cross-modal attention mechanism based on a historical accuracy rate dynamic adjustment weight; disaster risk prediction is carried out based on fusion features, and online model updating is realized through a sliding time window mechanism; and generating an early warning report according to the risk level. According to the invention, an early warning effect feedback closed loop mechanism is also established, and model iterative optimization is realized through postmortem verification of returned data. According to the method, a complete closed loop from data perception to model self-optimization is constructed, and the problems of poor semantic consistency, static feature fusion and insufficient early warning real-time performance of multi-source data are solved.

Description

A Dynamic Early Warning Method for Geological Disasters Driven by Multi-Source Data and Intelligent Models Technical Field

[0001] This invention belongs to the field of geological disaster early warning technology, specifically involving a dynamic early warning method for geological disasters driven by multi-source data and intelligent models. Background Technology

[0002] With the increasing frequency of geological disasters and the rise in extreme weather events, intelligent early warning systems based on multi-source data fusion have become a key technological direction for improving disaster prevention and mitigation capabilities. Dynamic early warning of geological disasters involves real-time perception and risk assessment of complex natural environmental changes. Its core lies in the efficient integration of data from different observation platforms and the use of advanced algorithms to achieve accurate modeling and prediction of disaster processes such as landslides and debris flows.

[0003] Currently, remote sensing imagery, meteorological observation, UAV aerial photography, and ground surveys are widely used in geological disaster monitoring. For example, patent publication number CN120893013A (hereinafter referred to as "Prior Art 1") discloses a geological disaster early warning method and system based on multi-source data fusion. This method acquires remote sensing data, sensor data, meteorological data, and historical geological data, performs spatiotemporal alignment, noise removal, and missing value imputation to construct a multi-source spatiotemporal data cube. It then employs a large-scale geological disaster prediction model that includes a spatial feature extraction layer, a temporal feature aggregation layer, and disaster classification and regression branches to achieve risk level classification and displacement change value prediction. However, Prior Art 1 suffers from semantic consistency issues when processing multi-source heterogeneous geological data, leading to semantic gaps during feature fusion. Furthermore, its intermodal attention mechanism cannot adaptively balance the contributions of each data source, resulting in insufficient generalization ability when data quality fluctuates. In addition, the model in Prior Art 1 cannot achieve real-time dynamic optimization, leading to a lag in early warning results.

[0004] For example, patent publication number CN120452170A (hereinafter referred to as "Prior Art 2") discloses a geological disaster early warning method and a precise early warning system based on multi-source data fusion. It integrates remote sensing, meteorological, and geological monitoring data, uses principal component analysis (PCA) and recursive feature elimination algorithms for feature extraction, and employs Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) for time series and spatial analysis to achieve disaster risk prediction. However, Prior Art 2 has limited feature extraction capabilities when processing multi-source heterogeneous geological data, making it difficult to effectively capture the nonlinear coupling relationship between surface deformation, precipitation accumulation, and soil stability. Furthermore, its early warning model relies excessively on traditional machine learning methods and lacks a deep fusion mechanism for multimodal data, resulting in poor generalization ability of the early warning results in complex terrain areas. It also suffers from problems of data spatiotemporal alignment and semantic consistency.

[0005] For example, patent publication number CN120448732A (hereinafter referred to as "Prior Art 3") discloses an automatic geological disaster identification system and method based on multi-source remote sensing data. It aims to improve the model's generalization ability across different geographical regions and uncover physical correlations between multi-source remote sensing data through a cross-domain physical fusion module (including domain adversarial network mechanisms and physical constraint mechanisms) and a spatiotemporal evolution decision module (including three-dimensional dilated convolution, dynamic graph evolution, and critical identification layers). However, Prior Art 3 still has significant shortcomings in practical applications: First, its dynamic adaptation mechanism mainly focuses on mitigating the differences in data distribution between different regions. When a single data source experiences noise pollution or temporary interruption, the system lacks explicit, performance-feedback-based intelligent fault tolerance capabilities. Second, its physical correlation mining is mainly limited to the remote sensing data modalities, making it difficult to achieve true multi-source collaborative perception from the perspective of disaster evolution mechanisms. Third, Prior Art 3 emphasizes one-time model optimization and static identification, resulting in insufficient early warning timeliness and long-term reliability in the face of dynamically changing environments.

[0006] It is evident that existing technologies generally suffer from the following problems when processing multi-source heterogeneous geological data: lack of semantic standardization leads to poor data fusion consistency; static feature fusion weight allocation cannot adapt to data changes; and the lack of a real-time dynamic update mechanism results in delayed early warnings. These problems restrict the accuracy and timeliness of geological disaster early warnings. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention provides a dynamic early warning method for geological disasters driven by multi-source data and intelligent models.

[0008] The technical solution adopted in this invention is as follows:

[0009] A dynamic early warning method for geological disasters driven by multi-source data and intelligent models includes the following steps:

[0010] Step 1: Acquire multi-source heterogeneous geological monitoring data, including regional surface deformation information, hourly meteorological data, high-resolution images of key hidden danger areas, and physical parameter data;

[0011] Step 2: Perform spatiotemporal alignment and semantic standardization on multi-source heterogeneous geological monitoring data. Using the geographic coordinate system as a reference, resample the data from different sources to a unified spatial grid, organize the data under a unified temporal reference, and assign unified geological semantic labels to the multi-source heterogeneous geological monitoring data to construct a semantically consistent standardized dataset.

[0012] Step 3: Construct a multimodal deep fusion neural network model consisting of four parallel feature extraction sub-networks and a cross-modal attention mechanism module. This model is used to extract features from the standardized multi-source heterogeneous geological monitoring data and adaptively weight and fuse all feature vectors to output a fused feature vector. The cross-modal attention mechanism module dynamically adjusts the weights based on the historical accuracy of each data source through a confidence generation network to achieve adaptive fusion of multimodal data.

[0013] Step 4: Based on the fused feature vector, identify the disaster evolution state and predict the risk level. Input the fused feature vector into the dual-task output layer with parallel classification head and regression head. The classification head outputs the disaster risk level within a specified time period in the future, and the regression head predicts the landslide displacement rate. The model is updated and the parameters are adjusted online through a sliding time window mechanism.

[0014] Step 5: Generate a dynamic early warning report based on the disaster risk level and trigger the response mechanism.

[0015] Specifically, in step 1, regional surface deformation information is collected through a remote sensing satellite system, hourly meteorological data is obtained through a meteorological observation network, high-resolution images of key hidden danger areas are collected through a drone aerial photography system, and physical parameter data are collected through ground survey equipment deployed at key locations of the landslide body.

[0016] Furthermore, in step 2, the size of the unified spatial grid is 25m×25m; the time reference is Coordinated Universal Time, and it is organized and filled under a unified temporal framework according to the actual update frequency of each data source; the geological semantic tags include "surface displacement", "precipitation intensity", "water content change" and "structural surface activity", and the semantic standardization processing is based on a pre-built geological ontology knowledge base containing multiple types of geological phenomena and parameter relationships for tag mapping and association.

[0017] In this invention, an automatic mapping model is used to map and associate tags with a geological ontology knowledge base that relates various geological phenomena to parameters. Specifically, the automatic mapping model is trained based on the knowledge base rule engine and combined with the correspondence between features and tags in historical data. After training, the input features are semantically analyzed according to the knowledge base rules and automatically mapped to unified geological semantic tags, outputting the corresponding quantitative values.

[0018] Specifically, in step 3, the four parallel feature extraction sub-networks are: a convolutional neural network for processing remote sensing data, a long short-term memory network for processing meteorological data, a residual network for processing UAV imagery, and a fully connected network for processing ground survey data.

[0019] Furthermore, in step 3, the cross-modal attention mechanism module adopts a multi-head attention structure with 8 heads and 128 dimensions for both the query, key, and value vectors.

[0020] Furthermore, in step 3, the process by which the cross-modal attention mechanism module dynamically adjusts the weights based on the historical accuracy of each data source through the confidence generation network is as follows:

[0021] Step a: Calculate the modal confidence factor for each data source using a confidence generation network; the input to the confidence generation network is the historical accuracy of each data source over the past 72 hours, which is calculated based on the statistical results of true positives, true negatives, false positives, and false negatives; the confidence generation network maps the input historical accuracy to a modal confidence factor between 0 and 1 using a fully connected network structure containing hidden layers and an output layer; this confidence generation network is trained and optimized together with the main network during the model training phase;

[0022] Step b: Split the initial fused feature vector output by the cross-modal attention mechanism module into four parts according to the modal source, with each part corresponding to the feature vector of a data source;

[0023] Step c: Multiply each split modal feature vector by its corresponding modal confidence factor to complete the dynamic weight adjustment based on historical accuracy;

[0024] Step d: Fuse the weighted modal feature vectors and obtain the final fused feature vector through linear layer transformation.

[0025] Specifically, in step 4, the disaster risk level output by the classification head is divided into four levels: "low risk", "medium risk", "high risk" and "emergency"; the regression head uses the sigmoid function to predict the maximum displacement rate of the landslide body; the sliding time window mechanism uses 72 hours of historical data as the input window to predict the risk in the next 24 hours; every 12 hours, the model fine-tunes some network layers based on the latest data and the model's own prediction consistency assessment; when no new data is obtained for a certain type of data within the update cycle, its features will use the most recent valid value to participate in the calculation, and this can be reflected in the confidence weight.

[0026] Specifically, in step 5, the dynamic early warning report includes risk areas overlaid on a digital map in the form of a heat map, the risk probability evolution trend over the past 72 hours shown in a line graph, and key indicators presented in the form of a dashboard.

[0027] In addition, step 5 also includes establishing a closed-loop feedback mechanism for geological disaster early warning effects, obtaining the accuracy of early warning results through post-event on-site verification and remote sensing verification, and labeling false alarm and missed alarm cases and sending them back to the training dataset for the next round of model iteration optimization.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] (1) In response to the problem that existing technologies only achieve spatiotemporal alignment of multi-source heterogeneous data and lack a unified expression at the mechanistic level, this invention achieves unified semantic fusion of multi-source data such as remote sensing deformation, meteorological parameters, and ground physical monitoring at the level of geophysical meaning through a semantic standardization method driven by a pre-constructed geological ontology knowledge base. This improvement maps raw data from different observation principles (such as radar interferometry phase, pore water pressure value, and rain gauge reading) into quantitative semantic units with clear geophysical meanings and capable of direct cross-modal computation, such as "surface displacement," "water content change," and "precipitation intensity." This solves the long-standing "semantic gap" problem in this field, provides feature inputs with consistent physical interpretation for subsequent models, and enables the model's decision-making process to be based on interpretable geological evolution mechanisms rather than unreliable statistical correlations.

[0030] (2) To address the inherent limitations of existing technologies where feature fusion weights are static or rely on implicit domain adaptation and cannot respond to real-time quality fluctuations in data sources, this invention proposes a confidence generation network based on historical accuracy, constructing a quantifiable and traceable "dynamic assessment and decision-making system for data source reliability." In this system, firstly, the assessment object undergoes a fundamental shift, changing from adaptation to "data distribution" in existing technology 3 to continuous assessment of the historical objective performance (calculated based on TP / TN / FP / FN) of each independent data source as an "information provider." Secondly, its decision-making mechanism achieves transparency and adaptability. The network and the main model are trained end-to-end, dynamically generating confidence factors based on the performance of each data source over the past 72 hours and explicitly adjusting their weights before fusion. Therefore, this invention possesses intelligent fault tolerance and adaptive capabilities: when the signal quality of a certain data source temporarily degrades, the system can automatically reduce its influence weight in the next fusion cycle based on its recent performance, while simultaneously increasing the decision weights of other reliable data sources, thereby improving the stability and reliability of early warning conclusions in complex and ever-changing monitoring environments.

[0031] (3) To address the problems of early warning lag and performance degradation caused by the "one-time training, static application" of existing early warning models, this invention constructs a hierarchical and collaborative self-evolving early warning model by combining "online fine-tuning with sliding time windows" and "closed-loop feedback iterative optimization." This model not only uses its own prediction consistency assessment as a trigger signal to achieve high-frequency, lightweight short-term environmental adaptation and ensure rapid response to dynamic disturbances such as rainstorms and engineering activities, but also periodically injects high-quality, precisely labeled "lessons learned" samples into the training set to drive full-scale model iteration and achieve systematic correction of long-term cognitive biases. The combination of these two approaches effectively overcomes the risk of "catastrophic forgetting" that may be caused by purely online learning, enabling the system to simultaneously consider short-term agility and long-term evolution, thereby significantly improving the timeliness of early warning and the reliability of continuous service in the time dimension.

[0032] (4) Each step of this invention is interconnected, complementary, and closely related. Semantic standardization provides a feature basis with unified physical meaning for confidence fusion, making performance-based weight adjustments comparable and interpretable. The high-quality decision-making ensured by confidence fusion provides a stable and reliable learning target for sliding window and closed-loop feedback, avoiding ineffective optimization on noisy data. These technical elements reinforce each other and form a closed loop, improving the accuracy, timeliness, robustness, and long-term sustainability of geological disaster early warning. Attached Figure Description

[0033] Figure 1 is a schematic diagram of the overall architecture of an embodiment of the present invention;

[0034] Figure 2 is a schematic diagram of the core principle framework of the multimodal deep fusion neural network model in an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0036] Example

[0037] This embodiment proposes a dynamic early warning method for geological disasters driven by multi-source data and intelligent models. By systematically integrating four types of heterogeneous data from remote sensing, meteorology, UAVs, and ground surveys, a data preprocessing process with spatiotemporal alignment and semantic standardization capabilities is constructed. A multimodal deep fusion neural network model is designed and a cross-modal attention mechanism is introduced to effectively model the nonlinear coupling relationship between surface deformation, precipitation accumulation, and the physical state of soil and rock. Dynamic identification and periodic online updates of disaster risks are achieved through a dual-task output structure and a sliding time window mechanism. Finally, a closed-loop feedback mechanism is constructed to continuously improve the model's generalization ability.

[0038] Referring to Figure 1, the overall architecture of this embodiment includes five functional modules: a data acquisition layer, a data preprocessing layer, an intelligent modeling layer, a risk prediction layer, and a response execution layer. The data acquisition layer is responsible for collecting multi-source heterogeneous geological monitoring data from remote sensing satellite systems, meteorological observation networks, UAV aerial photography systems, and ground exploration equipment. The data preprocessing layer performs spatiotemporal alignment and semantic standardization on the raw data to construct a standardized dataset with a unified format. The intelligent modeling layer deploys a multimodal deep fusion neural network model to extract and fuse features from the standardized data. The risk prediction layer uses fusion feature vector technology to identify the evolution state of disasters and predict risk levels, while utilizing a sliding time window mechanism to ensure that the model can be updated periodically to adapt to the dynamic changes of geological disasters. The response execution layer generates dynamic early warning reports and triggers a tiered response mechanism, while simultaneously feeding the early warning results back to the training dataset to form a closed-loop optimization system.

[0039] In the above-mentioned multi-source data and intelligent model-driven dynamic early warning method for geological disasters, step 1: acquire multi-source heterogeneous geological monitoring data. The regional surface deformation information is collected through the remote sensing satellite system with a spatial resolution of 3m. The time interval is dynamically adjusted according to the satellite revisit cycle. At the same time, hourly precipitation, temperature, humidity and wind speed data are acquired through the meteorological observation network. High-resolution images of key hidden danger areas are collected every 7 days through the UAV aerial photography system with a spatial resolution of 0.3m and a flight altitude of 180m. Ground survey equipment is deployed at key locations of the landslide body to collect pore water pressure, rock and soil displacement and tilt angle data. The sampling frequency is once every 20 minutes. All data are uniformly connected to the system data center.

[0040] Specifically, the remote sensing satellite system employs synthetic aperture radar interferometry (SAR) technology, with differential interferometry processing achieving an accuracy of 5mm under ideal conditions. Data originates from domestically developed Gaofen series satellites, with an orbital period of 12 days. Long-term deformation trends are extracted using permanent scatterer technology, suppressing atmospheric interference and noise. The UAV aerial photography system is equipped with multispectral sensors and lidar. Flight paths are generated by automatic planning algorithms, covering the main crack, rear edge, and leading edge areas of the landslide. Image stitching utilizes feature point matching and global optimization algorithms to generate orthophotos and digital elevation models, achieving an elevation accuracy better than 0.5m. Ground survey equipment includes deep displacement gauges, pore water pressure sensors, and inclinometers, deployed at depths of 5m, 8m, and 12m respectively. Data is automatically transmitted via 4G / 5G wireless networks, and data integrity verification employs a cyclic redundancy check mechanism to ensure controllable transmission errors. During data access, the data center is configured with a distributed message queue system to buffer, sort, and deduplicate data streams from different sources, ensuring data temporal integrity and logical consistency.

[0041] In the aforementioned multi-source data and intelligent model-driven dynamic early warning method for geological disasters, step 2 involves spatiotemporal alignment and semantic standardization of the multi-source data. Using a geographic coordinate system as the reference, bilinear interpolation is employed to resample remote sensing and meteorological data to a unified spatial grid with a grid size of 25m × 25m. The time axis uses Coordinated Universal Time (UTC) as the unified reference. For data with different update frequencies, processing is performed within a unified temporal framework: meteorological (hourly) and surface data (every 20 minutes) are processed using linear interpolation or aggregation to generate hourly data records; for remote sensing (approximately 12-day cycle) and UAV data (7-day cycle), the latest valid data represents their current state under that time reference, remaining unchanged before data updates to ensure all data are aligned within the unified spatiotemporal grid framework. Simultaneously, unified geological semantic labels are assigned to various data types, including "surface displacement," "precipitation intensity," "water content change," and "structural surface activity," constructing a semantically consistent standardized dataset.

[0042] Specifically, semantic standardization processing uses a pre-built geological ontology knowledge base for label mapping and relationship association. The knowledge base contains multiple types of geological phenomena and parameter relationships, supports rule-based automatic labeling and similarity-based fuzzy matching, and drives an automatic mapping model. In this embodiment, the geological ontology knowledge base includes the following parts: (1) Concept system: defines the concepts of "geological entities" (such as slopes, soil and rock masses), "monitoring parameters" (such as deformation rate, pore water pressure, cumulative rainfall), "geological events" (such as landslides, continuous rainfall), and "states" (such as stability, creep); (2) Relationship definition: clarifies the semantic relationships between concepts, such as "monitoring parameters" belonging to a certain "geological entity", "geological events" causing a certain "monitoring parameter" to change, and physical associations such as "positive correlation" or "negative correlation" between different "monitoring parameters"; (3) Logical rules: contains a series of "if-then" rules for semantic reasoning and label mapping, such as: "IFInSAR deformation rate > 30 mm / year AND 72-hour cumulative rainfall > 150 mm THEN state label = 'accelerated creep'". Furthermore, the geological ontology knowledge base construction process is as follows: First, natural language processing technology is used to perform thematic mining and semantic analysis on textual materials such as professional literature, technical standards, and historical disaster case reports in the field of geological hazards. This automatically identifies and extracts core geological concepts (such as landslides, debris flows, displacement rates, cumulative rainfall, pore water pressure, etc.) and their interrelationships, forming a preliminary conceptual system. Subsequently, this system is compared and integrated with a rule base containing prior knowledge from domain experts. Through human-machine collaboration, concepts are deduplicated, layered, attributes are defined, and constraints are clarified, forming a structured ontology model. For example, it can be defined that "when the '72-hour cumulative rainfall' attribute of the monitored area exceeds a certain empirical threshold (which can be determined based on historical regional data) and the 'surface deformation rate' attribute continues to increase, its 'hazard risk status' attribute value should be adjusted to 'accelerated creep'."

[0043] The automatic mapping model is a multilayer perceptron. Its semantic mapping process is as follows: The model receives multi-source raw or intermediate feature data (such as InSAR deformation phase values, meteorological station precipitation interpolation grids, sensor voltage sequences, etc.) from step 1 after spatiotemporal alignment. The model first calls the rule engine in the geological ontology knowledge base to match the input features with the concept definitions and logical rules in the knowledge base, calculating a preliminary rule-based semantic matching degree vector. For example, for a set of inputs, the rule engine calculates its conformity to the rule premise of "accelerated creep in landslides." Then, this semantic matching degree vector is concatenated with the raw input features and used as input to the multilayer perceptron. This multilayer perceptron has been trained using semantic label values ​​such as "surface displacement" and "precipitation intensity," labeled by experts based on historical data, as supervisory signals. Through training, the model learns the nonlinear mapping relationship from "raw features + rule matching degree" to the final precise quantification of semantic labels, thereby achieving efficient and intelligent semantic standardization of multi-source data. The automatic mapping model is trained using an expert annotation set during the pre-training stage, which is manually generated based on historical data. Label consistency verification is performed through cross-validation using an expert annotation set.

[0044] During spatial alignment, remote sensing imagery and UAV orthophotos are first geometrically corrected using ground control points, with the correction residual controlled within 1 pixel. Then, meteorological station data is extended to the entire regional grid using inverse distance weighted interpolation. The location coordinates of ground survey equipment are precisely positioned using RTK-GPS with an error of less than 0.02m, and its point observations are extended to the grid using inverse distance weighted interpolation. During temporal alignment, using Coordinated Universal Time 00:00 as the reference point, the ground survey data is aggregated (e.g., averaged) or interpolated every 20 minutes to obtain hourly values. Meteorological data itself is recorded hourly and does not require interpolation. Both remote sensing and UAV data use the most recently acquired valid data as the representative value for the current moment until new data arrives. After completing the spatiotemporal alignment, each grid cell contains all four types of data fields at each time step, forming a four-dimensional tensor structure with dimensions of [number of time steps, number of grid rows, number of grid columns, number of data channels]. The number of data channels is 4, corresponding to surface displacement, precipitation intensity, water content change, and structural surface activity, respectively.

[0045] In the above-mentioned dynamic early warning method for geological disasters driven by multi-source data and intelligent models, step 3 is to construct a multimodal deep fusion neural network model. The model consists of four parallel feature extraction sub-networks and a cross-modal attention mechanism module. Remote sensing data is input into a convolutional neural network to extract spatial texture and deformation gradient features, meteorological data is input into a long short-term memory network to extract time series trend features, UAV imagery is input into a residual network to extract local geomorphic evolution features, and ground survey data is input into a fully connected network to extract dynamic features of physical parameters. All feature vectors are normalized and then input into the cross-modal attention mechanism module to calculate the weight distribution among the features of each modality and output the fused feature vector.

[0046] Specifically, the cross-modal attention mechanism module adopts a multi-head attention structure with 8 heads. The query, key, and value vectors each have a dimension of 128. The attention weights are normalized using a softmax function. During the fusion process, a modality confidence factor is introduced, and the weights are dynamically adjusted based on the historical accuracy of each data source to avoid interference from low-quality data. The training of the multimodal deep fusion neural network model employs a transfer learning strategy. In the pre-training stage, historical data from 15 typical landslide areas across the country are used for model initialization. In the fine-tuning stage, measured data from the southern Sichuan region over the past 5 years are used. The training samples contain 6285 spatiotemporal data blocks, each covering 1 km. 2 Regional and 72-hour time series. In terms of model architecture details, the remote sensing data branch uses ResNet-34 as the backbone network, with an input size of 256×256 pixels and an output of a 512-dimensional feature vector; the meteorological data branch uses a 2-layer LSTM with 256 hidden units, outputting a 256-dimensional feature vector; the UAV imagery branch uses a lightweight MobileNetV3, with an input size of 512×512 pixels and an output of a 512-dimensional feature vector; and the ground survey data branch uses a 3-layer fully connected network with 128, 256, and 512 neurons per layer, outputting a 512-dimensional feature vector. The feature vectors output from all branches are L2 normalized and concatenated into a 1792-dimensional vector, which is then input to the cross-modal attention module. This module first projects the concatenated vector into eight independent 128-dimensional subspaces, calculating the query, key, and value matrices in each subspace. Then, it calculates the weights through a scaling dot product attention mechanism. Finally, it concatenates the outputs of the eight heads and performs a linear transformation to obtain a preliminary 512-dimensional fused feature vector.

[0047] Modal confidence factors are generated by a two-layer fully connected gating network: the input layer receives the prediction accuracy of each modality over the past 72 hours (accuracy calculation formula: (where TP represents true positives, TN represents true negatives, FP represents false positives, and FN represents false negatives). In the initial operation of the system or when historical data is insufficient or historical accuracy is unavailable, the input layer receives a preset initial confidence value, which is set based on the prior reliability assessment of the data source. The hidden layer (16 neurons, ReLU activation) calculates intermediate features, and the output layer (Sigmoid activation) generates confidence weights. This confidence generation network is trained and optimized end-to-end along with the main model during training. Furthermore, during the initial training and operation phases, the preset initial confidence bias can participate in learning to quickly adapt to the actual environment. Finally, the preliminary fused feature vector is split into four parts according to its modality source, multiplied by the corresponding modality confidence factor, and then weighted and fused. This is then passed through a linear layer to obtain the final 512-dimensional fused feature vector.

[0048] Figure 2 illustrates how, in this embodiment, four parallel inputs extract features through their respective dedicated sub-networks and achieve deep fusion in the cross-modal attention module. Specifically, remote sensing data flows through a convolutional neural network, meteorological data through a long short-term memory network, UAV imagery through a residual network, and ground survey data through a fully connected network. The four feature vectors are normalized and then input into the cross-modal attention mechanism module, ultimately outputting a fused feature vector. This architecture ensures that the characteristics of different data types are fully preserved. Simultaneously, the attention mechanism enables adaptive weighted fusion of information, effectively addressing the semantic gap problem in multi-source heterogeneous data fusion.

[0049] In the above-mentioned multi-source data and intelligent model-driven dynamic early warning method for geological disasters, step 4: Based on the fused feature vector, the disaster evolution state is identified and the risk level is predicted. The fused feature vector is input into the dual-task output layer of the classification head and the regression head in parallel. The classification head uses the softmax function to output the probability of whether a disaster will occur in the next 24 hours, which is divided into four levels: "low risk", "medium risk", "high risk" and "emergency". The regression head uses the sigmoid function to predict the maximum displacement rate of the landslide body, in millimeters per day. The model is updated online and adaptively adjusted through a sliding time window mechanism. Every 12 hours, the model is fine-tuned based on the latest monitoring data and the model's own prediction consistency assessment.

[0050] Specifically, the classification and regression heads share some fused features. The loss function uses a weighted combination: cross-entropy loss for classification and smoothed L1 loss for regression, with a total loss weight ratio of 3:1. The optimizer is AdamW, with an initial learning rate of 0.001, a batch size of 32, and 200 training iterations. Model validation uses 5-fold cross-validation to avoid overfitting. In this embodiment, the sliding window mechanism uses 72 hours of historical data as the input window and the prediction window for the next 24 hours. Every 12 hours, the model triggers a parameter fine-tuning process based on the latest incoming monitoring data and the model's own consistency evaluation of predictions over the continuous time window. Fine-tuning can cover the dual-task output layer as needed, and can selectively perform lightweight updates on some layers of the feature extraction subnetwork and the cross-modal attention module to balance adaptability and stability. In this embodiment, the fine-tuning priority rule is set as follows: when the historical accuracy of a data source is lower than 0.7, the last layer of its corresponding feature extraction sub-network (such as layer 4 of ResNet-34) and the query projection layer of the cross-modal attention module are updated; at the same time, for low-frequency data sources (such as remote sensing and UAVs) that have not obtained new data during the update cycle, the parameters of their feature extraction sub-network will be frozen, and their confidence weights in the fusion process can be adaptively downgraded by the confidence generation network according to their data freshness (such as the time since the last data). The update adopts gradient pruning (threshold 1.0) and fine-tuning the learning rate (0.0001), while freezing the parameters of the remaining network layers.

[0051] Furthermore, in this embodiment, the risk level classification standard is as follows: "low risk" is defined as a probability of disaster occurrence less than 0.3, "medium risk" as a probability of 0.3 to 0.6, "high risk" as a probability of 0.6 to 0.85, and "emergency" as a probability greater than 0.85. The displacement rate prediction is restored to the actual millimeter-per-day value through an inverse sigmoid transform. The risk level output by the classification head serves as the final early warning basis, while the regression head prediction serves as an auxiliary decision-making reference. During the online model update process, newly collected data first undergoes the same preprocessing procedure as the training data, and then forms a new training batch with historical data from the most recent 72 hours. The set updatable network layers are fine-tuned, with the learning rate decaying to 0.0001, and four rounds of fine-tuning are performed to ensure that the model quickly adapts to the latest geological condition changes without forgetting historical knowledge.

[0052] In the above-mentioned dynamic early warning method for geological disasters driven by multi-source data and intelligent models, step 5: generate a dynamic early warning report and trigger a response mechanism. The emergency plan is automatically matched according to the risk level. Early warnings of high risk and above are pushed to relevant management departments and residents in the affected areas simultaneously through SMS, broadcast and emergency platform. At the same time, a visual report containing risk area map, evolution trend curve and key indicator thresholds is generated, as shown in the meteorological risk early warning of geological disasters in Sichuan Province, which is used for dynamic early warning research of geological disasters in complex areas such as southern Sichuan.

[0053] Specifically, the dynamic early warning report includes multi-dimensional information display. Risk areas are displayed as heat maps overlaid on digital maps, and the evolution trend is shown as a line graph showing the change in risk probability over the past 72 hours. Key indicators are presented in the form of dashboards, including current precipitation intensity, displacement rate, and pore water pressure. The report is generated using a template engine, supports PDF and HTML format output, and the push delay is controlled within 2 minutes.

[0054] Furthermore, this embodiment establishes a closed-loop feedback mechanism for geological disaster early warning effectiveness. The accuracy of early warning results is obtained through post-event on-site verification and remote sensing validation. False alarms and missed alarms are labeled and fed back to the training dataset for the next round of model iteration and optimization, forming a self-evolving system of "monitoring-early warning-verification-learning" to continuously improve the model's generalization ability. This feedback closed-loop mechanism operates independently of the aforementioned online sliding window update, running at a fixed cycle (e.g., monthly or quarterly) to comprehensively retrain the model or update key modules.

[0055] During report generation, this embodiment first extracts the current warning results, historical risk sequences, and raw monitoring data from the database. Then, it calls a visualization template engine to populate the data into predefined chart components. The heatmap uses a red-yellow-green gradient to represent the level of risk, the line chart's horizontal axis represents time, and the vertical axis represents the probability of risk. The dashboard pointer position is dynamically adjusted according to the current indicator value. After the report is generated, it is distributed to multiple terminals through a message middleware. For example, the research team receives the full HTML report for analysis, grassroots managers receive a simplified PDF report, and affected residents receive a text summary and evacuation guidance via SMS. In the feedback closed-loop mechanism, for example, within 72 hours after each warning event, professionals conduct on-site verification to confirm whether the disaster actually occurred and mark the verification results as "correct warning," "false alarm," or "missed warning." These labeled data, along with the corresponding raw monitoring data, are stored in a dedicated feedback database. The model is periodically retrained, incorporating the feedback data into the training set to continuously optimize model performance.

[0056] The results of retrospective simulation tests based on historical data and trial runs in a field environment show that this embodiment can issue effective early warnings for multiple landslide evolution processes with precursory information in the region. In typical test cases, this embodiment demonstrates a good advance warning rate for different levels of warnings, with a risk level judgment accuracy rate of 87%, a false alarm rate of 9.5%, and a false negative rate of 3.5%, and its overall performance is superior to existing technologies.

[0057] In summary, this invention constructs a complete closed loop from data perception to model self-optimization, effectively improving the accuracy, timeliness, and reliability of geological disaster early warning.

[0058] The above embodiments are merely preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but which still solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A dynamic early warning method for geological disasters driven by multi-source data and intelligent models, characterized in that, Includes the following steps: Step 1: Acquire multi-source heterogeneous geological monitoring data, including regional surface deformation information, hourly meteorological data, high-resolution images of key hidden danger areas, and physical parameter data; Step 2: Perform spatiotemporal alignment and semantic standardization processing on the multi-source heterogeneous geological monitoring data. Using a geographic coordinate system as a reference, resample data from different sources to a unified spatial grid, organize the data under a unified temporal reference, and assign unified geological semantic labels to the multi-source heterogeneous geological monitoring data to construct a semantically consistent standardized dataset; Step 3: Construct a multimodal deep fusion neural network model consisting of four parallel feature extraction sub-networks and a cross-modal attention mechanism module to process the standardized multi-source heterogeneous geological monitoring data. The data is extracted separately, and all feature vectors are adaptively weighted and fused to output a fused feature vector. Among them, the cross-modal attention mechanism module dynamically adjusts the weights based on the historical accuracy of each data source through the confidence generation network to achieve adaptive fusion of multimodal data. Step 4: Based on the fused feature vector, the disaster evolution state is identified and the risk level is predicted. The fused feature vector is input into the dual-task output layer with parallel classification and regression heads. The classification head outputs the disaster risk level within a specified time period in the future, and the regression head predicts the landslide displacement rate. The model is updated and the parameters are adjusted online through a sliding time window mechanism. Step 5: A dynamic early warning report is generated according to the disaster risk level and the response mechanism is triggered.

2. The method for dynamic early warning of geological disasters driven by multi-source data and intelligent models according to claim 1, characterized in that, In step 1, regional surface deformation information is collected through a remote sensing satellite system, hourly meteorological data is obtained through a meteorological observation network, high-resolution images of key hidden danger areas are collected through a drone aerial photography system, and physical parameter data are collected through ground survey equipment deployed at key locations of the landslide body.

3. A dynamic early warning method for geological disasters driven by multi-source data and intelligent models according to claim 1 or 2, characterized in that, In step 2, the size of the unified spatial grid is 25m×25m; the time reference is Coordinated Universal Time, and it is organized and filled under a unified temporal framework according to the actual update frequency of each data source; the geological semantic tags include "surface displacement", "precipitation intensity", "water content change" and "structural surface activity". The semantic standardization process is based on a pre-built geological ontology knowledge base containing multiple geological phenomena and parameter relationships for tag mapping and association.

4. The method for dynamic early warning of geological disasters driven by multi-source data and intelligent models according to claim 3, characterized in that, The automatic mapping model enables the label mapping and association of a geological ontology knowledge base that relates various geological phenomena to parameters. Specifically, the automatic mapping model is trained based on the knowledge base rule engine and combines the correspondence between features and labels in historical data. After training, the input features are semantically analyzed according to the knowledge base rules and automatically mapped to unified geological semantic labels, outputting the corresponding quantitative values.

5. The method for dynamic early warning of geological disasters driven by multi-source data and intelligent models according to claim 4, characterized in that, In step 3, the four parallel feature extraction sub-networks are: a convolutional neural network for processing remote sensing data, a long short-term memory network for processing meteorological data, a residual network for processing UAV imagery, and a fully connected network for processing ground survey data.

6. The method for dynamic early warning of geological disasters driven by multi-source data and intelligent models according to claim 5, characterized in that, In step 3, the cross-modal attention mechanism module adopts a multi-head attention structure with 8 heads and 128 dimensions for query, key, and value vectors.

7. The method for dynamic early warning of geological disasters driven by multi-source data and intelligent models according to claim 6, characterized in that, In step 3, the process by which the cross-modal attention mechanism module dynamically adjusts weights based on the historical accuracy of each data source through the confidence generation network is as follows: Step a: Calculate the modal confidence factor of each data source through the confidence generation network; the input of the confidence generation network is the historical accuracy of each data source in the past 72 hours, which is calculated based on the statistical results of true positives, true negatives, false positives, and false negatives; the confidence generation network maps the input historical accuracy to a modal confidence factor between 0 and 1 through a fully connected network structure containing hidden layers and output layers; this confidence generation network is trained and optimized together with the main network during the model training phase; Step b: Split the preliminary fused feature vector output by the cross-modal attention mechanism module into four parts according to the modal source, each part corresponding to a feature vector of a data source; Step c: Multiply each split modal feature vector by its corresponding modal confidence factor to complete the dynamic weight adjustment based on historical accuracy; Step d: Fuse the weighted modal feature vectors and obtain the final fused feature vector through linear layer transformation.

8. The method for dynamic early warning of geological disasters driven by multi-source data and intelligent models according to claim 7, characterized in that, In step 4, the disaster risk level output by the classification head is divided into four levels: "low risk", "medium risk", "high risk" and "emergency". The regression head uses the sigmoid function to predict the maximum displacement rate of the landslide body. The sliding time window mechanism uses 72 hours of historical data as the input window to predict the risk in the next 24 hours. Every 12 hours, the model fine-tunes some network layers based on the latest data and the model's own prediction consistency assessment. When no new data is obtained for a certain type of data within the update cycle, its features will use the most recent valid value to participate in the calculation, which can be reflected in the confidence weight.

9. A dynamic early warning method for geological disasters driven by multi-source data and intelligent models according to claim 8, characterized in that, In step 5, the dynamic early warning report includes risk areas overlaid on a digital map in the form of a heat map, the risk probability evolution trend over the past 72 hours shown in a line graph, and key indicators presented in the form of a dashboard.

10. A dynamic early warning method for geological disasters driven by multi-source data and intelligent models according to claim 1, 2, 4, 5, 6, 7, 8, or 9, characterized in that, Step 5 also includes establishing a closed-loop feedback mechanism for geological disaster early warning effects. The accuracy of the early warning results is obtained through post-event on-site verification and remote sensing verification. False alarms and missed alarms are labeled and sent back to the training dataset for the next round of model iteration optimization.

Images