Landslide monitoring platform based on multi-source data

Abstract

The invention relates to the technical field of geological monitoring, in particular to a landslide monitoring platform based on multi-source data, which comprises the following steps of: acquiring space background data and real-time meteorological data of a monitoring area; obtaining the effective critical rainfall of the corresponding time window; constructing a standardized rainfall trigger index corresponding to the time window; carrying out reliability constraint on the rainfall triggering index to form a final rainfall induction intensity characterization value; and generating a comprehensive landslide risk index of each grid unit, and giving different risk grades according to a preset grading rule. According to the method, a dual mechanism of cross-time window consistency judgment and multi-moment continuous judgment is provided, and the reliability attenuation coefficient is introduced based on the judgment result to dynamically regulate and control the induction intensity, so that the problem of false high value output caused by short-time heavy rainfall, data abnormality or critical point fluctuation is effectively solved, and the stability and continuity of a landslide triggering index are improved.

Description

Technical Field

[0001] This invention relates to the field of geological monitoring technology, and in particular to a landslide monitoring platform based on multi-source data. Background Technology

[0002] Landslides are influenced by factors such as heavy rain, earthquakes, or engineering disturbances, and can evolve rapidly in a short period of time, causing damage. Therefore, improving landslide monitoring and early warning capabilities has become a key aspect of geological disaster prevention and control.

[0003] Currently, landslide early warning methods are mainly divided into two categories: one is based on historical statistical empirical critical rainfall threshold models, and the other is based on physical models or machine learning prediction methods using real-time monitoring data. The former usually uses fixed or regional average values ​​as trigger criteria, ignoring the differences in geological and topographical conditions at different locations, which can easily lead to "over-warning" or "under-warning" phenomena. Although the latter has a certain degree of intelligence, most models are based on rainfall data from a single time window, lacking the identification of temporal characteristics of rainfall evolution, and are easily affected by the randomness of short-term heavy rainfall. In addition, existing methods generally lack credibility constraints and spatial expression mechanisms for model output results, making it difficult to provide structured and highly interpretable early warning information for grassroots decision-making departments.

[0004] In terms of monitoring platform construction, traditional landslide monitoring systems are mostly limited to displaying a single data source or a single monitoring indicator, lacking the ability to integrate and express spatial background information, dynamic changes in rainfall, and the evolution of risk processes. In particular, in terms of multi-source data fusion and spatial-temporal joint risk assessment, there is still a lack of universal and practical system methods, which restricts the improvement of landslide risk intelligent perception, graded response, and visualized early warning capabilities. Summary of the Invention

[0005] This invention provides a landslide monitoring platform based on multi-source data, which integrates multi-source spatial data, real-time meteorological data and machine learning technology. It has the ability to dynamically identify risks, perform cross-temporal and spatial evolution analysis and provide interpretive output, thereby improving the stability and practicality of landslide early warning systems.

[0006] A landslide monitoring platform based on multi-source data, the platform operation includes performing the following monitoring: S1, acquire spatial background data and real-time meteorological data of the monitoring area. The spatial background data includes historical landslide samples, topographic and geological features and remote sensing indices. The real-time meteorological data includes rainfall sequences for each time window in the recent period. S2, based on the spatial background data, a background susceptibility index for each grid cell is generated through a pre-trained machine learning model, and the rainfall threshold for each time window is corrected according to the background susceptibility index to obtain the effective critical rainfall for the corresponding time window, forming an effective critical rainfall set corresponding to each grid cell. The machine learning model adopts an improved LightGBM model based on gradient boosting tree, and its training sample set includes spatial environmental variables corresponding to known landslide events. S3. Based on the real-time rainfall data, calculate the sliding cumulative rainfall for each time window, and combine it with the effective critical rainfall set to construct a standardized rainfall trigger index for the corresponding time window. Based on the generated rainfall trigger index, determine the consistency and persistence of the rainfall trigger state for each time window, identify whether the rainfall trigger is in a stable over-threshold state, and thus impose reliability constraints on the rainfall trigger index to form the final rainfall-induced intensity characterization value. S4, the baseline susceptibility index and the rainfall-induced intensity characterization value are multiplied grid by grid to generate a comprehensive landslide risk index for each grid cell, and different risk levels are assigned according to a preset classification rule; S5. The risk level results are displayed spatially and can be dynamically updated in chronological order. When the user selects any monitoring unit, risk interpretation information based on the background susceptibility, rainfall-induced state and its stability judgment results is output.

[0007] Optionally, the acquisition of the spatial baseline data includes: importing the historical landslide logging database of the geological survey department to obtain the location and boundary information of historical landslide samples; extracting topographic feature factors from digital elevation model data, extracting geological feature factors from regional geological map data, and performing index calculations on multi-temporal remote sensing images to generate a remote sensing index layer; and performing spatial registration, resampling, and normalization processing on the historical landslide samples, topographic and geological feature factors, and remote sensing index layer under a unified spatiotemporal reference to form a spatial baseline dataset that corresponds one-to-one with the raster units of the monitoring area.

[0008] Optionally, the acquisition of real-time meteorological data includes: accessing real-time rainfall monitoring data provided by meteorological departments or sensor networks deployed in the monitoring area through a data interface, and dynamically calculating and updating the recent rainfall sequence corresponding to each grid unit based on multiple preset time windows.

[0009] Optionally, the construction of the training sample set includes using the grid cells corresponding to the locations of the historical landslide samples as positive samples, randomly selecting grid cells in areas where no landslides have occurred as negative samples, and extracting the topographic and geological features and remote sensing indices corresponding to all samples as model feature variables; using the training sample set to train the improved LightGBM model, wherein the improvement includes introducing a weighting term for spatial sample imbalance into the loss function, and using feature importance ranking for recursive feature elimination to optimize the input feature subset, and obtaining the background susceptibility prediction model after training.

[0010] Optionally, the generation of the background susceptibility index includes inputting the spatial background dataset corresponding to all grid cells in the monitoring area into the background susceptibility prediction model, and the model outputs the probability value of landslides occurring in each grid cell, which is the background susceptibility index of the grid cell.

[0011] Optionally, the level correction of the rainfall threshold includes: defining a regional baseline critical rainfall for each preset time window; establishing a level correction function based on the background susceptibility index, mapping the background susceptibility index of the grid cell to a rainfall reduction coefficient between 0 and 1, wherein the higher the background susceptibility index, the smaller the reduction coefficient; multiplying the regional baseline critical rainfall by the rainfall reduction coefficient of the grid cell to obtain the effective critical rainfall of the grid cell in that time window; Iterate through all grid cells and all preset time windows, repeatedly adjust the rainfall threshold level, and generate the effective critical rainfall for each grid cell under different time windows, thereby forming a set of effective critical rainfall corresponding one-to-one with the grid cells and time windows.

[0012] Optionally, S3 specifically includes: S31, for each grid cell, according to its real-time rainfall sequence, a sliding summation calculation is performed according to the length of each preset continuous time window to obtain the sliding cumulative rainfall of the grid cell at the current moment corresponding to each time window; S32, for each grid cell and each time window, divide the sliding cumulative rainfall corresponding to the time window by the effective critical rainfall to obtain a ratio, which is the standardized rainfall triggering index of the current grid cell under the time window; when the ratio is less than or equal to 1, it indicates that the rainfall triggering condition has not been met; when the ratio is greater than 1, it indicates that the rainfall triggering condition has been met, and the larger the ratio, the higher the rainfall induction potential.

[0013] Optionally, S3 further includes: Consistency determination of triggering state: For the current moment, analyze the multiple standardized rainfall triggering indices calculated for the same grid cell under different time window lengths; if the index values ​​of the time windows exceeding a preset proportion are all greater than 1, then it is determined that the rainfall triggering state of the grid cell at the current moment has consistency across time scales. Triggering state persistence determination: Review the rainfall triggering state of the grid cell at multiple consecutive moments in the recent period; if the consistency determination result is consistently positive in the recent period, it is determined that the rainfall triggering of the grid cell is in a stable over-threshold state; Based on the combined results, for each grid cell, if its current standardized rainfall trigger index is greater than 1 and simultaneously satisfies both consistency and persistence criteria, it is marked as a stable overthreshold, and its standardized rainfall trigger index is directly used as the final rainfall-induced intensity characterization value; if its index is greater than 1 but does not satisfy consistency or persistence criteria, the index is constrained and reduced by multiplying it by a reliability attenuation coefficient to obtain the final rainfall-induced intensity characterization value; if its index is less than or equal to 1, the final rainfall-induced intensity characterization value is 0 or a baseline value representing no risk.

[0014] Optionally, S4 specifically includes: S41, For each grid cell within the monitoring area, the background susceptibility index of the grid cell is multiplied by the rainfall-induced intensity characterization value to obtain the initial comprehensive risk value of the grid cell. S42, normalize the initial comprehensive risk value of all grid cells in the region and map it to a continuous interval from 0 to 1 to obtain the comprehensive landslide risk index of each grid cell; S43, set multiple risk level thresholds to divide the value range of the comprehensive landslide risk index into corresponding risk level intervals; the risk levels include at least low risk, medium risk, high risk and extremely high risk; S44. Assign a corresponding risk level to each grid cell based on the risk level range of the comprehensive landslide risk index.

[0015] Optionally, S5 specifically includes: S51, based on the geographic information system platform, renders each grid cell according to the assigned risk level and different preset color symbols to generate a thematic map of the spatial distribution of landslide risk level in the monitoring area. S52, in the geographic information system platform, a time axis control is used to associate thematic maps of the spatial distribution of landslide risk levels corresponding to different calculation times; when the time axis control is operated, the platform interface dynamically switches and displays the risk spatial distribution at the corresponding time, realizing the playback and tracking of the spatiotemporal evolution of risk. S53, when a user selects any grid cell or aggregated administrative / watershed management unit on the platform interface through interactive operation, the information query and interpretation module is triggered to automatically extract and organize multi-source analysis data and intermediate results related to the corresponding unit, and generate a structured risk interpretation report.

[0016] The beneficial effects of this invention are: This invention couples a static spatial background susceptibility index with a dynamic standardized rainfall triggering index in multiple dimensions. It uses a machine learning model to generate susceptibility assessment results for each grid cell and combines multi-time window and multi-period sliding rainfall calculation with a threshold discrimination mechanism to construct a complete background-rainfall linkage triggering model. This significantly improves the accuracy of landslide identification in areas with strong spatial heterogeneity and complex local induction mechanisms, and avoids the misjudgment or missed judgment problem caused by the one-size-fits-all approach to critical rainfall in traditional methods.

[0017] Unlike most existing methods that rely solely on rainfall data within a single time window, this invention proposes a dual mechanism based on a standardized rainfall triggering index. This mechanism combines cross-time window consistency discrimination with multi-moment persistence discrimination. Furthermore, it introduces a reliability decay coefficient based on the discrimination results to dynamically adjust the induction intensity. This effectively overcomes the problem of false high-value outputs caused by short-duration heavy rainfall, data anomalies, or critical point fluctuations, thereby improving the stability, continuity, and explanatory power of landslide triggering indicators.

[0018] This invention integrates functions such as dynamic rendering of landslide risk level thematic maps, interactive playback of timelines, raster-level risk query, and automatic generation of interpretation reports based on a GIS platform. It proposes a structured multi-source interpretation model from raw data to intermediate indicators to the final risk level. Users can obtain a comprehensive report containing background indices, rainfall index sequences, consistency and persistence assessments, the final risk level, and recommended response measures by clicking on any raster. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the monitoring platform modules according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the operation steps of the monitoring platform according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0022] like Figures 1-2 As shown, the landslide monitoring platform based on multi-source data includes the following modules. Multi-source data access module: Unifies the access and preprocessing of spatial and meteorological data from different sources, providing high-quality input data support, mainly including: Spatial baseline data access submodule: Import historical landslide logging databases, regional geological maps, digital elevation models (DEMs), and multi-temporal remote sensing images from geological survey departments; Real-time meteorological data access submodule: Connects to the meteorological department interface or the sensor network deployed in the monitoring area to dynamically collect rainfall data; Data standardization and alignment submodule: Spatial registration, resampling and normalization of all raster cells are performed to form an input dataset with a unified format.

[0023] The susceptibility modeling module is used to construct a landslide baseline susceptibility model based on historical landslide samples and spatial factors, and to identify potentially high-risk areas. It includes: Training sample construction submodule: calibrate landslide and non-landslide raster samples, and extract topographic, geological and remote sensing features; Improved LightGBM modeling submodule: Introduced spatial sample weighting mechanism and feature recursive elimination strategy to enhance model robustness; Landslide susceptibility index calculation submodule: performs rasterized prediction of the entire area and outputs the landslide susceptibility index for each cell.

[0024] Rainfall Trigger Identification Module: Used to identify and assess the actual impact of current rainfall on landslide triggering, including: Sliding rainfall calculation submodule: Calculates sliding cumulative rainfall based on a preset time window; Standardized trigger index calculation submodule: Calculates the ratio of actual rainfall to the corresponding effective critical rainfall; Consistency determination submodule: Determines whether the triggering states are consistent across multiple time windows; Persistence determination submodule: Determines whether the state persists continuously over a continuous time period; Induced Intensity Characterization Submodule: Integrates the above factors and outputs a reliable rainfall induced intensity value for each grid cell.

[0025] Comprehensive Risk Index Calculation Module: This module integrates baseline risk with rainfall-induced risk to form a standardized comprehensive landslide risk index, specifically including: Product-coupled calculation submodule: Multiplies the susceptibility index with the induced intensity value grid by grid to obtain the initial risk value; Normalization mapping submodule: standardizes the initial risk value to the [0,1] interval; Risk level classification submodule: Based on set thresholds, it classifies risks into low, medium, high, and extremely high levels; Risk level assignment submodule: Assigns a corresponding risk label to each grid cell.

[0026] GIS Visualization and Interaction Module: This module, based on the GIS platform, implements spatial rendering, time-series control, and risk query and interpretation functions. It mainly includes: Thematic map rendering submodule: Draws a spatial distribution map of landslide risk based on risk level; Timing control submodule: Associated with timeline controls, supporting risk evolution playback and comparison; Interactive query submodule: Users can trigger information retrieval by clicking on any grid or management unit; Risk Explanation Report Generation Submodule: Automatically outputs structured explanatory information including susceptibility, rainfall process, trigger status, and response recommendations.

[0027] The platform operates based on the above modules and sub-modules, and performs the following monitoring: S1, acquire spatial background data and real-time meteorological data of the monitoring area. The spatial background data includes historical landslide samples, topographic and geological features and remote sensing indices. The real-time meteorological data includes rainfall sequences for each time window in the recent period.

[0028] S11, Acquisition of Spatial Baseline Data: By importing the historical landslide logging database provided by the geological survey department, the location and boundary information of historical landslide samples are obtained; topographic feature factors, including slope, aspect, curvature, and topographic humidity index, are extracted from the digital elevation model (DEM); geological feature factors, including lithology, fault distance, and stratigraphic age, are extracted from the regional geological map; remote sensing indices are calculated on remote sensing images of different time phases to obtain multiple index layers such as Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), and Normalized Difference Building Index (NDBI).

[0029] The aforementioned multi-source factors are registered, resampled, and normalized under a unified spatial reference frame to construct and monitor regional grid cells. —Corresponding spatial background dataset , represented as: ;in, Indicates the first Spatial background dataset of raster cells Represents a set of geological characteristic factors. Represents the set of terrain feature factors. This represents the set of remote sensing index factors.

[0030] S12, Real-time meteorological data acquisition: Obtain rainfall data streams by connecting to the meteorological bureau system through a data interface or by directly integrating with the meteorological sensor network deployed in the monitoring area.

[0031] Based on multiple preset time windows The rainfall within each time window is accumulated and statistically analyzed to obtain the rainfall sequence of each grid cell in the recent period. , represented as: ;in, Indicates the first Grid cells in time window The cumulative rainfall within the area, Indicates time The rainfall in this grid cell, Indicates the first There are several time windows, including rolling time periods such as 1 hour, 3 hours, 6 hours, and 24 hours.

[0032] S2, based on the spatial background data, a background susceptibility index for each grid cell is generated through a pre-trained machine learning model, and the rainfall threshold for each time window is corrected according to the background susceptibility index to obtain the effective critical rainfall for the corresponding time window, forming an effective critical rainfall set corresponding to each grid cell. The machine learning model adopts an improved LightGBM model based on gradient boosting trees, and its training sample set includes spatial environmental variables corresponding to known landslide events.

[0033] S21, Constructing a training sample set: In landslide susceptibility prediction, a machine learning model is needed to determine whether a location is prone to landslides. In order to train this model, a set of sample data with known answers must be prepared, which is the training sample set.

[0034] The first step is to construct positive and negative samples: Historical landslide locations are retrieved from existing landslide logging databases. These locations are represented on the platform as raster cells, i.e., map areas divided into regular squares. These raster cells representing landslide occurrences are defined as positive samples, as they represent landslide events. Simultaneously, some samples representing landslide-free areas are needed as negative samples for model learning comparison. These negative samples are raster cells randomly selected from areas where landslides did not occur. To ensure the model learning process is not severely affected by a significant imbalance in the number of samples, the number of positive and negative samples is usually kept roughly equal.

[0035] The second step is to extract input features: for each positive and negative sample, extract the spatial background data features of its corresponding location. Specifically, this means: Topographic factors extracted from digital elevation models, such as slope, aspect, and curvature; Geological factors extracted from geological maps, such as lithology and fault distance; Remote sensing indexes calculated from remote sensing images; These factors for each sample form a feature vector, representing the environmental conditions at that location.

[0036] The third step is to form a training sample set: each sample consists of two parts: one is its input features, which are the terrain and geological remote sensing factors mentioned earlier; the other is its label, which is used to indicate whether the sample is a landslide point, 1 indicates that a landslide has occurred, and 0 indicates that it has not occurred. This constitutes the basic unit of training data, that is, an input-output pair. Integrating all the samples together constitutes the entire training sample set, which is used for subsequent model training.

[0037] Specifically, this includes setting the raster cells corresponding to the landslide occurrence locations in the historical landslide logging database as positive samples. In areas where no landslides have occurred, a random number of grid cells of the same order of magnitude are selected as negative samples. .

[0038] Spatial background data features are extracted from all sample raster cells to form a training sample set: ;in, Indicates the first The input features for each sample include topographic features, geological features, and remote sensing indices. Indicates the first The labels for each sample are: 1 for landslides and 0 for non-landslides. For the first The set of spatial background data features corresponding to each sample.

[0039] S22, Model Training: Train the improved LightGBM model using the aforementioned training sample set. Improvements include: S221. Introduce the spatial sample weight loss function: ;in, The standard binary classification loss function, For the first The weighting coefficients for each sample are adjusted so that positive and negative samples have different weights to alleviate sample imbalance.

[0040] S222. Recursive Feature Elimination (RFE): Based on feature importance ranking, features with lower contributions are eliminated step by step, retaining the optimal subset. This is to improve the model's generalization ability.

[0041] After training, a baseline susceptibility prediction model is obtained. .

[0042] LightGBM is a machine learning algorithm based on gradient boosting decision trees. Its core idea is to gradually improve overall prediction performance through a series of weak classifiers, i.e., simple decision trees. Each newly generated tree attempts to fit the residual of the previous model, i.e., the prediction error, thus gradually approaching the true value. Its final output is the weighted set of all weak classifier results. LightGBM's main features compared to traditional gradient boosting decision trees include: using histogram optimization, which reduces memory consumption and accelerates training by discretizing continuous features into a fixed number of bins; a leaf-first growth strategy, which reduces loss faster than hierarchical growth; and efficient feature parallelism and data parallelism mechanisms, making it suitable for large-scale datasets.

[0043] LightGBM in this invention: Input data structure: The input for each sample is a vector containing spatial features such as terrain, geology, and remote sensing data; Features may include: slope, aspect, NDVI, fault distance, stratigraphic class code, etc. The output label is a binary variable (0 = no landslide, 1 = landslide occurred).

[0044] Model output structure: For each input grid cell, the model outputs a probability value between 0 and 1, representing the probability of a landslide occurring. This probability value is the background susceptibility index.

[0045] The overall process of model training is as follows: Step 1, Input training data: Use the training sample set prepared in the previous stage. Each sample consists of a set of feature vectors and a target label.

[0046] Step 2, introduce a spatial sample weighting mechanism: Landslides are low-probability events, and the number of actual landslide samples is far less than the number of non-landslide samples. Imbalanced samples may cause the model to be more biased towards predicting non-landslides, thus losing important early warning capabilities. During training, this invention assigns a sample weight wk to each sample and gives higher weight to landslide samples. LightGBM supports the sample weighting mechanism, integrating these weights into the loss function so that the model pays more attention to landslide samples during training.

[0047] Step 3, Building a weak classifier: The model is built starting from the first tree. Each tree is built to correct the error of the previous round. Each tree uses a small portion of features to split the nodes. The optimal split point is calculated by information gain or Gini coefficient. After the model is built, the outputs of all trees are combined into the final prediction.

[0048] Step 4, Recursive Feature Elimination Optimization: The original feature set may contain redundant or noisy features, which leads to increased model complexity and decreased generalization ability. During the training process of this invention, the importance score of each feature in the tree structure is recorded, the features are sorted according to the score, the features with the least contribution are gradually eliminated, the optimal feature subset is retained, and the new feature subset is repeatedly trained until the performance is optimal, and finally a model with a simpler structure and stronger predictive power is formed.

[0049] S23, Baseline Susceptibility Index: In S22, a machine learning model for predicting landslide susceptibility was trained based on historical landslide samples. The model's essential function is to determine the likelihood of a landslide occurring at a given location, considering its geological, topographical, and remote sensing environmental information. The trained model is then applied to every grid cell across the entire monitoring area, thereby creating a comprehensive landslide susceptibility map.

[0050] Input data preparation: Each grid cell has its spatial background feature data extracted in S1, including topographic factors, geological factors, and remote sensing indices; these features of each grid cell are organized into an input vector, similar to the sample features used during model training.

[0051] The spatial background features of each of the above grid cells are input into the trained model one by one. The model will output a value, which means the probability of a landslide occurring in the grid cell under the current environmental conditions. This probability value is defined as the background susceptibility index of the cell, and the value ranges from 0 to 1. The closer it is to 1, the higher the risk of landslide at that location, and the easier it is for the environment to trigger a landslide. The closer it is to 0, the more stable the geological and topographical conditions at that location are, and the less likely it is to landslide.

[0052] The final output is a two-dimensional raster map with the same spatial resolution as the monitored area, and each raster cell is assigned a value, namely the susceptibility index.

[0053] Formal expression means: all grid cells in the monitoring area Spatial background data characteristics Input into the trained model, and output the baseline susceptibility index of each grid cell. : ;in, Represents grid cells The probability of a landslide occurring, also known as the baseline susceptibility index.

[0054] S24, Rainfall Threshold Level Correction: In landslide monitoring, a critical rainfall threshold is needed to determine whether a rainfall event may trigger a landslide. Traditional methods often use a uniform regional critical rainfall threshold, such as setting a 24-hour cumulative rainfall of more than 100 mm as a possible trigger for a landslide. However, this approach ignores the significant differences in background conditions between different regions.

[0055] For example: Location A has loose geology, a steep slope, and thick topsoil; even 50mm of rain could cause a landslide. Location B has hard rock, a gentle slope, and good drainage; even if the rock is pushed down 120mm, a landslide is not necessarily possible.

[0056] Therefore, it is necessary to spatially modify the unified regional critical rainfall amount according to local conditions, so that each grid has its own effective critical rainfall amount.

[0057] The key idea behind this step is that areas with higher susceptibility to landslides are more sensitive to rainfall, and therefore require a lower rainfall threshold to trigger landslides; conversely, areas with lower susceptibility are less sensitive to rainfall. Therefore, a rainfall reduction factor is introduced. Specifically: For each time window Set the basic critical rainfall for the region A reduction function is constructed based on the background susceptibility index. Used to generate rainfall reduction factor ,satisfy: ;in, For grid cells In the time window The effective critical rainfall. , For reduction function This function will use the landslide susceptibility index. Mapped to rainfall reduction factor The design should meet the following requirements: when (Completely non-slip) , basically no reduction; when (Extremely slippery) For smaller values, the overall function is monotonically decreasing. Its functional form is: . This is a parameter used to control the degree of reduction, namely the reduction sensitivity coefficient. Let... At that time, if ,but: This indicates that the effective critical rainfall for this unit is 60% of the baseline value.

[0058] The reduction sensitivity coefficient is one of the most critical control parameters in the process of mapping the background susceptibility index to the rainfall reduction coefficient. It controls the intensity of rainfall threshold adjustment under different susceptibility levels. Its value is not fixed, but is obtained by fitting and analyzing historical data.

[0059] Collect a dataset of historical landslide events, including: Location of the landslide; The baseline susceptibility index at this location is calculated by the model; The actual cumulative rainfall before the event occurred occurred within different time windows; These samples were grouped and divided according to the susceptibility index range, and the average rainfall that triggered landslides in each group was calculated. By comparing the average rainfall of these groups with the regional unified baseline critical rainfall, we can deduce the optimal λ value that makes the following formula hold: The optimal fit is obtained using methods such as minimum mean square error or maximum likelihood estimation. .

[0060] S25, Traverse all grid cells and all time windows Perform the above correction steps to generate each pair The corresponding effective critical rainfall: Finally, a set of effective critical rainfall amounts corresponding to a dual index of spatial location and time window is constructed. This provides a threshold reference for the next step of standardizing the calculation of the rainfall trigger index.

[0061] S3. Based on the real-time rainfall data, calculate the sliding cumulative rainfall for each time window, and combine it with the effective critical rainfall set to construct a standardized rainfall trigger index for the corresponding time window. Based on the generated rainfall trigger index, determine the consistency and persistence of the rainfall trigger state for each time window, identify whether the rainfall trigger is in a stable over-threshold state, and thus impose reliability constraints on the rainfall trigger index to form the final rainfall-induced intensity characterization value.

[0062] S31, Sliding Accumulated Rainfall Calculation: In landslide-induced monitoring, knowing only the rainfall intensity at a specific moment is insufficient. More crucial is understanding whether there has been continuous heavy rainfall over a past period, as this is one of the core triggering factors for landslides. Therefore, sliding accumulated rainfall data is processed. This is done for each grid cell. Based on its real-time rainfall data series, according to multiple preset time window lengths Perform a sliding summation to calculate the sliding cumulative rainfall within each time window: ;in, Indicates at time , grid In the time window The sliding cumulative rainfall, Indicates time Instantaneous rainfall, For the first The length of each time window, such as 3h, 6h, 12h, 24h, etc., is used to calculate the lower limit of the summation. Indicates the start time of the sliding window.

[0063] S32, Standardized Rainfall Trigger Index Calculation: In landslide early warning, determining whether rainfall is sufficient to trigger a landslide cannot solely rely on the number of millimeters of rainfall. For example, 50mm of rain may exceed the landslide threshold in areas with fragile geological conditions and steep slopes, while in areas with stable rock and good drainage, it may be far from the risk level. Therefore, it is necessary to introduce a site-specific effective critical rainfall amount, which is the dividing line for each location and time window to determine whether the threshold has been exceeded (the threshold in S24). Then, the current actual cumulative rainfall can be divided by the local critical rainfall amount to obtain a ratio. This ratio is the Standardized Rainfall Trigger Index. This index transforms the original rainfall amount into a dimensionless, comparable risk indicator, enabling a fairer measurement of the trigger risk at different locations.

[0064] Specifically, this includes: calculating a standardized rainfall trigger index based on the aforementioned sliding cumulative rainfall and the generated effective critical rainfall. ;in, For grid In the time window ,time Standardized rainfall trigger index, For grid In the time window The effective critical rainfall.

[0065] like The rainfall has not yet reached the trigger threshold; like Rainfall has exceeded the critical threshold, posing a risk of triggering landslides. and The larger the size, the higher the potential for rainfall to induce precipitation.

[0066] S33, Consistency Judgment of Triggering States: In landslide induction analysis, rainfall exceeding a threshold within a single time window is indeed a risk signal, but it may not be stable enough. This is because, on the one hand, different time windows represent different responses to short-term heavy rainfall or long-term cumulative rainfall; on the other hand, in some scenarios, a certain time window exceeds the threshold while other time windows do not, indicating that the risk is not stable enough or is accidental. In order to further determine whether the risk is real and credible, a comprehensive judgment must be made across time windows. This is the purpose of consistency judgment, which analyzes whether this location generally exhibits a high risk under multiple different time windows.

[0067] Specifically, this includes: for grid cells At the present moment Analysis of standardized rainfall triggering indices under multiple time windows ,For example: 3-hour time window: Index = 0.95 (not exceeded); 6-hour time window: Index = 1.20 (over the threshold); 12-hour time window: Index = 1.10 (over the threshold); 24-hour time window: Index = 1.30 (beyond the threshold); Statistical calculations satisfy the threshold condition The time window ratio is expressed as: ;in: For grid cells At any moment The consistency discrimination ratio, This is an indicator function; it returns 1 if the condition is met, and 0 otherwise. The total number of all time windows, if: Then it is determined that the grid has a consistent trigger state across time scales at the current moment. The value is set to 0.7, meaning that the threshold is exceeded 70% of the time window.

[0068] S34, following S33, determines whether a consistent triggering state exists at a certain moment. That is, whether the rainfall triggering threshold is generally exceeded across multiple time windows at that moment. However, observing consistency only at one moment is insufficient to conclude that the area is truly in a dangerous state, as it could be due to short-term fluctuations, data errors, or simply a few time windows exceeding the threshold. Therefore, it is also necessary to determine whether this consistent state has persisted for a period of time. If a grid cell consistently exceeds the threshold across multiple time windows, then the risk is stable and persistent, and its credibility is higher.

[0069] Therefore, the specific solution is: review the past consecutive... If, at any given moment, such as the consistency judgment results of the past 6 hours, the consistency condition is met across all review times, then it is determined to be in a stable overthreshold state. This formula means: only if the sum of the last M time steps is... Both are greater than The product must be 1 for the state to be valid; otherwise, at least one element in the product must be 0, resulting in an overall result of 0, indicating that the state is not persistent. If the product is 1, it means that in the most recent... If the consistency criterion is consistently met within a given time period, it indicates that the triggered state is persistent in time. It is the final persistence discrimination result, representing the grid cell. At any moment Is it in a stable overthreshold state? "t" represents the time range of the review, i.e., the most recent consecutive M moments. It is the first Consistency index at any given time. It is the threshold for consistency judgment. It is an indicator function; its value is 1 when the condition within the parentheses is true, and 0 when it is false.

[0070] S35, in the preceding steps, has completed the following: S32: The standardized rainfall triggering index was calculated for the current time and different time windows; S33: Determines whether these triggering indices are consistent across different time windows; S34: Further determined whether this triggering state is continuous in time.

[0071] However, at this point, we only have some intermediate variables. In order to eventually use them for functions such as risk level determination, spatial visualization, and landslide early warning, we need to give a unified value for the comprehensive risk intensity of each grid cell, namely the rainfall-induced intensity characterization value, which indicates how likely it is that a landslide will be triggered at this location due to rainfall at the current moment. This value takes into account both rainfall intensity and spatiotemporal stability.

[0072] Combining S32-S34, for each grid cell at time... The final rainfall-induced intensity was characterized as follows: Case 1: If and and If the standardized trigger index is >1 and meets the consistency and persistence criteria, it is marked as a stable exceedance, indicating that the area not only has current rainfall exceeding the standard, but also consistently exceeds the standard across multiple time windows, and this has been ongoing for some time. This is the most typical truly dangerous state. Final trigger intensity: ;in, This indicates the time window that triggers the largest exponent; Case 2: If only the following conditions are met If the trigger index is greater than 1, but the consistency or persistence conditions are not met simultaneously, it indicates that the current rainfall exceeds the threshold. However, this exceedance is not stable or comprehensive; it may only be a short-term fluctuation or a local anomaly. To avoid false alarms or over-warnings, this risk value needs to be appropriately lowered, hence the introduction of a decay factor. : ; Scenario 3: If There is no risk of rainfall triggering the event. ; in, This represents the final rainfall-induced intensity characterization value. This is the reliability attenuation factor, set to 0.7. This means a 30% reduction in confidence level to prevent false positives, while still retaining risk information because the index remains greater than 0. This is the baseline value for the default risk-free state; it can be set to 0 or other small values ​​for consistent layer visualization.

[0073] S4. The baseline susceptibility index and the rainfall-induced intensity characterization value are multiplied grid by grid to generate a comprehensive landslide risk index for each grid cell, and different risk levels are assigned according to a preset classification rule.

[0074] S41, Grid-by-grid product calculation: For each grid cell within the monitoring area... Its baseline susceptibility index Characteristic values ​​of rainfall-induced intensity at corresponding times Multiplying these together yields the initial overall risk value: ;in, This represents the background susceptibility index of the grid cell. This represents the intensity characterization value induced by rainfall. This is the initial comprehensive risk value for this grid cell, which has not yet been normalized. This product reflects the combined effect of static background risk and dynamic rainfall risk; if any factor is low, the overall risk will also be low, emphasizing the filtering effect of spatial factors. S42, Index Normalization: To unify the initial risk values ​​across different grids into a comparable standardized range, all initial composite risk values ​​are normalized to obtain a standardized composite landslide risk index. ;in, This is the normalized comprehensive landslide risk index. This represents the minimum initial comprehensive risk value among all grid cells within the current monitoring area at time t. This represents the maximum value at the current moment.

[0075] S43, Risk Level Classification: Based on practical application needs, the comprehensive landslide risk index range [0,1] is divided into several risk level intervals, for example: ; in, As the risk level threshold, satisfying The classification of landslide levels can be flexibly set based on historical landslide experience, distribution characteristics, or early warning needs. Typical values ​​are: , , .

[0076] S44, apply the above division rules to each grid cell, based on its comprehensive landslide risk index. The range in which a stock falls is assigned a corresponding risk level label: ;in, For grid cells At any moment Risk level labels: Low / Medium / High / Very High This is the interval classification function performed according to the aforementioned level thresholds.

[0077] S5. The risk level results are displayed spatially and can be dynamically updated in chronological order. When the user selects any monitoring unit, risk interpretation information based on the background susceptibility, rainfall-induced state and its stability judgment results is output.

[0078] S51: Based on a Geographic Information System (GIS) platform, the risk level assigned to each raster cell is used as the basis for visualization, and rendering is performed according to a preset color coding scheme. Different risk levels correspond to different colors or color bands, for example: Low risk → Light green; Medium risk → Yellow; High risk → Orange; Extremely high risk → Red; The rendering process will divide each raster unit Risk level The mapping is converted into color codes to generate a spatial distribution thematic map of landslide risk levels, thereby realizing a spatial visualization of landslide risk in the monitoring area.

[0079] S52 establishes a time-series risk map display mechanism based on a time axis, linking the thematic maps of the spatial distribution of landslide risk levels for each time segment with the time axis controls. When users drag or select time nodes on the platform interface, the system dynamically loads and displays the thematic map for the corresponding time, achieving the following functions: Replay of risk situations at different times; Visualization of the spatiotemporal evolution of landslide risk; It helps determine whether a risk has a tendency to spread, transfer, or accumulate.

[0080] This process relies on risk level data that is continuously generated by the system. Furthermore, the time control and spatial layer are synchronized and linked in the GIS front end.

[0081] S53: In the GIS platform interface, when a user specifies a target area through interactive operations such as clicking or selecting, which can be a single raster cell, an administrative division, or a sub-region of a watershed, the risk interpretation module is automatically triggered. This module extracts multi-source analysis data and intermediate results related to the cell and generates a structured interpretation report. The report content should include at least: 1. The basic spatial positioning information of this unit; 2. Baseline susceptibility index and its corresponding susceptibility level; 3. Standardized rainfall triggering index sequences for current and recent times; 4. Determine the consistency and persistence of the aforementioned rainfall triggering states; 5. The determined final rainfall-induced intensity characterization value; 6. The comprehensive landslide risk index and the final risk level; 7. A comprehensive risk assessment and prevention recommendations based on the above information.

[0082] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0083] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A landslide monitoring platform based on multi-source data, characterized in that, The platform operation includes performing the following monitoring: S1, acquire spatial background data and real-time meteorological data of the monitoring area. The spatial background data includes historical landslide samples, topographic and geological features and remote sensing indices. The real-time meteorological data includes rainfall sequences for each time window in the recent period. S2, based on the spatial background data, a background susceptibility index for each grid cell is generated through a pre-trained machine learning model, and the rainfall threshold for each time window is corrected according to the background susceptibility index to obtain the effective critical rainfall for the corresponding time window, forming an effective critical rainfall set corresponding to each grid cell. The machine learning model adopts an improved LightGBM model based on gradient boosting tree, and its training sample set includes spatial environmental variables corresponding to known landslide events. S3. Based on the real-time rainfall data, calculate the sliding cumulative rainfall for each time window, and combine it with the effective critical rainfall set to construct a standardized rainfall trigger index for the corresponding time window. Based on the generated rainfall trigger index, determine the consistency and persistence of the rainfall trigger state for each time window, identify whether the rainfall trigger is in a stable over-threshold state, and thus impose reliability constraints on the rainfall trigger index to form the final rainfall-induced intensity characterization value. S4, the baseline susceptibility index and the rainfall-induced intensity characterization value are multiplied grid by grid to generate a comprehensive landslide risk index for each grid cell, and different risk levels are assigned according to a preset classification rule; S5. The risk level results are displayed spatially and can be dynamically updated in chronological order. When the user selects any monitoring unit, risk interpretation information based on the background susceptibility, rainfall-induced state and its stability judgment results is output.

2. The landslide monitoring platform based on multi-source data according to claim 1, characterized in that, The acquisition of the spatial baseline data includes: importing the historical landslide logging database of the geological survey department to obtain the location and boundary information of historical landslide samples; extracting topographic feature factors from digital elevation model data, extracting geological feature factors from regional geological map data, and performing index calculations on multi-temporal remote sensing images to generate a remote sensing index layer; and performing spatial registration, resampling, and normalization processing on the historical landslide samples, topographic and geological feature factors, and remote sensing index layer under a unified spatiotemporal reference to form a spatial baseline dataset that corresponds one-to-one with the raster units of the monitoring area.

3. The landslide monitoring platform based on multi-source data according to claim 1, characterized in that, The acquisition of real-time meteorological data includes: accessing real-time rainfall monitoring data provided by meteorological departments or sensor networks deployed in the monitoring area through a data interface; and dynamically calculating and updating the recent rainfall sequence corresponding to each grid unit based on multiple preset time windows.

4. The landslide monitoring platform based on multi-source data according to claim 1, characterized in that, The construction of the training sample set includes taking the grid cells corresponding to the location of the historical landslide as positive samples, randomly selecting grid cells in areas where no landslides have occurred as negative samples, and extracting the topographic and geological features and remote sensing indices corresponding to all samples as model feature variables. The improved LightGBM model is trained using the training sample set, wherein the improvement includes introducing a weighting term for spatial sample imbalance in the loss function and performing recursive feature elimination by ranking feature importance to optimize the input feature subset. After training, a background susceptibility prediction model is obtained.

5. The landslide monitoring platform based on multi-source data according to claim 4, characterized in that, The generation of the background susceptibility index includes inputting the spatial background dataset corresponding to all grid cells in the monitoring area into the background susceptibility prediction model, and the model outputs the probability value of landslides occurring in each grid cell. The probability value is the background susceptibility index of the grid cell.

6. The landslide monitoring platform based on multi-source data according to claim 5, characterized in that, The level correction of the rainfall threshold includes: defining a regional baseline critical rainfall for each preset time window; establishing a level correction function based on the background susceptibility index, mapping the background susceptibility index of the grid cell to a rainfall reduction coefficient between 0 and 1, wherein the higher the background susceptibility index, the smaller the reduction coefficient; multiplying the regional baseline critical rainfall by the rainfall reduction coefficient of the grid cell to obtain the effective critical rainfall of the grid cell in that time window; Iterate through all grid cells and all preset time windows, repeatedly adjust the rainfall threshold level, and generate the effective critical rainfall for each grid cell under different time windows, thereby forming a set of effective critical rainfall corresponding one-to-one with the grid cells and time windows.

7. The landslide monitoring platform based on multi-source data according to claim 1, characterized in that, S3 specifically includes: S31, for each grid cell, according to its real-time rainfall sequence, a sliding summation calculation is performed according to the length of each preset continuous time window to obtain the sliding cumulative rainfall of the grid cell at the current moment corresponding to each time window; S32, for each grid cell and each time window, divide the sliding cumulative rainfall corresponding to the time window by the effective critical rainfall to obtain a ratio, which is the standardized rainfall triggering index of the current grid cell under the time window; when the ratio is less than or equal to 1, it indicates that the rainfall triggering condition has not been met; when the ratio is greater than 1, it indicates that the rainfall triggering condition has been met, and the larger the ratio, the higher the rainfall induction potential.

8. The landslide monitoring platform based on multi-source data according to claim 7, characterized in that, S3 further includes: Consistency determination of triggering state: For the current moment, analyze the multiple standardized rainfall triggering indices calculated for the same grid cell under different time window lengths; if the index values ​​of the time windows exceeding a preset proportion are all greater than 1, then it is determined that the rainfall triggering state of the grid cell at the current moment has consistency across time scales. Triggering state persistence determination: Review the rainfall triggering state of the grid cell at multiple consecutive moments in the recent period; if the consistency determination result is consistently positive in the recent period, it is determined that the rainfall triggering of the grid cell is in a stable over-threshold state; Based on the combined results, for each grid cell, if its current standardized rainfall trigger index is greater than 1 and simultaneously satisfies both consistency and persistence criteria, it is marked as a stable overthreshold, and its standardized rainfall trigger index is directly used as the final rainfall-induced intensity characterization value; if its index is greater than 1 but does not satisfy consistency or persistence criteria, the index is constrained and reduced by multiplying it by a reliability attenuation coefficient to obtain the final rainfall-induced intensity characterization value; if its index is less than or equal to 1, the final rainfall-induced intensity characterization value is 0 or a baseline value representing no risk.

9. The landslide monitoring platform based on multi-source data according to claim 1, characterized in that, S4 specifically includes: S41, For each grid cell within the monitoring area, the background susceptibility index of the grid cell is multiplied by the rainfall-induced intensity characterization value to obtain the initial comprehensive risk value of the grid cell. S42, normalize the initial comprehensive risk value of all grid cells in the region and map it to a continuous interval from 0 to 1 to obtain the comprehensive landslide risk index of each grid cell; S43, set multiple risk level thresholds to divide the value range of the comprehensive landslide risk index into corresponding risk level intervals; the risk levels include at least low risk, medium risk, high risk and extremely high risk; S44. Assign a corresponding risk level to each grid cell based on the risk level range of the comprehensive landslide risk index.

10. The landslide monitoring platform based on multi-source data according to claim 1, characterized in that, S5 specifically includes: S51, based on the geographic information system platform, renders each grid cell according to the assigned risk level and different preset color symbols to generate a thematic map of the spatial distribution of landslide risk level in the monitoring area. S52, in the geographic information system platform, a time axis control is used to associate thematic maps of the spatial distribution of landslide risk levels corresponding to different calculation times; when the time axis control is operated, the platform interface dynamically switches and displays the risk spatial distribution at the corresponding time, realizing the playback and tracking of the spatiotemporal evolution of risk. S53, when a user selects any grid cell or aggregated administrative / watershed management unit on the platform interface through interactive operation, the information query and interpretation module is triggered to automatically extract and organize multi-source analysis data and intermediate results related to the corresponding unit, and generate a structured risk interpretation report.

Images