Intelligent dynamic monitoring and early warning method for rainfall type landslide-debris flow chain disaster

Abstract

The invention belongs to the technical field of geological disaster monitoring and early warning, and relates to a rainfall type landslide-debris flow chain disaster intelligent dynamic monitoring and early warning method, which comprises the following steps: data acquisition: generating multi-source data including an optical image, an SAR image, an unmanned aerial vehicle image and ground data; preprocessing the multi-source data; performing space-time registration, data format and organization on the multi-source data after data processing; key characteristic parameters are calculated, and data fusion is carried out; constructing a multi-dimensional feature vector, and generating a multi-dimensional feature image; constructing a risk assessment model by using a deep learning model, and performing real-time prediction; dynamically determining the total risk level of each area according to a preset comprehensive research and judgment rule by integrating the real-time prediction result; and when the total risk level reaches a'high risk 'or'extremely high risk' threshold value, generating early warning information, and performing early warning release. According to the invention, comprehensive, whole-process three-dimensional perception, dynamic and self-adaptive intelligent evaluation and disaster prediction are realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geological disaster monitoring and early warning, and particularly relates to an intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters. BACKGROUND

[0002] Rainfall-induced landslide-debris flow chain disasters are common and extremely destructive geological disasters in China and even the world. The typical disaster formation mode is as follows: sustained heavy rainfall induces slope instability to form a landslide, and the landslide material is disintegrated and saturated during migration and further transformed into a debris flow, forming a chain disaster with a wider range of destruction and greater harm.

[0003] At present, the traditional monitoring and early warning methods have obvious limitations: first, they rely on a single data source (such as only relying on rainfall or only relying on displacement), which makes it difficult to fully capture the complete chain of disaster evolution and precursor information; second, they use static empirical thresholds as early warning criteria, which cannot dynamically reflect the real-time state changes of the disaster body, resulting in high false alarm and missed alarm rates; third, they analyze data separately, lack effective multi-source information fusion mechanisms, and result in insufficient early warning accuracy and timeliness. Specifically:

[0004] 1. The existing technology of monitoring and early warning methods based on a single data source relies on a single type of monitoring data and fails to fully utilize the complementarity of multi-source information. For example, a static rainfall threshold model is established for early warning based on the relationship between historical rainfall data and disaster occurrence: this method is simple and easy to implement, but it ignores key factors such as geological environment, soil moisture content, and the cumulative effect of previous rainfall, resulting in poor adaptability to disaster early warning under specific geological conditions and high false alarm and missed alarm rates. The early warning is mainly based on GPS displacement monitoring data by setting a displacement rate threshold: this method directly captures surface displacement, but it cannot perceive precursor information such as deep deformation and soil moisture content changes, and the cost of equipment deployment is high, the coverage is limited, and it is difficult to achieve early warning at a regional scale.

[0005] The evolution of landslide-debris flow chain disasters is a complex system of multiple physical processes (seepage-stress-deformation coupling). Single data source analysis can only reflect one side of the system.

[0006] Relying only on rainfall data cannot perceive the specific response of rock-soil mass (such as pore water pressure rise and deformation acceleration), and cannot distinguish the stability differences of different geological bodies under the same rainfall, resulting in early warning based on "possibility" rather than "certainty" and high false alarm rate.

[0007] Relying only on surface displacement data often only issues an alarm in the deformation acceleration stage (i.e., the pre-slide stage), missing the best opportunity for early warning based on earlier precursors such as deep deformation and moisture content changes, and the early warning time is extremely urgent.

[0008] This "blind men and the elephant" style of monitoring cannot build a complete cognitive map of the entire process of disaster gestation, occurrence, and development, and cannot effectively identify the critical conditions for the chain transformation from landslide to debris flow.

[0009] 2. Early Warning Methods Based on Static Empirical Models: Most existing early warning models rely on historical statistical experience and lack dynamic adjustment and learning capabilities. For example, early warning methods that monitor groundwater level changes and combine them with empirical water level thresholds typically have fixed or seasonally adjusted thresholds, failing to dynamically respond to drastic changes in external conditions such as extreme rainfall and human engineering activities, and exhibiting weak model generalization ability. Many traditional systems employ deterministic mechanical or statistical models, whose parameters require manual setting and adjustment, making them difficult to adapt to complex situations in different regions and with varying geological conditions. The accuracy and reliability of early warning results highly depend on the prior knowledge of experts.

[0010] Early warning models are mostly static empirical models, with poor adaptability and generalization ability. Existing technologies widely use fixed thresholds based on historical statistics (such as critical rainfall and critical displacement rate) as early warning criteria, which are essentially static models.

[0011] Spatiotemporal variability: The spatial heterogeneity of geological conditions (such as soil type, structure, and slope) leads to significant differences in critical thresholds at different locations. Empirical thresholds for one region are difficult to apply directly to another region, resulting in poor universality.

[0012] Dynamic evolution: The properties of the disaster body itself (such as the degree of fracture development and shear strength decay) evolve dynamically over time. Static thresholds cannot reflect this dynamic change and cannot achieve timely early warning. In cases of continuous rainfall or when previous rainfall has led to increased soil saturation, the critical rainfall amount that triggers the disaster will decrease significantly, and static models cannot dynamically adjust this threshold.

[0013] Therefore, static models are prone to failure in unfamiliar areas or under extreme operating conditions, resulting in a large number of false alarms (threshold too low) or false alarms (threshold too high).

[0014] 3. Separate Analysis of Multi-Source Data, Lack of Deep Integration: Although some advanced technologies have begun to incorporate multi-source data, they mostly remain at the level of separate processing and result comparison, failing to achieve deep integration and intelligent mining of information. For example, while using satellite remote sensing and ground monitoring data simultaneously, the two types of data are usually processed and analyzed separately first (e.g., InSAR deformation calculation, ground sensor analysis of rainfall), and then the results are simply overlaid or compared. This fails to integrate the data at the underlying data level, to uncover the intrinsic correlations and co-evolutionary patterns between multiple parameters, and to form a holistic understanding of the disaster chain process.

[0015] Data analysis is isolated and lacks mechanisms for multi-source information fusion and collaborative analysis. Even with the deployment of multiple sensors, existing technologies often adopt a "data silo" analysis model, that is, each data source is analyzed and thresholded independently, with only simple logical superposition at the very end (such as triggering an alarm when "rainfall exceeds the threshold and displacement exceeds the threshold").

[0016] Information Redundancy and Conflicts: While multi-source data can be complementary, they also contain redundancies and even conflicts. Isolated analysis cannot effectively identify and utilize these relationships. For example, a significant displacement signal may be triggered by a landslide or by other disturbances (such as construction or vegetation swaying). Isolated analysis of displacement data cannot make an accurate judgment, but if rainfall and soil moisture data for the area can be analyzed simultaneously (to determine if they also reach high-risk levels), the accuracy of identification can be significantly improved.

[0017] Failure to uncover deep-seated correlations: The evolution of disasters is essentially a process of coordinated changes in multiple parameters. For example, accelerated displacement is often accompanied by a sudden increase in pore water pressure. Isolated analysis cannot uncover the inherent, nonlinear coupling relationships between these parameters, yet these very relationships contain the core mechanisms of disaster evolution. Current technologies lack the ability to extract unified, high-dimensional features from multi-source data, resulting in a significant waste of information value and an inability to quickly infer the occurrence of disaster chains from existing data.

[0018] 4. Isolated System Architecture and Delayed Early Warning Information Dissemination: Existing technical solutions often focus on data collection and simple analysis, failing to form a closed-loop system from perception, analysis, decision-making to dissemination. Many deployed systems have independent data collection, transmission, processing, and analysis processes with low automation, requiring significant manual intervention. This results in an excessively long cycle from data acquisition to early warning information generation, failing to meet the requirements of real-time dynamic early warning. Furthermore, early warning information dissemination channels are limited, typically confined to internal professional platforms, failing to effectively link with public emergency response systems, leading to a prominent "last mile" problem in early warning information delivery.

[0019] Therefore, there is an urgent need for a new method that can integrate observation data from multiple platforms (space, air, and ground) and intelligently and dynamically assess the risk of rainfall-induced landslide-debris flow chain disasters, thereby achieving accurate early warning. Summary of the Invention

[0020] The purpose of this invention is to provide an intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters. This method achieves comprehensive and full-process three-dimensional perception, dynamic and adaptive intelligent assessment, and forward-looking chain disaster prediction. At the same time, it realizes an automated and intelligent decision-making closed loop, which can process data and make decisions in real time.

[0021] Technical solution to achieve the purpose of this invention:

[0022] A method for intelligent dynamic monitoring and early warning of rainfall-induced landslide-debris flow chain disasters includes:

[0023] S1. Data Acquisition: Acquire satellite data, UAV data, and ground sensor monitoring data to generate multi-source data including optical images, SAR images, UAV images, and ground data;

[0024] S2. Data Processing: Preprocess the multi-source data to generate optical DOM, geocoded backscattering coefficient map, UAV DOM and DEM, and a set of surface observation data values ​​arranged in chronological order;

[0025] S3. Formatted Data: Spatiotemporal registration, data formatting, and organization are performed on the multi-source data after data processing to obtain multi-source data with a unified and standardized data format;

[0026] S4. Feature Extraction: Calculate key feature parameters and perform data fusion on them;

[0027] S5. Feature Construction: Construct multi-dimensional feature vectors and generate multi-dimensional feature images;

[0028] S6. Model Building: Using a deep learning model, a risk assessment model is built based on the multi-dimensional feature vector constructed in S5 to perform real-time prediction.

[0029] S7. Risk Assessment: Based on real-time forecast results and preset comprehensive assessment rules, dynamically determine the total risk level of each region.

[0030] S8. Warning Information Release: When the overall risk level reaches the "high risk" or "extremely high risk" threshold, a warning information is generated and a warning is released.

[0031] Further, S2 includes:

[0032] S2.1 Optical image preprocessing to generate an optical orthophoto with accurate planar position;

[0033] Radiometric calibration: converting the raw digital quantization values ​​of high-resolution optical images into top-atmosphere radiance values ​​or surface reflectance;

[0034] Atmospheric correction: converting radiance values ​​or surface reflectance into true surface reflectance;

[0035] Orthorectification: Orthorectifying the true surface reflectance to generate an optical orthophoto with accurate planar position;

[0036] S2.2 SAR image preprocessing to generate geocoded backscattering coefficient map;

[0037] Radiometric calibration: converting the intensity information of single-look complex data from SAR satellite imagery into backscattering coefficients;

[0038] Speckle noise filtering: An adaptive filtering algorithm is used to process SAR images to suppress their inherent speckle noise while preserving edge and texture details to the greatest extent possible.

[0039] Geocoding: Based on satellite precise orbit data and DEM data, SAR images are accurately transformed from slant range projection coordinate system to ground range projection coordinate system through range-Doppler model or rational polynomial coefficient model, eliminating terrain distortion and generating geocoded backscattering coefficient map.

[0040] S2.3. UAV image preprocessing to generate UAV digital orthophoto maps and digital elevation models;

[0041] Aerial triangulation and dense matching: Aerial triangulation is performed on the sequence of images acquired by UAVs to solve the high-precision exterior orientation elements of each image, and a high-density point cloud is generated through a dense matching algorithm;

[0042] Generate digital products: Based on point cloud data, generate high-resolution UAV digital orthophoto maps and digital elevation models;

[0043] S2.4 Ground data preprocessing: The measurements from rain gauges, soil moisture meters, pore water pressure sensors, and GNSS displacement meters are used to generate a set of surface observation data values ​​arranged in chronological order.

[0044] S2.4.1 Rain gauge measurement and data preprocessing to generate standardized hourly rainfall time series;

[0045] S2.4.2 Soil moisture meter measurement, data preprocessing, and generation of standardized soil moisture content and change rate time series;

[0046] S2.4.3 pore water pressure sensor measurement, data preprocessing, and generation of standardized pore water pressure and rate of change time series;

[0047] S2.4.4 GNSS displacement gauge measurement, data preprocessing, and generation of standardized displacement rate and acceleration time series;

[0048] Finally, a data set was formed with 1-hour intervals for hourly rainfall, cumulative rainfall, soil moisture content, rate of change of moisture content, deformation rate, deformation acceleration, hourly pore water pressure, and rate of change of pore water pressure.

[0049] Further, S3 includes:

[0050] S3.1 Multi-source data spatiotemporal registration: unify the multi-source data after S2 data processing to the same spatiotemporal reference;

[0051] S3.2 Data Format and Organization: The multi-source data processed in S3.1 is converted into a unified and standardized data format, and organized and managed using a spatiotemporal grid data model to obtain multi-source data with a unified and standardized data format.

[0052] Further, S3.1 includes:

[0053] S3.1.1 Unification of Spatial Reference Standards:

[0054] Coordinate System 1: Convert all data processed from S2 data from plane coordinate system 1 to geodetic coordinate system, and unify the elevation system to geodetic height system;

[0055] Pixel resampling: Resample the raster data in all data after S2 data processing to the same spatial resolution;

[0056] Precise registration: Using a high-precision image that has undergone orthorectification as the reference image, the affine transformation or polynomial transformation model between other images and this reference image is calculated through feature point matching or region matching algorithms to achieve sub-pixel level geometric precision registration.

[0057] S3.1.2, Unified Time Base:

[0058] The timestamps of all data collected after S2 data processing will be uniformly converted to Coordinated Universal Time.

[0059] For data acquired at different times, the time difference should be recorded and considered during time series analysis.

[0060] Further, S4 includes:

[0061] S4.1 Feature extraction: Extract feature parameters such as deformation rate, NDVI change, surface slope, hourly rainfall, and soil moisture content change rate to form multi-dimensional feature parameters.

[0062] S4.2 Feature fusion is used to transform heterogeneous features extracted from multiple platforms ("air-space-ground") into structured data that can be directly input into deep learning models.

[0063] Furthermore, the key feature parameters in S4 include: deformation features, vegetation and land cover features, topographic features, and hydrological and mechanical features; deformation features include InSAR deformation rate, InSAR deformation acceleration, GNSS deformation rate, and GNSS deformation acceleration; vegetation and land cover features include NDVI variation and backscattering coefficient variation; topographic features include land slope, aspect, topographic curvature, and topographic humidity index; hydrological and mechanical features include hourly rainfall, cumulative rainfall, soil volumetric water content change rate, and pore water pressure; and feature-level fusion of the above key feature parameters is performed using "pixel-alignment and band-stacking".

[0064] Further, S5 includes:

[0065] S5.1 Feature Vector Construction: Each of the four multi-dimensional features extracted in S4 is regarded as an independent dimension; for each cell or each regular grid in the target area, the values ​​of all its corresponding feature parameters are combined in a fixed order to form a high-dimensional feature vector.

[0066] S5.2 Multidimensional Image Storage: Reorganize the feature vectors of all pixels in space and store them in a multidimensional image or data cube format.

[0067] Further, S6 includes:

[0068] S6.1 Construction and Training of Risk Assessment Model

[0069] S6.2, the online deployment phase of dynamic risk assessment based on multi-model fusion, utilizes the model trained in S6.1 for real-time prediction.

[0070] Further, S6.1 includes:

[0071] S6.1.1 Training Dataset Construction: Collect multi-dimensional feature vector datasets from historical periods and their corresponding disaster occurrence labels that have been verified in the field. Clean, enhance, and divide the datasets to form training sets, validation sets, and test sets.

[0072] S6.1.2 Training a deep convolutional neural network model to extract spatial context information and local features from multi-dimensional feature images and output a spatial probability map of landslide occurrence;

[0073] S6.1.3 Training of Long Short-Term Memory Network Model: This is used to process time-series data, capture the long-term dependencies of data in the time dimension, and output the probability of a landslide turning into a debris flow within a specific time window in the future.

[0074] S6.2 includes:

[0075] S6.2.1 Landslide Spatial Probability Prediction: The latest multi-dimensional feature image generated in real time is input into the trained deep convolutional neural network model to generate and output the spatial probability map of landslide disaster in the study area at the current moment.

[0076] S6.2.2 Chain Disaster Probability Prediction: For high-risk landslide areas or key valley units, extract their temporal feature sequences, input them into a pre-trained long short-term memory network model, and output the probability of a "landslide-debris flow" chain disaster occurring in the area.

[0077] Further, S8 includes:

[0078] S8.1 Warning Generation: When the system determines that the total risk level of any area reaches the "high risk" or "extremely high risk" threshold, the warning generation mechanism is automatically triggered to generate warning information.

[0079] S8.2 Warning Release: Warning information is automatically pushed to the geological disaster warning information platform through the application programming interface. The platform releases the warning information to relevant management departments and threatened people through multiple channels as soon as possible.

[0080] The beneficial technical effects of this invention are as follows:

[0081] 1. This invention provides an intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters, achieving comprehensive and full-process three-dimensional perception, and solving the fundamental problem of "partial monitoring and incomplete information." Current technology relies heavily on single data sources (such as rainfall or displacement alone), like "blind men touching an elephant," failing to comprehensively capture the physical precursor signals at different stages and levels during the evolution of the disaster chain (landslide-debris flow). The principle and effects of this invention: This invention innovatively integrates observation data from multiple platforms including air (satellites), space (UAVs), and ground (sensor networks), simultaneously acquiring multi-dimensional information such as deformation (InSAR), land cover (optical), moisture (SAR, sensors), topography (DEM), and rainfall (rain gauges) from a physical mechanism perspective. This enables the system not only to perceive the deformation of the landslide body (the result), but also to capture its inherent driving factors (such as a sharp increase in soil moisture content) and potential chain effects (such as changes in the dynamic storage of material sources in the gully). This achieves a three-dimensional and holographic perception of the entire process of chain disaster "gestation-occurrence-development", fundamentally eliminating information blind spots.

[0082] 2. This invention provides an intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters, achieving dynamic and adaptive intelligent assessment and solving the core pain point of "static models and poor adaptability" in existing technologies. Most existing early warning models are based on static empirical thresholds (such as critical rainfall), with fixed parameters, unable to respond to the spatiotemporal variability of geological conditions and the dynamic evolution of disaster bodies, and are prone to failure in unfamiliar areas or extreme conditions. The principle and effects of this invention: This invention utilizes deep learning models (CNN and LSTM) to autonomously learn the complex nonlinear laws of disaster evolution from massive historical data. CNN networks excel at mining spatial correlation patterns between multi-source features, accurately identifying feature combinations related to instability even in different geological environments; LSTM networks are adept at capturing the temporal evolution laws of parameters such as rainfall, deformation, and water content, and can dynamically predict their development trends. This data-driven intelligent model possesses strong adaptive capabilities, with its assessment results updated in real time according to the dynamic changes of input data, greatly improving the accuracy and generalization ability of early warnings in different regions and scenarios.

[0083] 3. This invention provides an intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters, achieving forward-looking prediction of chain disasters and solving the major problems of "isolated analysis and delayed early warning." Current technology mostly treats landslides and debris flows as isolated events, with a single analysis model, failing to reveal the chain relationship between them. Early warnings are often targeted at a single disaster and are usually triggered after deformation acceleration or the disaster itself, resulting in extremely tight warning times. Invention principle and effects: This invention uses LSTM temporal modeling to specifically analyze the temporal correlation between landslide occurrence conditions and subsequent debris flow initiation conditions. The model can determine the probability and critical timing of transformation into a debris flow after the landslide source is initiated, under conditions such as continuous rainfall. This enables the system not only to provide early warning of the primary disaster of "landslide" but also to proactively predict the risk of the chain disaster of "debris flow," achieving true chain disaster early warning and gaining crucial additional time for the evacuation of people in danger zones and emergency response.

[0084] 4. This invention provides an intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters, realizing an automated and intelligent decision-making closed loop. It solves the practical bottleneck of "reliant on manual labor and low efficiency" in existing technologies: traditional methods often require extensive manual intervention from data analysis to early warning decision-making, resulting in low efficiency, high subjectivity, and difficulty in achieving 24 / 7 uninterrupted automatic operation, failing to meet the emergency response needs of sudden disasters. The principle and effects of this invention: This invention constructs a complete automated closed loop from automatic access, fusion processing, feature extraction, model reasoning to automatic generation and dissemination of early warning information. The entire process requires no manual intervention; the intelligent model acts as a "tireless analysis expert," capable of processing data and making decisions in real time. This greatly improves the timeliness, objectivity, and operational capability of early warnings, laying a core technological foundation for building an unattended, all-weather automated monitoring and early warning system. Attached Figure Description

[0085] Figure 1 The flowchart illustrates a method for intelligent dynamic monitoring and early warning of rainfall-induced landslide-debris flow chain disasters provided by this invention. Detailed Implementation

[0086] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0087] The method of this invention involves the collaborative processing of sensor data from multiple platforms (space, air, and ground), the deep fusion of heterogeneous information from multiple sources, a dynamic probability assessment model for disaster risk based on deep learning, and an automatic generation technology for early warning information of chain disasters. It is a comprehensive system method that integrates remote sensing science, sensor technology, Internet of Things, and deep learning.

[0088] like Figure 1 As shown, this invention provides an intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters, specifically including the following steps:

[0089] S1. Data Acquisition: Acquire satellite data, UAV data, and ground sensor monitoring data to generate multi-source data including optical imagery, SAR imagery, UAV imagery, and ground data.

[0090] Satellite data: Acquire GF-1 / 2 optical satellite imagery (for NDVI extraction) and LUTAN-1 SAR satellite imagery (for deformation InSAR solution) covering the target area.

[0091] Drone data: During rainfall, high-resolution images are acquired and DEMs are generated by taking advantage of gaps in the clouds and rain.

[0092] Ground sensor monitoring data: Real-time reception of monitoring data from rain gauges, soil moisture meters, and GNSS displacement meters.

[0093] S2. Data Processing: Preprocess the multi-source data to generate optical DOM, geocoded backscattering coefficient maps, UAV DOM and DEM, and a set of surface observation data values ​​arranged in chronological order.

[0094] This step aims to address issues such as inconsistencies in spatiotemporal references and units of measurement in multi-source heterogeneous remote sensing data, and is a prerequisite for achieving high-precision data fusion and analysis. This step specifically includes the following sub-steps:

[0095] This step involves specialized processing based on the characteristics of different data sources to extract accurate and reliable surface information.

[0096] S2.1 Optical satellite image preprocessing to generate orthophotos with accurate planar positions.

[0097] Radiometric calibration: This involves converting the raw digital quantization (DN) values ​​of high-resolution satellite imagery (such as GF-1 and GF-2) into physically meaningful values ​​for top-level atmospheric radiance or surface reflectance, thus eliminating the influence of differences in sensor response. The calculation formula is as follows:

[0098] Lλ = Gain * DN + Offset

[0099] Where Lλ is the radiance value, DN is the original digital quantization value of the optical image, and Gain and Offset are calibration coefficients provided by the satellite platform.

[0100] Atmospheric correction: Using physical models such as FLAASH and 6S, or empirical models based on dark pixels, the effects of scattering and absorption by atmospheric molecules and aerosols are eliminated, and the radiance values ​​or surface reflectance are inverted to the true surface reflectance. This process is crucial for subsequent quantitative inversion of vegetation indices and surface components.

[0101] Orthorectification: Using satellite orbital parameters, sensor attitude parameters, and digital elevation models (DEMs), orthorectification is performed on the true surface reflectivity to eliminate geometric distortions caused by terrain undulations and sensor perspective geometry, generating orthophotos with accurate planar positions.

[0102] S2.2 SAR satellite image preprocessing to generate geocoded backscattering coefficient maps

[0103] Radiometric calibration: The intensity information of single-look complex data (SLC) from SAR satellite images such as LUTAN-1 is converted into backscattering coefficients (σ°) using calibration constants provided by the satellites. This gives the backscattering coefficients a clear physical meaning and allows for comparisons between different time phases and regions. The calculation formula is as follows:

[0104] σ°=|DN|2 / K

[0105] Where K is the absolute scaling constant.

[0106] Speckle noise filtering: Adaptive filtering algorithms (such as Refined Lee and Gamma MAP filters) are used to process SAR images to suppress their inherent speckle noise while preserving edge and texture details to the greatest extent possible.

[0107] Geocoding: Based on satellite precise orbit data and DEM data, SAR images are accurately transformed from slant range projection coordinate system to ground distance projection coordinate system through range-Doppler model or rational polynomial coefficient (RPC) model, eliminating terrain distortion and generating geocoded backscattering coefficient map.

[0108] S2.3. UAV image preprocessing to generate high-resolution digital orthophoto maps (DOM) and digital elevation models (DEM).

[0109] Aerial triangulation and dense matching: Aerial triangulation is performed on the sequence of images acquired by UAVs to solve the high-precision exterior orientation elements of each image, and a high-density point cloud is generated through a dense matching algorithm.

[0110] Generate digital products: Based on point cloud data, generate high-resolution digital orthophoto maps (DOM) and digital elevation models (DEM) as the basis for subsequent terrain factor extraction and fine change detection.

[0111] S2.4 Ground data preprocessing: Measurements are performed using rain gauges, soil moisture meters, pore water pressure sensors, and GNSS displacement meters to generate a set of surface observation data arranged in chronological order.

[0112] S2.4.1 Rain gauge measurement and data preprocessing to generate standardized hourly rainfall time series raw data: minute-level rainfall (mm), cumulative rainfall (mm), and rainfall intensity (mm / h).

[0113] Preprocessing steps:

[0114] Data decoding: Convert binary / JSON format into a structured data table (fields: sensor ID, timestamp, rainfall value, signal strength, battery voltage).

[0115] Time synchronization: Calibrated to Coordinated Universal Time (UTC).

[0116] Resampling: Data is standardized to 1-hour intervals (e.g., hourly rainfall).

[0117] Standardization: Units are unified to millimeters (mm) and normalized to [0,1] (e.g., scaled according to the historical maximum rainfall).

[0118] Output: Standardized hourly rainfall time series.

[0119] S2.4.2 Soil moisture meter measurement, data preprocessing, and generation of standardized soil moisture content and change rate time series.

[0120] Raw data: Soil volumetric water content (%), soil temperature (°C), sensor burial depth (cm).

[0121] Preprocessing steps:

[0122] Data decoding: Similar to rain gauge, it is parsed into structured data (fields: sensor ID, timestamp, moisture content, temperature, depth).

[0123] Time synchronization and resampling: Aligned with the rain gauge at 1-hour intervals.

[0124] Standardization: Moisture content converted to international standard units (m) 3 / m 3 ), calculate the rate of change (such as the difference between the current hour and the previous hour).

[0125] Output: Standardized soil moisture content and rate of change time series.

[0126] S2.4.3. Pore water pressure sensor measurement, data preprocessing, and generation of standardized pore water pressure and rate of change time series.

[0127] Raw data: Frame signals containing sensor ID, local timestamp, pore water pressure u (kPa), temperature T (°C), burial depth (m), signal strength RSSI, battery voltage Vbat (V), etc.

[0128] Preprocessing steps: Extract the value of each field in the frame signal according to the field name and convert it into a structured data table.

[0129] Time synchronization and resampling: Resampling is uniformly done at 1-hour intervals.

[0130] Standardization: Z-score standardization is performed on the hourly change in pore water pressure.

[0131] Output: Time series of standardized pore water pressure and rate of change.

[0132] S2.4.4 GNSS displacement gauge measurement, data preprocessing, and generation of standardized displacement rate and acceleration time series.

[0133] Raw data: 3D displacement (mm, North / East / Elevation direction), displacement rate (mm / day), satellite signal quality (e.g., PDOP value).

[0134] Preprocessing steps:

[0135] Data decoding: Parsing into structured data (fields: sensor ID, timestamp, displacement component, velocity, PDOP).

[0136] Time synchronization and resampling: uniformly set to 1-hour intervals (interpolation processing required for high-frequency GNSS data).

[0137] Standardization: The unit of displacement is standardized to millimeters, and deformation acceleration (such as the first-order difference of velocity) is calculated.

[0138] Output: Standardized displacement rate and acceleration time series.

[0139] Finally, hourly rainfall (mm), cumulative rainfall (mm), and soil moisture content (m³) are calculated at 1-hour intervals. 3 / m 3 ), rate of change in water content (m) 3 / m 3 / h), deformation rate (mm / h), deformation acceleration (mm / h) 2 A data set of hourly pore water pressure (kPa) and pore water pressure change rate (kPa / h).

[0140] S3. Formatted data: This involves spatiotemporal registration, data formatting, and organization of multi-source data to obtain multi-source data in a unified and standardized format.

[0141] S3.1 Spatiotemporal Registration of Multi-Source Data

[0142] To achieve pixel-level fusion and analysis, all data must be unified to the same spatiotemporal reference.

[0143] S3.1.1 Unification of Spatial Reference Standards:

[0144] 1. Coordinate System 1: Convert the plane coordinate system of all data after S2 data processing (optical DOM, SAR geocoding products, UAV DOM / DEM, ground sensor locations) to WGS84 or CGCS2000 geodetic coordinate system, and unify the elevation system to EGM96 or CGCS2000 geodetic height system.

[0145] 2. Pixel Resampling: All raster data (optical, SAR, UAV DOM) from the S2 data processing are resampled to the same spatial resolution. Depending on the fusion objective, resampling to the highest resolution (to retain the most detail) or the lowest resolution (to reduce data volume) can be selected. The resampling method can be chosen based on requirements: nearest neighbor (suitable for discrete data), bilinear interpolation (balancing accuracy and efficiency), or cubic convolution (maximizing the preservation of spectral and texture information).

[0146] 3. Precise registration: Using a high-precision image that has undergone orthorectification (such as a UAV DOM) as the reference image, the affine transformation or polynomial transformation model between other images and this reference image is calculated through feature point matching (such as SIFT, ORB algorithm) or region matching algorithm, to achieve sub-pixel level geometric precision registration.

[0147] S3.1.2, Unified Time Base:

[0148] 1. Convert the collection timestamps of all data after S2 data processing to Coordinated Universal Time (UTC) to eliminate time confusion caused by different time zones.

[0149] 2. For data acquired at different times, the time difference must be recorded and considered during time series analysis to ensure the accuracy of multi-temporal analysis.

[0150] S3.2 Data Format and Organization

[0151] The multi-source data processed by S3.1 is converted into a unified and standardized data format (such as GeoTIFF) and organized and managed using a spatiotemporal grid data model to obtain multi-source data in a unified and standardized format. This ensures that each data layer is fully aligned in terms of spatial range and cell size, laying a solid foundation for subsequent data fusion and feature extraction.

[0152] S4. Feature Extraction: Calculate key feature parameters and perform data fusion.

[0153] S4.1 Feature Extraction

[0154] Feature parameters such as deformation rate, NDVI change, surface slope, hourly rainfall, and soil moisture content change rate are extracted to form multi-dimensional feature parameters.

[0155] Deformation characteristics:

[0156] Deformation Rate: This parameter (unit: mm / year) is obtained by processing multi-temporal SAR data using techniques such as SBAS-InSAR to measure the surface deformation rate field of the target area along the radar line of sight. This parameter is the most direct indicator reflecting deep slope creep and potential instability. The specific calculation process is as follows:

[0157] Interferometric processing of the multi-temporal SAR backscattering coefficient map output by S3 was performed using SBAS-InSAR technology:

[0158] (1) Generate interferograms for each image pair, and remove phase from the DEM terrain using precise orbit and DEM;

[0159] (2) Use 30×30 multi-view noise suppression and select stable coherence points using the amplitude deviation threshold (≤0.25);

[0160] (3) Under the constraints of the spatiotemporal baseline (vertical baseline ≤ 300m, time baseline ≤ 72d), a linear system of equations is established and solved by singular value decomposition (SVD) to obtain the radar line-of-sight (LOS) deformation rate v_LOS, in millimeters / year;

[0161] (4) Project v_LOS onto the vertical direction using the local incident angle θ: v_vert=v_LOS / cosθ, and finally generate a 10m resolution deformation rate grid to characterize the deep creep intensity of the slope.

[0162] Deformation Acceleration: Deformation acceleration is obtained by performing second-order differential calculations on time-series deformation data. A sudden increase in acceleration is a very strong signal for impending slippage.

[0163] Based on the deformation rate field, a quadratic polynomial fit is performed along the time dimension for each coherent point: d(t)=a0+a1t+a2t 2 Where a1 is the linear velocity and a2 is the acceleration term; a2 is taken as the deformation acceleration at that point, with units of mm / yr. 2 After fitting the coherent points across the province, Kriging interpolation was used to generate a 10m acceleration grid with the same resolution as the velocity field. When the acceleration of a certain pixel is greater than or equal to the historical average plus 2σ and lasts for more than or equal to 3 time phases, it is marked as "acceleration anomaly" as a key early warning signal for the imminent slip phase.

[0164] The surface deformation rate (mm / h) and deformation acceleration (mm / h) of the surface points obtained by GNSS in step S2.4.4 2 ).

[0165] Vegetation and ground cover characteristics:

[0166] NDVI change (ΔNDVI): Calculated from multi-temporal optical imagery. The formula for NDVI (Normalized Difference Vegetation Index) is: NDVI = (NIR - RED) / (NIR + RED), where NIR represents the reflectance of the target feature in the near-infrared spectral band, and RED represents the reflectance of the target feature in the red visible spectral band. By calculating the difference between the current and previous NDVI values, a sudden decrease or withering of vegetation cover can be identified, which may be indirect evidence of slope slippage and damage to vegetation root systems.

[0167] Change in backscattering coefficient (Δσ°): Calculated from multi-temporal SAR data. A significant increase in the backscattering coefficient (Δσ°>0) can indicate a sharp increase in surface or shallow soil moisture content. Specific calculations are as follows:

[0168] The variation in backscattering coefficients is extracted from the multi-temporal SAR geocoding map output by S3 through the following steps:

[0169] (1) Radiation uniformity correction

[0170] Using the first image as a reference, the gain offset coefficients of the remaining time phases are calculated using the "statistical uniform region" method, and relative radiometric correction is performed image by image to ensure the comparability of the σ° sequence.

[0171] (2) Topographic radiation correction

[0172] By introducing the S3 fused DEM, the area integral approach is used to remove the terrain modulation caused by local incident angle differences, and the terrain normalized backscattering coefficient σ°_norm is obtained.

[0173] (3) Calculation of change

[0174] Calculate the pixel-by-pixel difference between any two σ°_norm graticets:

[0175] Δσ°=σ°_norm(t2)-σ°_norm(t1)

[0176] Unit: dB. If Δσ°>2dB (empirical threshold, which can be calibrated by region), it is judged as "significantly enhanced", indicating a sharp increase in surface or 0–5cm shallow water content, and is stored in the multidimensional characteristic band as a precursor indicator of debris flow source saturation.

[0177] Topographic features:

[0178] Surface slope and aspect: These are directly calculated from DEM data through spatial analysis (such as neighborhood algorithms). Slope is the core static factor controlling slope stability. The specific calculation process is as follows:

[0179] Using the DEM output by S3 as input, the first-order partial derivatives of the center pixel elevation in the x and y directions are calculated using the 3×3 neighborhood algorithm. slope: Unit: °;

[0180] Slope aspect: The results are represented as 0° due north and 0–360° clockwise. A 10m resolution slope and aspect grid is generated to characterize the static control effect of slope geometry on stability.

[0181] Topographic curvature: Calculated from the DEM, including planar curvature and profile curvature, used to identify catchment areas, ridges, or depressions in the terrain. These areas are sensitive to material accumulation and stability changes. Specific calculation process:

[0182] Calculating the second-order partial derivative based on the slope, we obtain:

[0183] Plan curvature – the curvature of contour lines perpendicular to the slope direction;

[0184] Profile curvature – the rate of change of slope shape along the slope direction;

[0185] Both curvature units are 1 / m, with positive values ​​indicating convexity and negative values ​​indicating depression. Sensitive geomorphic units such as drainage valleys, ridges, and steep slopes can be quickly identified by threshold segmentation (|Curvature|>0.01).

[0186] Topographic Moisture Index (TWI): Used to simulate the distribution trend of soil moisture under topographic control. Specific calculation process:

[0187] (1) Catchment area α: The upstream contribution area was calculated pixel by pixel using the D8 unidirectional flow algorithm (unit: m²). 2 );

[0188] (2) Slope β: Convert the above Slope value into radians;

[0189] (3) TWI = ln(α / tanβ), dimensionless.

[0190] The results are stored at a 10m resolution. Areas with high TWI values ​​correspond to potential soil moisture accumulation zones and can be used as static indicators for evaluating the susceptibility of shallow landslides.

[0191] Hydrological and mechanical characteristics:

[0192] Hourly Precipitation and Cumulative Precipitation: These are the main external driving forces for disaster triggering, obtained through S2.4.1.

[0193] Rate of VWC Change: Derived from a ground-based sensor network. This measure monitors the rate of change in shallow soil moisture content in real time and is calculated using S2.4.2. It reflects the real-time effect of rainfall infiltration and is a key dynamic indicator for determining whether the soil is approaching saturation.

[0194] Pore ​​water pressure: Calculated from a pore water pressure sensor, S2.4.3. An increase in pore water pressure effectively reduces the effective stress on the slip surface and is a direct mechanical indicator triggering instability.

[0195] S4.2 Data Fusion

[0196] To transform heterogeneous features extracted from multiple platforms ("air-space-ground") into structured data that can be directly input into deep learning models, this invention performs feature-level fusion ("pixel-alignment, band-stacking") on the following four categories and a total of 16 feature parameters in stage S4.1, resulting in four multi-dimensional features:

[0197] Deformation fusion features

[0198] The four bands ①InSAR deformation rate, ②InSAR deformation acceleration, ③GNSS deformation rate, and ④GNSS deformation acceleration were resampled to 10m using the "nearest neighbor" method and stacked pixel by pixel to form a 4-band DFL. The units were unified as mm / yr (or mm / h) and mm / yr. 2 (or mm / h) 2 ), used to extract deformation feature vectors in one go.

[0199] Vegetation and ground cover integration characteristics

[0200] The two bands ⑤ΔNDVI and ⑥Δσ° are directly spliced ​​together with a resolution of 10m and the band order is fixed as "NDVI first, σ° second" to generate a 2-band VCFL for use in the vegetation-humidity change sub-vector.

[0201] Terrain fusion features

[0202] Consider the following parameters: ⑦ slope, ⑧ aspect, ⑨ plane curvature, and ⑩ profile curvature. The five TWI bands are stacked in a fixed order (Slope|Aspect|PlanCur|ProfileCur|TWI) with a resolution of 10m to generate a 5-band TFL, enabling one-time reading of terrain static factors.

[0203] Features of the integration of hydrology and mechanics

[0204] Will Hourly rainfall Cumulative rainfall Soil moisture content change rate The four pore water pressure bands are stacked in a "time-mechanical" order, with units of mm, mm, and m respectively. 3 / m 3 / h, kPa, generate 4-band HMFL for hydro-mechanical subvector extraction.

[0205] S5. Feature Construction: Construct multi-dimensional feature vectors and generate multi-dimensional feature images.

[0206] S5.1 Feature Vector Construction: Each of the four multi-dimensional features extracted in S4 (such as deformation fusion feature, vegetation and land cover fusion feature, topography fusion feature, hydrology and mechanics fusion feature, etc.) is considered as an independent dimension. For each pixel (or each regular grid) in the target area, the values ​​of all its corresponding feature parameters are combined in a fixed order to form a high-dimensional feature vector, i.e.: V = [Deformation_Rate, ΔNDVI, Slope, Hourly_Precipitation, Rate_of_VWC_Change, ...]. This feature vector comprehensively describes the current geological hazard risk status of the pixel's location from multiple dimensions such as deformation, vegetation, topography, hydrology, and mechanics.

[0207] S5.2 Multidimensional Image Storage: The feature vectors of all pixels are spatially reorganized and stored using a multidimensional raster or data cube format. Storage Structure: This multidimensional image maintains consistency with the original image in the horizontal space (X, Y dimensions). In the vertical direction (Z dimension or band dimension), each "layer" stores a feature parameter (e.g., the first layer is deformation rate, the second layer is slope, the third layer is hourly rainfall, etc.). This storage method greatly facilitates the subsequent reading, management, and analysis of large batches of multidimensional data by machine learning models.

[0208] S6, Model Building

[0209] By utilizing deep learning models and constructing multi-dimensional feature vectors based on S5, a risk assessment model is built for intelligent prediction, dynamically quantifying the risk probability of landslides and their associated debris flows, and generating tiered early warning information accordingly. This step represents a leap from "data perception" to "intelligent decision-making," completely changing the traditional early warning model that relies on static thresholds.

[0210] S6.1 Construction and Training of Intelligent Risk Assessment Model

[0211] This step is performed offline and aims to build a high-precision risk assessment model.

[0212] S6.1.1 Training Dataset Construction: Collect a multi-dimensional feature vector dataset from historical periods (as input features) and its corresponding, field-verified disaster occurrence labels (as supervision signals, e.g., 1 indicates a landslide / debris flow occurred, 0 indicates no landslide occurred). Clean, augment, and partition the dataset to form a training set, a validation set, and a test set.

[0213] S6.1.2. Construction and Training of Deep Convolutional Neural Network (CNN) Model (for Landslide Spatial Probability Prediction):

[0214] (1) Network Architecture and Initialization

[0215] Adopt the U-Net encoder-decoder structure:

[0216] Encoder: Initialize with the pre-trained weights of ResNet-34, remove the top fully connected layer, and retain the feature maps of 1 / 2, 1 / 4, 1 / 8, and 1 / 16;

[0217] Decoder: 2× transposed convolution + skip connection of the same-scale encoded features at each level, and the number of channels is 512 → 256 → 128 → 64 → 32 in sequence;

[0218] Output layer: Reduce the dimension to 1 channel through 1×1 convolution, activate through Sigmoid, and output a landslide probability map of 0–1.

[0219] (2) Loss Function and Optimization Strategy

[0220] Joint loss L = λ1L_Focal + λ2L_Dice, λ1 = 1.0, λ2 = 0.5;

[0221] Optimizer: AdamW, initial learning rate 1×10 -3 , weight decay 1×10 -4 ;

[0222] Scheduler: Cosine annealing, T_max = 50 epochs, minimum lr = 1×10 -5 ;

[0223] Early stopping strategy: Terminate if the F1-score on the validation set has not improved for 10 consecutive epochs, and save the best model.

[0224] (3) Data Augmentation and Regularization

[0225] Randomly rotate 90° online, horizontally flip, adjust brightness by ±10%, add Gaussian noise with σ = 0.01 to prevent overfitting; use DropBlock (keep_prob = 0.8) at the end of the decoder to enhance spatial regularization.

[0226] (4) Model Validation and Output

[0227] Calculate Precision, Recall, F1, and AUC for each epoch; if AUC ≥ 0.92 and F1 ≥ 0.85 on the test set after training, it is considered qualified; enable the model to learn to identify typical patterns of landslide gestation from complex multi-source features and generate a landslide spatial probability map in real time.

[0228] S6.1.3. Training of Long Short-Term Memory (LSTM) Model (for Sequential Probability Prediction of Chain Disasters)

[0229] (1) Network Architecture and Hyperparameters

[0230] Adopt 2-layer stacked LSTM with a hidden layer dimension of 128 and dropout = 0.2;

[0231] Input shape: (batch, 72, 15);

[0232] Output layer: Sigmoid unit, outputting the "probability of chain-generated debris flow within 6 hours" ranging from 0 to 1.

[0233] (2) Loss Function and Optimization

[0234] Use Binary Cross-Entropy + class weights (positive sample weight = number of negative samples / number of positive samples ≈ 10);

[0235] Optimizer: Adam, lr = 1×10 -3 with weight decay of 1×10 -4 ;

[0236] Scheduler: ReduceLROnPlateau, patience = 5 epochs, factor = 0.5, minimum lr = 1×10 -5 ;

[0237] Early stopping: Terminate if the AUC of the validation set does not improve for 8 consecutive epochs, and save the optimal weights.

[0238] (3) Data Augmentation and Regularization

[0239] Randomly translate along the time axis by ±2 hours, add Gaussian noise with σ = 0.02, and randomly deactivate 10% of the time steps to improve robustness;

[0240] Adopt Temporal Dropout (rate = 0.1) in the LSTM input layer to suppress overfitting.

[0241] (4) Training Process and Evaluation

[0242] Batch size is 256, maximum number of epochs is 100;

[0243] Calculate AUC, Precision, Recall, and F1 for each epoch;

[0244] If the AUC of the test set ≥ 0.90 and F1 ≥ 0.80, it is considered qualified;

[0245] Finally, the model is enabled to learn to recognize the critical timing law of the initiation, movement, and transformation of substances into debris flows under continuous rainfall and other conditions after a landslide occurs, and can infer online to output the probability of chain-generated debris flows in real time.

[0246] S6.2. In the online deployment stage of dynamic risk assessment based on multi-model fusion, the trained model in S6.1 is used for real-time prediction.

[0247] S6.2.1. Prediction of landslide spatial probability: The newly generated real-time multi-dimensional feature images are input into the trained deep convolutional neural network (CNN) model. The model performs forward inference and generates and outputs the landslide disaster spatial probability map of the study area at the current moment in real time. This map accurately depicts the spatial distribution of risks.

[0248] S6.2.2. Prediction of chain disaster probability: For high-risk landslide areas or key gully units, the time series feature sequences are extracted and input into the trained long short-term memory network (LSTM) model. The model outputs the probability of "landslide-debris flow" chain disasters occurring in this area, realizing the prediction of the disaster evolution process.

[0249] S7. Risk judgment

[0250] The system comprehensively combines the landslide spatial probability (P_landslide) output by the CNN and the probability of chain-generated debris flows (P_debris_flow) output by the LSTM, and dynamically determines the total risk level (Risk_Level) of each area according to the preset comprehensive judgment rules. Example of judgment rules: Extremely high risk level: P_landslide > 0.7 and P_debris_flow > 0.6; High risk level: P_landslide > 0.7; Medium risk level: 0.4 < P_landslide <= 0.7; Low risk level: P_landslide <= 0.4. These rules can be determined by learning and optimizing through historical data according to specific application scenarios.

[0251] S8. Release of early warning information: When the total risk level reaches the "high risk" or "extremely high risk" threshold, early warning information is generated and released.

[0252] S8.1. Early warning generation: When the system determines that the total risk level of any area reaches the "high risk" or "extremely high risk" threshold, the early warning generation mechanism is automatically triggered to generate early warning information. The early warning information is structured data, including at least: early warning level, core disaster-causing factors (such as "rainfall reached 150 mm in the past 24 hours"), risk probability value, geographical scope of the main risk area (GeoJSON vector boundary), recommended measures, and effective period.

[0253] S8.2 Early Warning Issuance: Early warning information is automatically pushed to the geological disaster early warning information platform via the application programming interface (API). The platform disseminates the early warning information to relevant management departments and threatened populations in a timely manner through multiple channels (such as web, mobile app, SMS, and broadcast). The system simultaneously generates a thematic early warning map, intuitively displaying the spatial distribution of risks and providing support for command and decision-making.

[0254] The present invention has been described in detail above with reference to the accompanying drawings and embodiments. However, the present invention is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention. All contents not described in detail in the present invention can be derived from existing technologies.

Claims

1. A method for intelligent dynamic monitoring and early warning of rainfall-induced landslide-debris flow chain disasters, characterized in that, include: S1. Data Acquisition: Acquire satellite data, UAV data, and ground sensor monitoring data to generate multi-source data including optical images, SAR images, UAV images, and ground data; S2. Data Processing: Preprocess the multi-source data to generate optical DOM, geocoded backscattering coefficient map, UAV DOM and DEM, and a set of surface observation data values ​​arranged in chronological order; S3. Formatted Data: Spatiotemporal registration, data formatting, and organization are performed on the multi-source data after data processing to obtain multi-source data with a unified and standardized data format; S4. Feature Extraction: Calculate key feature parameters and perform data fusion on them; S5. Feature Construction: Construct multi-dimensional feature vectors and generate multi-dimensional feature images; S6. Model Building: Using a deep learning model, a risk assessment model is built based on the multi-dimensional feature vector constructed in S5 to perform real-time prediction. S7. Risk Assessment: Based on real-time forecast results and preset comprehensive assessment rules, dynamically determine the total risk level of each region. S8. Warning Information Release: When the overall risk level reaches the "high risk" or "extremely high risk" threshold, a warning information is generated and a warning is released.

2. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 1, characterized in that, S2 includes: S2.1 Optical image preprocessing to generate an optical orthophoto with accurate planar position; Radiometric calibration: converting the raw digital quantization values ​​of high-resolution optical images into top-atmosphere radiance values ​​or surface reflectance; Atmospheric correction: converting radiance values ​​or surface reflectance into true surface reflectance; Orthorectification: Orthorectifying the true surface reflectance to generate an optical orthophoto with accurate planar position; S2.2 SAR image preprocessing to generate geocoded backscattering coefficient map; Radiometric calibration: converting the intensity information of single-look complex data from SAR satellite imagery into backscattering coefficients; Speckle noise filtering: An adaptive filtering algorithm is used to process SAR images to suppress their inherent speckle noise while preserving edge and texture details to the greatest extent possible. Geocoding: Based on satellite precise orbit data and DEM data, SAR images are accurately transformed from slant range projection coordinate system to ground range projection coordinate system through range-Doppler model or rational polynomial coefficient model, eliminating terrain distortion and generating geocoded backscattering coefficient map. S2.

3. UAV image preprocessing to generate UAV digital orthophoto maps and digital elevation models; Aerial triangulation and dense matching: Aerial triangulation is performed on the sequence of images acquired by UAVs to solve the high-precision exterior orientation elements of each image, and a high-density point cloud is generated through a dense matching algorithm; Generate digital products: Based on point cloud data, generate high-resolution UAV digital orthophoto maps and digital elevation models; S2.4 Ground data preprocessing: The measurements from rain gauges, soil moisture meters, pore water pressure sensors, and GNSS displacement meters are used to generate a set of surface observation data values ​​arranged in chronological order. S2.4.1 Rain gauge measurement and data preprocessing to generate standardized hourly rainfall time series; S2.4.2 Soil moisture meter measurement, data preprocessing, and generation of standardized soil moisture content and change rate time series; S2.4.3 pore water pressure sensor measurement, data preprocessing, and generation of standardized pore water pressure and rate of change time series; S2.4.4 GNSS displacement gauge measurement, data preprocessing, and generation of standardized displacement rate and acceleration time series; Finally, a data set was formed with 1-hour intervals for hourly rainfall, cumulative rainfall, soil moisture content, rate of change of moisture content, deformation rate, deformation acceleration, hourly pore water pressure, and rate of change of pore water pressure.

3. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 2, characterized in that, S3 includes: S3.1 Multi-source data spatiotemporal registration: unify the multi-source data after S2 data processing to the same spatiotemporal reference; S3.2 Data Format and Organization: The multi-source data processed in S3.1 is converted into a unified and standardized data format, and organized and managed using a spatiotemporal grid data model to obtain multi-source data with a unified and standardized data format.

4. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 3, characterized in that, S3.1 includes: S3.1.1 Unification of Spatial Reference Standards: Coordinate System 1: Convert all data processed from S2 data from plane coordinate system 1 to geodetic coordinate system, and unify the elevation system to geodetic height system; Pixel resampling: Resample the raster data in all data after S2 data processing to the same spatial resolution; Precise registration: Using a high-precision image that has undergone orthorectification as the reference image, the affine transformation or polynomial transformation model between other images and this reference image is calculated through feature point matching or region matching algorithms to achieve sub-pixel level geometric precision registration. S3.1.2, Unified Time Base: The timestamps of all data collected after S2 data processing will be uniformly converted to Coordinated Universal Time. For data acquired at different times, the time difference should be recorded and considered during time series analysis.

5. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 4, characterized in that, S4 includes: S4.1 Feature extraction: Extract feature parameters such as deformation rate, NDVI change, surface slope, hourly rainfall, and soil moisture content change rate to form multi-dimensional feature parameters. S4.2 Feature fusion is used to transform heterogeneous features extracted from multiple platforms ("air-space-ground") into structured data that can be directly input into deep learning models.

6. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 5, characterized in that, The key feature parameters in S4 include: deformation features, vegetation and land cover features, topographic features, and hydrological and mechanical features; deformation features include InSAR deformation rate, InSAR deformation acceleration, GNSS deformation rate, and GNSS deformation acceleration; vegetation and land cover features include NDVI variation and backscattering coefficient variation; topographic features include land slope, aspect, topographic curvature, and topographic humidity index; hydrological and mechanical features include hourly rainfall, cumulative rainfall, soil volumetric water content change rate, and pore water pressure; feature-level fusion of the above key feature parameters is performed using "pixel-alignment and band-stacking".

7. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 6, characterized in that, S5 includes: S5.1 Feature Vector Construction: Each of the four multi-dimensional features extracted in S4 is regarded as an independent dimension; for each cell or each regular grid in the target area, the values ​​of all its corresponding feature parameters are combined in a fixed order to form a high-dimensional feature vector. S5.2 Multidimensional Image Storage: Reorganize the feature vectors of all pixels in space and store them in a multidimensional image or data cube format.

8. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 7, characterized in that, S6 includes: S6.1 Construction and Training of Risk Assessment Model S6.2, the online deployment phase of dynamic risk assessment based on multi-model fusion, utilizes the model trained in S6.1 for real-time prediction.

9. A method for intelligent dynamic monitoring and early warning of rainfall-induced landslide-debris flow chain disasters according to claim 8, characterized in that, S6.1 includes: S6.1.1 Training Dataset Construction: Collect multi-dimensional feature vector datasets from historical periods and their corresponding disaster occurrence labels that have been verified in the field. Clean, enhance, and divide the datasets to form training sets, validation sets, and test sets. S6.1.2 Training a deep convolutional neural network model to extract spatial context information and local features from multi-dimensional feature images and output a spatial probability map of landslide occurrence; S6.1.3 Training of Long Short-Term Memory Network Model: This is used to process time-series data, capture the long-term dependencies of data in the time dimension, and output the probability of a landslide turning into a debris flow within a specific time window in the future. S6.2 includes: S6.2.1 Landslide Spatial Probability Prediction: The latest multi-dimensional feature image generated in real time is input into the trained deep convolutional neural network model to generate and output the spatial probability map of landslide disaster in the study area at the current moment. S6.2.2 Chain Disaster Probability Prediction: For high-risk landslide areas or key valley units, extract their temporal feature sequences, input them into a pre-trained long short-term memory network model, and output the probability of a "landslide-debris flow" chain disaster occurring in the area.

10. The intelligent dynamic monitoring and early warning method for rainfall-induced landslide-debris flow chain disasters according to claim 9, characterized in that, S8 includes: S8.1 Warning Generation: When the system determines that the total risk level of any area reaches the "high risk" or "extremely high risk" threshold, the warning generation mechanism is automatically triggered to generate warning information. S8.2 Warning Release: Warning information is automatically pushed to the geological disaster warning information platform through the application programming interface. The platform releases the warning information to relevant management departments and threatened people through multiple channels as soon as possible.

Images