Landslide prediction and early warning system and method based on remote sensing technology

Abstract

The invention discloses a landslide prediction and early warning system and method based on a remote sensing technology, and is applied to the technical field of landslide early warning. Comprising a data acquisition module for acquiring multi-source data in a monitoring area; the cloud computing module is used for large-range landslide risk assessment and deformation monitoring; the edge calculation module is used for landslide risk assessment and deformation monitoring of a specific area; the end-cloud cooperative communication module is used for realizing data interaction between the cloud computing module and the edge computing module; the federal learning module is used for realizing cooperative training under data privacy protection of different areas; and the landslide early warning module is used for carrying out early warning based on landslide risk assessment and deformation monitoring results of the cloud computing module and the edge computing module. According to the invention, by using the deep learning model, the end-cloud collaborative architecture and the federal learning driven data fusion framework, efficient, low-cost and high-precision landslide risk prediction is realized.

Description

Technical Field

[0001] This invention relates to the field of landslide early warning technology, and more specifically to a landslide prediction and early warning system and method based on remote sensing technology. Background Technology

[0002] Monitoring and early warning are crucial measures for mitigating geological disaster risks. Existing geological disaster early warning schemes can be categorized into three main types based on monitoring methods: manual, simplified, and professional monitoring. Manual monitoring is a core component of community-based monitoring and prevention, characterized by its strong intuitiveness, adaptability, and high reliability. Manual monitoring involves human intervention at geological disaster sites, such as landslide monitoring, regularly observing the occurrence and development of cracks in landslide bodies, ground subsidence, expansion, uplift, collapse, and building deformation, and investigating abnormal phenomena such as ground sounds and groundwater anomalies. Simplified monitoring, with human assistance, uses basic instruments for geological disaster monitoring and early warning. This method often utilizes simple monitoring instruments and has achieved good results in geological disaster monitoring. Professional monitoring methods utilize instruments and equipment to achieve remote automatic telemetry and data collection. However, most existing monitoring schemes are applicable to geological disaster risk assessment at single work sites or localized areas, making it difficult to achieve large-scale monitoring of geological disasters along highways. Highways are linear transportation projects with long routes, traversing different geological and geomorphological units, and complex geological environments, requiring large-scale monitoring technologies. Therefore, how to provide a landslide prediction and early warning system and method based on remote sensing technology is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0003] In view of this, the present invention provides a landslide prediction and early warning system and method based on remote sensing technology, which achieves efficient, low-cost and high-precision landslide risk prediction by using deep learning models, edge-cloud collaborative architecture and federated learning-driven data fusion framework.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A landslide prediction and early warning system based on remote sensing technology includes:

[0006] The data acquisition module acquires multi-source data within the monitoring area;

[0007] The cloud computing module is used for large-scale landslide risk assessment and deformation monitoring. It is equipped with the induced feature threshold model STH-M, the low-resolution fusion assessment model LSA-M1, the medium-to-high resolution regional landslide risk assessment model LSA-M2, the deformation threshold model DISP-M, and the landslide identification model SEG-M.

[0008] The edge computing module is used for landslide risk assessment and deformation monitoring in specific areas, and is equipped with the lightweight landslide risk assessment model LSA-M2D and the lightweight landslide identification model SEG-MD.

[0009] The edge-cloud collaborative communication module enables data interaction between the cloud computing module and the edge computing module.

[0010] The federated learning module enables collaborative training under the protection of data privacy in different regions;

[0011] The landslide early warning module provides early warnings based on landslide risk assessment and deformation monitoring results from cloud computing and edge computing modules.

[0012] Optionally, the cloud computing module assesses landslide risk based on the Induced Feature Threshold Model (STH-M), the Low-Resolution Fusion Assessment Model (LSA-M1), and the Medium-to-High Resolution Regional Landslide Risk Assessment Model (LSA-M2), specifically:

[0013] The STH-M induced feature threshold model rapidly generates a large-scale landslide risk classification map R1 based on induced feature risk thresholds, meteorological forecast data, and satellite remote sensing data.

[0014] The low-resolution fusion assessment model LSA-M1 performs a fusion assessment of regional landslide risk based on the large-scale landslide risk classification map R1, and obtains the regional landslide risk assessment result R2.

[0015] The medium-to-high resolution regional landslide risk assessment model LSA-M2 integrates the regional landslide risk assessment results R2 to obtain the detailed assessment results R3 of small-scale landslide risk.

[0016] Optionally, the input data for the STH-M induced feature threshold model is the induced feature vector. A landslide risk classification map R1 is generated by weighting the induced feature vector with risk feature thresholds based on dynamic statistics of historical landslide events. The induced feature vector includes three remote sensing derived parameters: meteorological, geological, topographical, and soil environmental parameters. The geospatial weight matrix W... g With geological condition weight matrix W s Dynamically adjust risk characteristic thresholds to achieve regional differentiation adaptation;

[0017] The specific steps for constructing the induced feature threshold model STH-M are as follows:

[0018] Construction of a historical landslide event database: Collecting historical landslide event data and correlating it with multi-source remote sensing data at the time of the events;

[0019] Extract the induced feature vector for each event point

[0020]

[0021] Where, f rain For rainfall characteristics, f soil For soil moisture characteristics, f terrain For terrain features, f ndvi f represents the vegetation cover index characteristic. lith Geological lithology classification characteristics;

[0022] Risk characteristic threshold quantification:

[0023] Based on statistical analysis, the critical threshold ranges for each induced characteristic were determined: Features exceeding the threshold are marked as high-risk.

[0024] Introducing geospatial heterogeneity weights ω g and geological condition weight ω s Generate a differential weight matrix W = ω g ×ω s ;

[0025] Weighted risk fusion model:

[0026] Establish risk index R index Calculation formula:

[0027]

[0028] In the formula, The midpoint of the threshold interval;

[0029] According to the risk index R index Risk is divided into four levels: 0-0.3 is low risk, 0.3-0.6 is medium risk, 0.6-0.9 is high risk, and greater than 0.9 is extremely high risk.

[0030] Optionally, the low-resolution fusion evaluation model LSA-M1 is as follows:

[0031] Dual-channel input architecture:

[0032] Dynamic remote sensing channel: Input NDVI, surface temperature and daily cumulative precipitation (K), and unify them to 1 km resolution through spatiotemporal alignment;

[0033] Static geological channel: Input slope, lithology classification and geological fracture density, and normalize and fuse them through feature importance weight matrix;

[0034] Intelligent data imputation unit: For missing areas caused by cloud cover, a complete data field is generated using historical time series similarity matching and regression imputation algorithms;

[0035] Risk assessment dynamic triggering mechanism:

[0036] In the large-scale landslide risk classification map R1, low-risk areas undergo regular assessments; high-risk areas are automatically triggered for detailed assessment using the medium-to-high resolution model LSA-M2.

[0037] Optional, the medium-to-high resolution regional landslide risk assessment model LSA-M2 includes:

[0038] Spatiotemporal feature decoupling unit:

[0039] Time-series sensing head: Analyzes periodic patterns in meteorological data and extracts the evolution characteristics of inducing factors;

[0040] Spatial correlation head: captures the spatial coupling relationship of terrain features and outputs an attention weight matrix to identify key inducing factors;

[0041] Geological deviation correction unit:

[0042] Dynamic correction coefficients are generated based on static geological data, and systematic biases in weather forecasts are calibrated through residual connections. The calculation process is as follows:

[0043]

[0044] In the formula, This is the corrected predicted value; Y base The baseline model prediction; G static This is a static geological feature vector; Δ g eo represents the geological deviation compensation amount; U represents the learnable parameter matrix; σ represents the Sigmoid activation function; This is element-wise multiplication;

[0045] Resolution adaptive mechanism:

[0046] By default, risk assessments are performed at a spatial resolution of 10 meters, which can be dynamically switched to a range of 5 to 30 meters based on data availability.

[0047] The parameter matrix U is generated through the following steps: constructing a historical landslide event-geological feature mapping table; using ridge regression to solve for the coefficient matrix that minimizes the prediction bias; and aggregating parameters from multiple regions through federated learning to achieve generalization enhancement.

[0048] The temporal sensing head in the spatiotemporal feature decoupling unit supports sliding time window analysis, with the window length dynamically adapting to the monsoon / dry season cycle; the spatial correlation head fuses InSAR coherence maps to improve the accuracy of active fault zone identification.

[0049] Geological deviation compensation amount Δ geo The calculation satisfies:

[0050] Δ geo=β·log(1+CrackDensity)×RockSoftnessIndex

[0051] In the formula, β is the regional calibration coefficient, CrackDensity is the crack density, RockSoftnessIndex is the rock softening index, and the remote sensing calculation steps for RockSoftnessIndex are as follows:

[0052] Clay mineral content index (CMI) was calculated using near-infrared (SWIR) band data from multispectral remote sensing data.

[0053]

[0054] In the formula, ρ is the surface reflectance, and SWIR1 and SWIR2 are adjacent near-infrared bands, respectively.

[0055] Elastic modulus ratio estimation:

[0056] Acquire synthetic aperture radar (SAR) images and extract the surface deformation rate V using temporal interferometry (InSAR). def Combined with meteorological precipitation data V rain Calculate the modulus ratio:

[0057]

[0058] In the formula, k1 is the regional calibration coefficient;

[0059] Rock softening index generation:

[0060] Normalize the CMI to the 0-1 range to obtain the CMI. norm The final index is calculated as follows:

[0061]

[0062] Where, E sat E represents the elastic modulus of the rock mass under saturated conditions. dry This represents the elastic modulus of the rock mass under dry conditions.

[0063] Optionally, the cloud computing module utilizes the deformation threshold model DISP-M and the landslide identification model SEG-M to achieve refined monitoring and verification of high-risk areas, specifically including:

[0064] Deformation monitoring and dynamic weight calibration:

[0065] The deformation threshold model DISP-M extracts areas with high risk for three consecutive days from the regional landslide risk assessment result R2, generates a boundary mask MASK-1, collects Sentinel-1InS AR data of the MASK-1 area and performs preprocessing, including deformation index quantification and dynamic weight calibration.

[0066] Dynamic weight calibration is based on an InSAR deformation feature library of historical landslide events. A transfer learning algorithm is used to adaptively adjust the weights of surface deformation sensitivity factors, where the weight coefficient W... dynamic (t) is dynamically updated according to seasonal rainfall changes;

[0067]

[0068] In the formula, ΔR(t) is the rate of change of rainfall, α1 and β are the federated learning adaptive parameters, and C hist The contribution value of the sensitive factor to the InSAR deformation feature library in historical landslide events;

[0069] Output surface deformation risk distribution map R4 based on preprocessed Sentinel-1 InSAR data;

[0070] Landslide boundary identification constrained by topographic gradient:

[0071] The landslide identification model SEG-M is based on the Swin-Transformer architecture and introduces terrain gradient constraints through a spatial attention mechanism, specifically:

[0072]

[0073] In the formula, Q i For the query vector, K j Let d be the key vector. k Represents the vector dimension. The elevation gradient is given by α2 = 0.3, which is the adjustment coefficient.

[0074] Optionally, the landslide recognition model SEG-M is a landslide recognition model based on the Swin-Transformer architecture. By default, it uses Sentinel-2 multi-band data and STRM topographic data to identify and label landslide areas in images. When Sentinel-2 data for a region is unavailable, it uses 30-meter-level Landsat images and super-resolution reconstruction technology to fill in the gaps by default. In addition, adversarial examples such as cloud and fog obscuration and shadow noise are added during the training phase to ensure the model's accuracy under adverse weather conditions.

[0075] Optionally, the edge-cloud collaborative communication module enables data interaction between the cloud computing module and the edge computing module as follows:

[0076] The cloud computing module utilizes GPU clusters for parallel computing, dividing tasks by geographical blocks; based on the idle rate of computing nodes, high-risk areas are prioritized for scheduling to high-performance GPU nodes; static data is stored in a distributed manner, cached by region to reduce I / O latency; dynamic data is managed through a time-series database, supporting rapid retrieval and updates; multi-source data is uniformly converted to UTC timestamps and WGS84 coordinates through an ETL pipeline; spatiotemporal alignment errors are controlled within 10 meters and 1 hour; high-risk areas trigger cascaded computation of the medium-to-high resolution regional landslide risk assessment model LSA-M2 and the landslide identification model SEG-M.

[0077] The edge computing module uses distillation technology to transform the medium-to-high resolution regional landslide risk assessment model LSA-M2 into the lightweight landslide risk assessment model LSA-M2D, retaining the core parameters of the residual correction module. It also rectifies the landslide identification model SEG-M into the lightweight landslide identification model SEG-MD, removing redundant attention heads while retaining multi-scale skip connections. When memory is less than 4GB, the residual correction module of the lightweight landslide risk assessment model LSA-M2D is disabled; when computation latency is greater than 5 seconds, the input data is dynamically downsampled to a 30-meter resolution to ensure real-time response. Static data is pre-stored on edge devices and updated weekly via cloud synchronization; dynamic data is acquired from local weather stations via the LoRaWAN protocol.

[0078] The edge computing module publishes the coordinates of high-risk areas via the MQTT protocol. Upon receiving an MQTT request, the cloud computing module prioritizes scheduling idle nodes for detection and transmits the data in chunks via HTTP / 2 streaming. The transmission priority is: deformation early warning data > risk distribution map > historical data.

[0079] When the network is down, the edge computing module caches computing resources locally for 72 hours, supporting fully offline operation; after the network is restored, the cloud load is reduced through differential synchronization.

[0080] Optionally, the federated learning module enables collaborative training under data privacy protection in different regions as follows:

[0081] The edge computing module server uses Paillier homomorphic encryption for local gradients, and the ciphertext is directly aggregated in the cloud to prevent the leakage of the original data. The encrypted gradient is transmitted through threshold signature to prevent man-in-the-middle attacks.

[0082] Weights are dynamically allocated based on data quality scores:

[0083]

[0084] Q i =αRecall i +βPrecision i +γCompleteness i

[0085] In the formula, w i As the weight, Q i For quality scoring, α, β, and γ are weighting coefficients, which are dynamically adjusted according to the regional landslide monitoring needs. i Precision is the key factor for data recall. i For data accuracy, Completeness i For data integrity, Q i Low quality region (Q < 0.5) i <0.5) The weight is forcibly decayed to below 0.1 to suppress noise interference;

[0086] Cross-region feature alignment: Before federated aggregation, maximum mean difference regularization is applied to the parameters of each edge model to minimize the regional feature distribution differences;

[0087] Elastic weight solidification: Shared layer parameters are solidified to retain global feature extraction capabilities; Local adaptation layers allow for edge fine-tuning, adapting to specific regions.

[0088] Training process closed loop: After each round of federated training, the cloud computing module distributes the global model to the edge computing module and updates the local distillation model synchronously; the edge computing module detects the local data distribution shift through adversarial verification and dynamically adjusts the training sampling weights.

[0089] Optionally, it also includes an iterative optimization module, which achieves self-iterative optimization of the system based on a dynamic optimization mechanism using a supervised sample library and incremental learning and catastrophic forgetting suppression.

[0090] Dynamic optimization mechanism of semi-supervised sample libraries:

[0091] Automatic calibration of classification threshold C-TH: from the prediction sample library PL i The 10% samples with the highest uncertainty are selected, manually labeled, and added to the training set. The remaining 90% of samples are then reclassified using the updated threshold, and another 10% are selected for manual review. If the review recall rate R... 10p If the value is ≥0.95, the remaining samples are automatically labeled; otherwise, iteration continues. After each iteration, based on the recall and precision curves of the labeled samples, a binary search method is used to determine the minimum threshold C-THmin, ensuring R0. 10p ≥0.95; when R is used for two consecutive rounds 10p When the value is ≥0.95, the remaining unlabeled samples are automatically assigned pseudo-labels and added to the training set by combining the terrain mutation rate and deformation rate.

[0092] During the iteration process, 10% of the samples are labeled in the first round, 9% remain in the second round, and so on, until the total amount of manual labeling is ≤5%. Samples with an automatic label confidence score <0.7 are marked as "to be reviewed" and are randomly selected by experts for inspection. When the inspection error is >5%, the model is fully retrained.

[0093] Incremental learning and catastrophic forgetting inhibition:

[0094] The importance of model parameters is calculated using the Fisher information matrix, with the following formula:

[0095]

[0096] In the formula, F i Here is the Fisher matrix; θ i Let be the i-th parameter of the model, representing the historically optimal parameter; This is an old task dataset; For the expected value of the data, represent the expectation of the data. The average of all samples in the sample;

[0097] With the shared layer parameters fixed, only the local adaptation layer is fine-tuned; the loss function is:

[0098]

[0099] in, λ is the loss function; λ is the regularization coefficient, used to balance the learning weights of new and old tasks, with a value of 0.5. The optimal parameter values ​​after training the old task; F i Fisher information for parameters, reflecting their importance;

[0100] Adversarial example generation: Perturbations are added to the input data using the fast gradient sign method to simulate distribution shift;

[0101]

[0102] Where ∈ represents the perturbation intensity, controlling the adversarial sample x. 对抗 The offset range is 0.05. This represents the gradient of the loss function with respect to the input;

[0103] Distribution offset detection:

[0104]

[0105] Among them, D KL P represents the Kullback-Leibler divergence, quantifying the difference between the old and new distributions. 新 P represents the distribution of the new data; 旧This represents the distribution of old data. When the distribution shifts, the historical sample sampling weight is increased by 1.5. Regions with severe shifts trigger full model retraining.

[0106] A landslide prediction and early warning method based on remote sensing technology includes the following steps:

[0107] S1. Using historical data and the STH-M induced characteristic threshold model, a rapid classification of landslide risk over a large area is completed, and the classification result R1 is obtained.

[0108] S2. The regional landslide risk is assessed by combining the graded result R1 and the pre-trained medium- and low-resolution fusion assessment model LSA-M1, resulting in the assessment result R2.

[0109] S3, the medium-to-high resolution regional landslide risk assessment model LSA-M2 is based on the regional landslide risk assessment result R2 to perform a fusion assessment of regional landslide risk and obtain the fine assessment result R3 of small-scale landslide risk;

[0110] S4. The deformation threshold model DISP-M extracts areas with high risk for three consecutive days from the regional landslide risk assessment result R2, generates a boundary mask MASK-1, collects Sentinel-1 InSAR data of the MASK-1 area and preprocesses it, and makes an early warning based on the preprocessed Sentinel-1 InSAR data, outputting a deformation risk distribution map R4.

[0111] S5. The landslide identification model SEG-M extracts all landslide boundaries within the region based on the variable risk distribution map R4, and obtains the landslide boundary extraction result R5.

[0112] S6. Using the regional landslide risk assessment result R2 and the landslide identification model SEG-M, the landslide boundary is extracted again to obtain the landslide boundary extraction result R6. R6 is used as a supplement to R5. Duplicate areas are removed, and the result of R5 is used first for conflict areas to generate the final landslide boundary map.

[0113] As can be seen from the above technical solution, compared with the prior art, the present invention provides a landslide prediction and early warning system and method based on remote sensing technology, which has the following beneficial effects:

[0114] 1. This invention captures the synergistic effect of precipitation periodicity and topography. The residual correction introduces static geological parameters (lithology, slope) to generate dynamic correction coefficients, calibrates the systematic bias of weather forecasts, and makes up for the shortcomings of traditional models in not making full use of static geological features. It can achieve efficient synergy between large-scale and refined risk assessment.

[0115] 2. This invention deploys the full model in the cloud computing module to complete complex calculations, while the edge computing module runs a lightweight model for real-time preprocessing and preliminary analysis, dynamically allocating computing tasks to ensure efficient and stable system operation. The edge computing module applies distillation to compress model parameters and shuts down non-critical modules (such as SWIR band processing) when memory is insufficient, prioritizing real-time response at the cost of accuracy. Tasks are divided by geographical blocks, with high-risk areas prioritized for scheduling to high-performance GPU nodes to improve computing efficiency.

[0116] 3. This invention achieves collaborative training under multi-regional data privacy protection through federated learning, and combines dynamic weight allocation with MMD feature alignment to resolve the contradiction between regional specificity and generalization ability. Attached Figure Description

[0117] To more clearly illustrate the technical solutions in the embodiments of the present 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0118] Figure 1 This is a schematic diagram of the landslide prediction and early warning system based on remote sensing technology of the present invention.

[0119] Figure 2 This is a flowchart of the landslide prediction and early warning method based on remote sensing technology of the present invention. Detailed Implementation

[0120] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0121] This invention discloses a landslide prediction and early warning system based on remote sensing technology, such as... Figure 1 As shown, it includes:

[0122] The data acquisition module acquires multi-source data within the monitoring area;

[0123] The cloud computing module is used for large-scale landslide risk assessment and deformation monitoring. It is equipped with the induced feature threshold model STH-M, the low-resolution fusion assessment model LSA-M1, the medium-to-high resolution regional landslide risk assessment model LSA-M2, the deformation threshold model DISP-M, and the landslide identification model SEG-M.

[0124] The edge computing module is used for landslide risk assessment and deformation monitoring in specific areas, and is equipped with the lightweight landslide risk assessment model LSA-M2D and the lightweight landslide identification model SEG-MD.

[0125] The edge-cloud collaborative communication module enables data interaction between the cloud computing module and the edge computing module.

[0126] The federated learning module enables collaborative training under the protection of data privacy in different regions;

[0127] The landslide early warning module provides early warnings based on landslide risk assessment and deformation monitoring results from cloud computing and edge computing modules.

[0128] In this embodiment of the invention, the multi-source data includes static geographic data, dynamic remote sensing meteorological data, and dynamic geological parameters; the static geographic data includes water system distance, road network topology data, and crack data; the dynamic remote sensing meteorological data includes rainfall prediction data, cumulative rainfall intensity, daily average temperature, diurnal temperature range, probability of freeze-thaw events, rate of temperature change, maximum wind speed, wind direction, regional wind force level forecast, relative humidity forecast, evaporation estimation, 24-hour air pressure change rate, sudden air pressure change forecast, sunshine duration forecast, cloud cover, historical precipitation threshold, soil moisture content, volumetric water content at a depth of 5 cm below the surface, optical image data, SAR image data, thermal infrared data, displacement monitoring values, groundwater level, and geotechnical parameters; and the dynamic geological parameters include slope, lithology data, and lithology distribution.

[0129] In static geographic data, water system distances are extracted using Geographic Information System (GIS) to determine the Euclidean distances between rivers, lakes, and other water bodies within the monitoring area and the roadside, quantifying the impact of groundwater infiltration on slope stability. Road network topology data is obtained by acquiring the spatial distribution of infrastructure such as road alignment, bridges, and tunnels, and combined with a digital elevation model (DEM) to analyze the sensitivity of the surrounding terrain to landslides. A dynamic crack database is constructed by integrating crack location, length, direction, and evolution time-series data recorded by UAV aerial surveys or ground inspections.

[0130] In the dynamic remote sensing meteorological data, the rainfall forecast data uses hourly rainfall forecasts for the next 1-3 days, the cumulative rainfall intensity is the cumulative rainfall intensity in the typhoon track forecast, the daily average temperature is the daily average temperature for the next 3 days, the diurnal temperature range and the probability of freeze-thaw events are obtained based on the meteorological bureau's numerical model, the temperature change rate is fitted from historical temperature curves, the maximum wind speed and wind direction are obtained based on the meteorological bureau's typhoon forecast system, the regional wind force level forecast is obtained based on the WRF mesoscale meteorological model, the relative humidity forecast is the relative humidity forecast for the next 72 hours, the evaporation is estimated using the Penman-Monteith model, the 24-hour air pressure change rate is the meteorological bureau's real-time monitoring data, the sunshine duration forecast is obtained based on meteorological satellite inversion, the cloud cover rate is Himawari-8 satellite data, the historical precipitation threshold is obtained based on the critical rainfall curves of different geological units statistically analyzed from historical disaster events, the volumetric water content at a depth of 5 cm below the surface is obtained using SMAP satellite data inversion, and the displacement monitoring values, groundwater level and geotechnical parameters are monitored in real time through monitoring stations.

[0131] In dynamic geological parameters, slope and lithology data are updated in real time through UAV aerial surveys, and lithology distribution maps are retrieved through hyperspectral remote sensing.

[0132] Furthermore, the cloud computing module assesses landslide risk based on the Induced Feature Threshold Model (STH-M), the Low-Resolution Fusion Assessment Model (LSA-M1), and the Medium-to-High Resolution Regional Landslide Risk Assessment Model (LSA-M2), specifically as follows:

[0133] The STH-M induced feature threshold model rapidly generates a large-scale landslide risk classification map R1 based on induced feature risk thresholds, meteorological forecast data, and satellite remote sensing data.

[0134] The low-resolution fusion assessment model LSA-M1 performs a fusion assessment of regional landslide risk based on the large-scale landslide risk classification map R1, and obtains the regional landslide risk assessment result R2.

[0135] The medium-to-high resolution regional landslide risk assessment model LSA-M2 integrates the regional landslide risk assessment results R2 to obtain the detailed assessment results R3 of small-scale landslide risk.

[0136] Furthermore, the input data for the induced feature threshold model STH-M is the induced feature vector. By weighting the induced feature vector with risk feature thresholds based on dynamic statistics of historical landslide events, a landslide risk classification map R1 is generated. The induced feature vector includes three remote sensing derived parameters: meteorological, geological and topographical, and soil environmental. The geospatial weight matrix W... g With geological condition weight matrix W s Dynamically adjust risk characteristic thresholds to achieve regional differentiation adaptation;

[0137] The specific steps for constructing the induced feature threshold model STH-M are as follows:

[0138] Construction of a historical landslide event database: Collecting historical landslide event data and correlating it with multi-source remote sensing data at the time of the events;

[0139] Extract the induced feature vector for each event point

[0140]

[0141] Where, f rain For rainfall characteristics, f soil For soil moisture characteristics, f terrain For terrain features, f ndvi f represents the vegetation cover index characteristic. lith Geological lithology classification characteristics;

[0142] Risk characteristic threshold quantification:

[0143] Based on statistical analysis, the critical threshold ranges for each induced characteristic were determined: Features exceeding the threshold are marked as high-risk.

[0144] Introducing geospatial heterogeneity weights ω g and geological condition weight ω s Generate a differential weight matrix W = ω g ×ω s ;

[0145] Weighted risk fusion model:

[0146] Establish risk index R index Calculation formula:

[0147]

[0148] In the formula, The midpoint of the threshold interval;

[0149] According to the risk index R index Risk is divided into four levels: 0-0.3 is low risk, 0.3-0.6 is medium risk, 0.6-0.9 is high risk, and greater than 0.9 is extremely high risk.

[0150] Furthermore, the low-resolution fusion evaluation model LSA-M1 is as follows:

[0151] Dual-channel input architecture:

[0152] Dynamic remote sensing channel: Input NDVI, surface temperature and daily cumulative precipitation (K), and unify them to 1 km resolution through spatiotemporal alignment;

[0153] Static geological channel: Input slope, lithology classification and geological fracture density, and normalize and fuse them through feature importance weight matrix;

[0154] Intelligent data imputation unit: For missing areas caused by cloud cover, a complete data field is generated using historical time series similarity matching and regression imputation algorithms;

[0155] Risk assessment dynamic triggering mechanism:

[0156] In the large-scale landslide risk classification map R1, low-risk areas undergo regular assessments; high-risk areas are automatically triggered for detailed assessment using the medium-to-high resolution model LSA-M2.

[0157] In this embodiment of the invention, the lightweight CatBoost ensemble learning framework has the following advantages: 1. Feature importance driven: By prioritizing key parameters through feature ranking (slope, land surface type, and cumulative precipitation accounting for more than 60% of the weight), the model efficiency is improved; 2. Dynamic update mechanism: The latest satellite data (such as MODIS land surface temperature) is automatically fused every 3 days, and random forest regression is used to fill in data in cloud-covered areas to ensure the timeliness of the evaluation; 3. Adaptive downsampling: For low-risk areas (R1=0), spatial interpolation is used to simplify the calculation and free up edge device resources;

[0158] When conducting a fusion assessment of regional landslide risk, if the edge computing module has insufficient memory, non-critical feature channels (such as wind speed) are shut down to prioritize accuracy over speed. The output result R2 of the edge computing module is synchronized to the cloud computing module to update the global risk map.

[0159] Furthermore, the medium-to-high resolution regional landslide risk assessment model LSA-M2 includes:

[0160] Spatiotemporal feature decoupling unit:

[0161] Time-series sensing head: Analyzes periodic patterns in meteorological data and extracts the evolution characteristics of inducing factors;

[0162] Spatial correlation head: captures the spatial coupling relationship of terrain features and outputs an attention weight matrix to identify key inducing factors;

[0163] Geological deviation correction unit:

[0164] Dynamic correction coefficients are generated based on static geological data, and systematic biases in weather forecasts are calibrated through residual connections. The calculation process is as follows:

[0165]

[0166] In the formula, This is the corrected predicted value; Y baseThe baseline model prediction; G static This is a static geological feature vector; Δ g eo represents the geological deviation compensation amount; U represents the learnable parameter matrix; σ represents the Sigmoid activation function; This is element-wise multiplication;

[0167] Resolution adaptive mechanism:

[0168] By default, risk assessments are performed at a spatial resolution of 10 meters, which can be dynamically switched to a range of 5 to 30 meters based on data availability.

[0169] The parameter matrix U is generated through the following steps: constructing a historical landslide event-geological feature mapping table; using ridge regression to solve for the coefficient matrix that minimizes the prediction bias; and aggregating parameters from multiple regions through federated learning to achieve generalization enhancement.

[0170] The temporal sensing head in the spatiotemporal feature decoupling unit supports sliding time window analysis, with the window length dynamically adapting to the monsoon / dry season cycle; the spatial correlation head fuses InSAR coherence maps to improve the accuracy of active fault zone identification.

[0171] Geological deviation compensation amount Δ geo The calculation satisfies:

[0172] Δ geo =β·log(1+CrackDensity)×RockSoftnessIndex

[0173] In the formula, β is the regional calibration coefficient, CrackDensity is the crack density, RockSoftnessIndex is the rock softening index, and the remote sensing calculation steps for RockSoftnessIndex are as follows:

[0174] Clay mineral content index (CMI) was calculated using near-infrared (SWIR) band data from multispectral remote sensing data.

[0175]

[0176] In the formula, ρ is the surface reflectance, and SWIR1 and SWIR2 are adjacent near-infrared bands, respectively.

[0177] Elastic modulus ratio estimation:

[0178] Acquire synthetic aperture radar (SAR) images and extract the surface deformation rate V using temporal interferometry (InSAR). def Combined with meteorological precipitation data V rain Calculate the modulus ratio:

[0179]

[0180] In the formula, k1 is the regional calibration coefficient;

[0181] Rock softening index generation:

[0182] Normalize the CMI to the 0-1 range to obtain the CMI. norm The final index is calculated as follows:

[0183]

[0184] Where, E sat E represents the elastic modulus of the rock mass under saturated conditions. dry This represents the elastic modulus of the rock mass under dry conditions. The detailed assessment process for the LSA-M2 medium-to-high resolution regional landslide risk assessment model is as follows:

[0185] 1. Data Acquisition and Preprocessing:

[0186] Temporal characteristics:

[0187] Regional surface deformation data: deformation time series covering the 3 months prior to the landslide (Sentinel-1 InSAR data); regional surface type change: dynamic surface type classification over 30 days (bare land S, vegetation V, buildings B); NDVI and drought index: 30-day time series data used for vegetation cover and soil moisture monitoring; meteorological time series data: including rainfall, precipitation intensity, wind speed and direction, and soil moisture, covering 30 days before to 7 days after the landslide (ECMWF forecast data).

[0188] Non-temporal characteristics:

[0189] Geological and topographic data: slope, aspect, and lithology coding (e.g., loose rock areas are coded as 1, and hard rock areas as 2); Remote sensing proximity observation data: Sentinel-2 multispectral bands (shortwave infrared SWIR, etc.) of the adjacent area; Statistical threshold data: dynamic threshold matrix TH~t,d~ (time × space × land surface type).

[0190] Data source and resolution:

[0191] In high-risk areas (R2=1), Sentinel-2 multispectral data (10 meters) were collected first, with Landsat data used as a historical supplement; meteorological data (ECMWF) and topographic data (slope, lithology) were spatiotemporally aligned to a 10-meter resolution.

[0192] 2. Model Inference:

[0193] Time window: Meteorological data covers 30 days before to 7 days after the landslide, and deformation data covers 6 months; Enhanced adversarial training: Adversarial samples such as cloud and fog obstruction and shadow noise are added to improve robustness under severe weather conditions.

[0194] 3. Collaborative Workflow:

[0195] Daily assessments of high-risk areas are performed, and the results (R3) are streamed to edge devices via a REST API. In conjunction with the DISP-M model, InSAR deformation monitoring is triggered when the R3 probability is greater than 0.8.

[0196] Furthermore, the cloud computing module, based on the deformation threshold model DISP-M and the landslide identification model SE GM, enables refined monitoring and verification of high-risk areas, specifically including:

[0197] Deformation monitoring and dynamic weight calibration:

[0198] The deformation threshold model DISP-M extracts areas with high risk for three consecutive days from the regional landslide risk assessment result R2, generates a boundary mask MASK-1, collects Sentinel-1InS AR data of the MASK-1 area and performs preprocessing, including deformation index quantification and dynamic weight calibration.

[0199] Dynamic weight calibration is based on an InSAR deformation feature library of historical landslide events. A transfer learning algorithm is used to adaptively adjust the weights of surface deformation sensitivity factors, where the weight coefficient W... dynamic (t) is dynamically updated according to seasonal rainfall changes;

[0200]

[0201] In the formula, ΔR(t) is the rate of change of rainfall, α1 and β are the federated learning adaptive parameters, and C hist The contribution value of the sensitive factor to the InSAR deformation feature library in historical landslide events;

[0202] Output surface deformation risk distribution map R4 based on preprocessed Sentinel-1 InSAR data;

[0203] Landslide boundary identification constrained by topographic gradient:

[0204] The landslide identification model SEG-M is based on the Swin-Transformer architecture and introduces terrain gradient constraints through a spatial attention mechanism, specifically:

[0205]

[0206] In the formula, Q i For the query vector, K j Let d be the key vector. k Represents the vector dimension. The elevation gradient is given by α2 = 0.3, which is the adjustment coefficient.

[0207] In this embodiment of the invention, the deformation index quantification of the pretreated seed is specifically as follows:

[0208] Deformation index quantification: Calculation of maximum deformation rate (r) 6m-max ), cumulative deformation (d) total ), daily average deformation (d) avg ) and nearest-neighbor day shape variables (d latest The contribution of each indicator to the landslide probability (e.g., d in loose rock areas) was quantified using Monte Carlo simulation (1000 iterations). total (Weight w1 = 0.4).

[0209] The early warning mechanism of the Deformation Threshold Model (DISP-M) is as follows:

[0210] 1. When r 6m-max Exceeding the threshold and with precipitation forecast to reach TH in the next 3 days t,3d At that time, a red alert was triggered;

[0211] 2. The early warning results are pushed to the edge computing module in real time via the MQTT protocol, triggering the emergency response process.

[0212] Furthermore, the landslide identification model SEG-M is a landslide identification model based on the Swin-Transformer architecture. By default, it uses Sentinel-2 multi-band data (such as shortwave infrared SWIR) and STRM topographic data (slope and aspect) to identify and label landslide areas in images. When Sentinel-2 data for a region is unavailable, it uses 30-meter-level Landsat images and super-resolution reconstruction technology to fill in the gaps by default. In addition, adversarial examples such as cloud and fog obscuration and shadow noise are added during the training phase to ensure the model's accuracy under adverse weather conditions.

[0213] In this embodiment of the invention, the Swin-Transformer architecture supports small-scale shallow landslides (area <100m²). 2 Fine segmentation reduces the false negative rate;

[0214] The specific process of boundary extraction for the SEG-M landslide identification model is as follows:

[0215] 1. Input data: R3 high-risk area mask (MASK-1); Sentinel-2 multispectral image (10-meter resolution, bands 1-10); Topographic data: slope, aspect (10-meter grid);

[0216] 2. Model Inference: Normalize and augment the MASK-1 region image; output a binary segmentation map (landslide area = 1, non-landslide area = 0), and optimize the boundary smoothness through morphological filtering.

[0217] 3. Result verification: Cross-validation with GNSS field monitoring data; missed areas were manually reviewed and added to the sample library to drive incremental learning of the model.

[0218] Furthermore, the edge-cloud collaborative communication module enables data interaction between the cloud computing module and the edge computing module as follows:

[0219] The cloud computing module utilizes GPU clusters for parallel computing, dividing tasks by geographical blocks; based on the idle rate of computing nodes, high-risk areas are prioritized for scheduling to high-performance GPU nodes; static data (topography, lithology) is stored in a distributed manner, cached by region to reduce I / O latency; dynamic data (meteorology, InSAR) is managed through a time-series database (such as InfluxDB), supporting rapid retrieval and updates; multi-source data (satellite imagery, weather forecasts, geological data) are uniformly converted to UTC timestamps and WGS84 coordinate system through an ETL pipeline; spatiotemporal alignment errors are controlled within 10 meters (spatial) and 1 hour (time); high-risk areas (R2=1) trigger cascaded computation of the medium-to-high resolution regional landslide risk assessment model LSA-M2 and the landslide identification model SEG-M; in this embodiment of the invention, the cloud computing module regularly performs full model parameter updates daily, and incremental training data comes from a semi-supervised sample library;

[0220] The edge computing module uses distillation technology to transform the medium-to-high resolution regional landslide risk assessment model LSA-M2 into the lightweight landslide risk assessment model LSA-M2D, retaining the core parameters of the residual correction module. It also rectifies the landslide identification model SEG-M into the lightweight landslide identification model SEG-MD, removing redundant attention heads while retaining multi-scale skip connections. When memory is less than 4GB, the residual correction module of the lightweight landslide risk assessment model LSA-M2D is disabled; when computation latency is greater than 5 seconds, the input data is dynamically downsampled to a 30-meter resolution to ensure real-time response. Static data is pre-stored on edge devices and updated weekly via cloud synchronization; dynamic data is acquired from local weather stations via the LoRaWAN protocol.

[0221] In this embodiment of the invention, the edge computing device includes a drone payload, an Internet of Things terminal, etc.

[0222] The edge computing module publishes the coordinates of high-risk areas (GeoJSON format) via the MQTT protocol. The data packet size is compressed to support transmission in weak network environments (2G / narrowband IoT). Upon receiving an MQTT request, the cloud computing module prioritizes scheduling idle nodes for detection and transmits the data in chunks via HTTP / 2 streaming. The transmission priority is deformation warning data (R3) > risk distribution map (R2) > historical data.

[0223] When the network is down, the edge computing module caches 72 hours of computing resources (model parameters + static data) locally, supporting fully offline operation; after the network is restored, the cloud load is reduced through differential synchronization (only uploading changed data).

[0224] Furthermore, the federated learning module enables collaborative training under data privacy protection in different regions as follows:

[0225] The edge computing module server uses Paillier homomorphic encryption for local gradients, and the ciphertext is directly aggregated in the cloud to prevent the leakage of the original data. The encrypted gradient is transmitted through threshold signature to prevent man-in-the-middle attacks.

[0226] Weights are dynamically allocated based on data quality scores:

[0227]

[0228] Q i =αRecall i +βPrecision i +γCompleteness i

[0229] In the formula, w i As the weight, Q i For quality scoring, α, β, and γ are weighting coefficients, which are dynamically adjusted according to the regional landslide monitoring needs. i Precision is the key factor for data recall. i For data accuracy, Completeness i For data integrity, Q i Low quality region (Q < 0.5) i <0.5) The weight is forcibly decayed to below 0.1 to suppress noise interference;

[0230] Cross-region feature alignment: Before federated aggregation, maximum mean difference (MMD) regularization is applied to the parameters of each edge model to minimize the difference in regional feature distribution;

[0231] Elastic Weight Solidification (EWC): Shared layer parameters (such as the spatiotemporal attention layer of LSA-M2) are solidified to retain global feature extraction capabilities; Local adaptation layers (such as regional lithology coding modules) allow for edge fine-tuning to adapt to regional specificity;

[0232] Training process closed loop: After each round of federated training, the cloud computing module distributes the global model to the edge computing module and updates the local distillation model synchronously; the edge computing module detects the local data distribution shift through adversarial verification and dynamically adjusts the training sampling weights (such as the weight of data with sudden deformation ×2).

[0233] Furthermore, it also includes an iterative optimization module, which achieves self-iterative optimization of the system based on a dynamic optimization mechanism using a supervised sample library and incremental learning and catastrophic forgetting suppression.

[0234] Dynamic optimization mechanism of semi-supervised sample libraries:

[0235] Automatic calibration of classification threshold C-TH: from the prediction sample library PL i Select 10% of the samples with the highest uncertainty (model predicted probability between 0.4 and 0.6), manually label them, and add them to the training set; reclassify the remaining 90% of samples using the updated threshold, and then select 10% of these new samples for manual review; if the review recall rate R... 10p If the value is ≥0.95, the remaining samples are automatically labeled; otherwise, iteration continues. After each iteration, based on the recall and precision curves of the labeled samples, a binary search method is used to determine the minimum threshold C-THmin, ensuring R0. 10p ≥0.95; when R is used for two consecutive rounds 10p When ≥0.95, the combination of terrain abrupt change rate (e.g., slope change >15%) and deformation rate (r) 6m-max >Threshold), automatically assign pseudo-labels to the remaining unlabeled samples and add them to the training set;

[0236] During the iteration process, 10% of the samples are labeled in the first round, 9% remain in the second round, and so on, until the total amount of manual labeling is ≤5%. Samples with an automatic label confidence score <0.7 are marked as "to be reviewed" and are randomly selected by experts for inspection. When the inspection error is >5%, the model is fully retrained.

[0237] Incremental learning and catastrophic forgetting inhibition:

[0238] The importance of model parameters is calculated using the Fisher information matrix, with the following formula:

[0239]

[0240] In the formula, F i Here is the Fisher matrix; θ i Let be the i-th parameter of the model, representing the historically optimal parameter; This is an old task dataset; For the expected value of the data, represent the expectation of the data. The average of all samples in the sample;

[0241] With the shared layer parameters fixed, only the local adaptation layer is fine-tuned; the loss function is:

[0242]

[0243] in, λ is the loss function; λ is the regularization coefficient, used to balance the learning weights of new and old tasks, with a value of 0.5. The optimal parameter values ​​after training the old task; F i Fisher information for parameters, reflecting their importance;

[0244] Adversarial example generation: Perturbations are added to the input data using the fast gradient sign method to simulate distribution shift;

[0245]

[0246] Where ∈ represents the perturbation intensity, controlling the adversarial sample x. 对抗 The offset range is 0.05. This represents the gradient of the loss function with respect to the input;

[0247] Distribution offset detection:

[0248]

[0249] Among them, D KL P represents the Kullback-Leibler divergence, quantifying the difference between the old and new distributions. 新 P represents the distribution of the new data; 旧 This represents the distribution of old data. When the distribution shifts, the historical sample sampling weight is increased by 1.5. Regions with severe shifts trigger full model retraining.

[0250] In this embodiment of the invention, self-iterative optimization also includes local model fine-tuning and transformation of the edge computing device, specifically:

[0251] Local distillation model optimization:

[0252] 1. Fine-tuning data sources: historical landslide samples within the regional boundary (R4-Edge / R5-Edge labeled data); local sensor data (GNSS deformation, real-time soil moisture monitoring).

[0253] 2. Lightweight fine-tuning strategy: Freeze the encoder parameters of SEG-MD and fine-tune only the cross-scale connection layers of the decoder; the amount of training data per round is ≤1000S13-2:

[0254] Transformation matrix construction:

[0255] 1. Principal component features are extracted from the cloud-based global model and the local distillation model using Singular Value Decomposition (SVD);

[0256]

[0257] In the formula, W 全局 The parameter matrix of the global model in the cloud; W 本地Let V represent the parameter matrix of the edge-local model; U is the left singular vector matrix, representing the orthogonal basis of the feature space; Σ represents the singular value matrix, with diagonal elements indicating feature importance; V T Let be the transpose of the right singular vector matrix, and let represent the orthogonal basis of the sample space.

[0258] 2. Construct an objective function that minimizes the differences and optimize the transformation matrix:

[0259]

[0260] Where F represents the Frobenius norm, which measures the difference between matrices; W represents the transformation matrix to be optimized, used to align features between cloud and local models.

[0261] 3. Closed-form solution:

[0262]

[0263] Performance evaluation and feedback:

[0264] The fine-tuned model improves accuracy by ≥5%, allowing replacement of the original edge model; the optimized parameters are uploaded encrypted through the federated learning framework and participate in global model aggregation.

[0265] and Figure 1 Corresponding to the aforementioned system, this invention also discloses a landslide prediction and early warning method based on remote sensing technology, such as... Figure 2 As shown, it includes the following steps:

[0266] S1. Using historical data and the STH-M induced characteristic threshold model, a rapid classification of landslide risk over a large area is completed, and the classification result R1 is obtained.

[0267] S2. The regional landslide risk is assessed by combining the graded result R1 and the pre-trained medium- and low-resolution fusion assessment model LSA-M1, resulting in the assessment result R2.

[0268] S3, the medium-to-high resolution regional landslide risk assessment model LSA-M2 is based on the regional landslide risk assessment result R2 to perform a fusion assessment of regional landslide risk and obtain the fine assessment result R3 of small-scale landslide risk;

[0269] S4. The deformation threshold model DISP-M extracts areas with high risk for three consecutive days from the regional landslide risk assessment result R2, generates a boundary mask MASK-1, collects Sentinel-1 InSAR data of the MASK-1 area and preprocesses it, and makes an early warning based on the preprocessed Sentinel-1 InSAR data, outputting a deformation risk distribution map R4.

[0270] S5. The landslide identification model SEG-M extracts all landslide boundaries within the region based on the variable risk distribution map R4, and obtains the landslide boundary extraction result R5.

[0271] S6. Using the regional landslide risk assessment result R2 and the landslide identification model SEG-M, the landslide boundary is extracted again to obtain the landslide boundary extraction result R6. R6 is used as a supplement to R5. Duplicate areas are removed, and the result of R5 is used first for conflict areas to generate the final landslide boundary map.

[0272] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Regarding the methods disclosed in the embodiments, since they correspond to the systems disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the system section description.

[0273] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A landslide prediction and early warning system based on remote sensing technology, characterized in that, include: The data acquisition module acquires multi-source data within the monitoring area; The cloud computing module is used for large-scale landslide risk assessment and deformation monitoring. It is equipped with the induced feature threshold model STH-M, the low-resolution fusion assessment model LSA-M1, the medium-to-high resolution regional landslide risk assessment model LSA-M2, the deformation threshold model DISP-M, and the landslide identification model SEG-M. The edge computing module is used for landslide risk assessment and deformation monitoring in specific areas, and is equipped with the lightweight landslide risk assessment model LSA-M2D and the lightweight landslide identification model SEG-MD. The edge-cloud collaborative communication module enables data interaction between the cloud computing module and the edge computing module. The federated learning module enables collaborative training under the protection of data privacy in different regions; The landslide early warning module provides early warnings based on landslide risk assessment and deformation monitoring results from cloud computing and edge computing modules.

2. The landslide prediction and early warning system based on remote sensing technology according to claim 1, characterized in that, The cloud computing module assesses landslide risk based on the Induced Feature Threshold Model (STH-M), the Low-Resolution Fusion Assessment Model (LSA-M1), and the Medium-to-High-Resolution Regional Landslide Risk Assessment Model (LSA-M2). Specifically: The STH-M induced feature threshold model rapidly generates a large-scale landslide risk classification map R1 based on induced feature risk thresholds, meteorological forecast data, and satellite remote sensing data. The low-resolution fusion assessment model LSA-M1 performs a fusion assessment of regional landslide risk based on the large-scale landslide risk classification map R1, and obtains the regional landslide risk assessment result R2. The medium-to-high resolution regional landslide risk assessment model LSA-M2 integrates the regional landslide risk assessment results R2 to obtain the detailed assessment results R3 of small-scale landslide risk.

3. A landslide prediction and early warning system based on remote sensing technology according to claim 2, characterized in that, The input data for the STH-M induced feature threshold model is the induced feature vector. A landslide risk classification map R1 is generated by weighting the induced feature vector with risk feature thresholds based on dynamic statistics of historical landslide events. The induced feature vector includes three remote sensing derived parameters: meteorological, geological and topographical, and soil environmental parameters. The geospatial weight matrix W... g With geological condition weight matrix W s Dynamically adjust risk characteristic thresholds to achieve regional differentiation adaptation; The specific steps for constructing the induced feature threshold model STH-M are as follows: Construction of a historical landslide event database: Collecting historical landslide event data and correlating it with multi-source remote sensing data at the time of the events; Extract the induced feature vector for each event point Where, f rain For rainfall characteristics, f soil For soil moisture characteristics, f terrain For terrain features, f ndvi f represents the vegetation cover index characteristic. lith Geological lithology classification characteristics; Risk characteristic threshold quantification: Based on statistical analysis, the critical threshold ranges for each induced characteristic were determined: Features exceeding the threshold are marked as high-risk. Introducing geospatial heterogeneity weights ω g and geological condition weight ω s Generate a differential weight matrix W = ω g ×ω s ; Weighted risk fusion model: Establish risk index R index Calculation formula: In the formula, The midpoint of the threshold interval; According to the risk index R index Risk is divided into four levels: 0-0.3 is low risk, 0.3-0.6 is medium risk, 0.6-0.9 is high risk, and greater than 0.9 is extremely high risk.

4. A landslide prediction and early warning system based on remote sensing technology according to claim 2, characterized in that, The low-resolution fusion evaluation model LSA-M1 is as follows: Dual-channel input architecture: Dynamic remote sensing channel: Input NDVI, surface temperature and daily cumulative precipitation (K), and unify them to 1 km resolution through spatiotemporal alignment; Static geological channel: Input slope, lithology classification and geological fracture density, and normalize and fuse them through feature importance weight matrix; Intelligent data imputation unit: For missing areas caused by cloud cover, a complete data field is generated using historical time series similarity matching and regression imputation algorithms; Risk assessment dynamic triggering mechanism: In the large-scale landslide risk classification map R1, low-risk areas undergo regular assessments; high-risk areas are automatically triggered for detailed assessment using the medium-to-high resolution model LSA-M2.

5. A landslide prediction and early warning system based on remote sensing technology according to claim 2, characterized in that, The medium-to-high resolution regional landslide risk assessment model LSA-M2 includes: Spatiotemporal feature decoupling unit: Time-series sensing head: Analyzes periodic patterns in meteorological data and extracts the evolution characteristics of inducing factors; Spatial correlation head: captures the spatial coupling relationship of terrain features and outputs an attention weight matrix to identify key inducing factors; Geological deviation correction unit: Dynamic correction coefficients are generated based on static geological data, and systematic biases in weather forecasts are calibrated through residual connections. The calculation process is as follows: In the formula, This is the corrected predicted value; Y base The baseline model prediction; G static This is a static geological feature vector; Δ g eo represents the geological deviation compensation amount; U represents the learnable parameter matrix; σ represents the Sigmoid activation function; This is element-wise multiplication; Resolution adaptive mechanism: By default, risk assessments are performed at a spatial resolution of 10 meters, which can be dynamically switched to a range of 5 to 30 meters based on data availability. The parameter matrix U is generated through the following steps: constructing a historical landslide event-geological feature mapping table; using ridge regression to solve for the coefficient matrix that minimizes the prediction bias; and aggregating parameters from multiple regions through federated learning to achieve generalization enhancement. The temporal sensing head in the spatiotemporal feature decoupling unit supports sliding time window analysis, with the window length dynamically adapting to the monsoon / dry season cycle; the spatial correlation head fuses InSAR coherence maps to improve the accuracy of active fault zone identification. Geological deviation compensation amount Δ geo The calculation satisfies: Δ geo =β·log(1+CrackDensity)×RockSoftnessIndex In the formula, β is the regional calibration coefficient, CrackDensity is the crack density, RockSoftnessIndex is the rock softening index, and the remote sensing calculation steps for RockSoftnessIndex are as follows: Clay mineral content index (CMI) was calculated using near-infrared (SWIR) band data from multispectral remote sensing data. In the formula, ρ is the surface reflectance, and SWIR1 and SWIR2 are adjacent near-infrared bands, respectively. Elastic modulus ratio estimation: Acquire synthetic aperture radar (SAR) images and extract the surface deformation rate V using temporal interferometry (InSAR). def Combined with meteorological precipitation data V rain Calculate the modulus ratio: In the formula, k1 is the regional calibration coefficient; Rock softening index generation: Normalize the CMI to the 0-1 range to obtain the CMI. norm The final index is calculated as follows: Where, E sat E represents the elastic modulus of the rock mass under saturated conditions. dry This represents the elastic modulus of the rock mass under dry conditions.

6. A landslide prediction and early warning system based on remote sensing technology according to claim 2, characterized in that, The cloud computing module, based on the deformation threshold model DISP-M and the landslide identification model SEG-M, enables refined monitoring and verification of high-risk areas, specifically including: Deformation monitoring and dynamic weight calibration: The deformation threshold model DISP-M extracts areas with high risk for three consecutive days from the regional landslide risk assessment result R2, generates a boundary mask MASK-1, collects Sentinel-1 InS AR data of the MASK-1 area and performs preprocessing, including deformation index quantification and dynamic weight calibration. Dynamic weight calibration is based on an InSAR deformation feature library of historical landslide events. A transfer learning algorithm is used to adaptively adjust the weights of surface deformation sensitivity factors, where the weight coefficient W... dynamic (t) is dynamically updated according to seasonal rainfall changes; In the formula, ΔR(t) is the rate of change of rainfall, α1 and β are the federated learning adaptive parameters, and C hist The contribution value of the sensitive factor to the InSAR deformation feature library in historical landslide events; Output surface deformation risk distribution map R4 based on preprocessed Sentinel-1 InSAR data; Landslide boundary identification constrained by topographic gradient: The landslide identification model SEG-M is based on the Swin-Transformer architecture and introduces terrain gradient constraints through a spatial attention mechanism, specifically: In the formula, Q i For the query vector, K j Let d be the key vector. k Represents the vector dimension. The elevation gradient is given by α2 = 0.3, which is the adjustment coefficient.

7. A landslide prediction and early warning system based on remote sensing technology according to claim 6, characterized in that, The landslide identification model SEG-M is a landslide identification model based on the Swin-Transformer architecture. By default, it uses Sentinel-2 multi-band data and STRM topographic data to identify and label landslide areas in images. When Sentinel-2 data for a region is unavailable, it uses 30-meter-level Landsat images and super-resolution reconstruction technology to fill in the gaps. In addition, adversarial examples such as cloud and fog obscuration and shadow noise are added during the training phase to ensure the model's accuracy under adverse weather conditions.

8. A landslide prediction and early warning system based on remote sensing technology according to claim 1, characterized in that, The edge-cloud collaborative communication module enables data interaction between the cloud computing module and the edge computing module in the following ways: The cloud computing module utilizes GPU clusters for parallel computing, dividing tasks by geographical blocks; based on the idle rate of computing nodes, high-risk areas are prioritized for scheduling to high-performance GPU nodes; static data is stored in a distributed manner, cached by region to reduce I / O latency; dynamic data is managed through a time-series database, supporting rapid retrieval and updates; multi-source data is uniformly converted to UTC timestamps and WGS84 coordinates through an ETL pipeline; spatiotemporal alignment errors are controlled within 10 meters and 1 hour; high-risk areas trigger cascaded computation of the medium-to-high resolution regional landslide risk assessment model LSA-M2 and the landslide identification model SEG-M. The edge computing module uses distillation technology to transform the medium-to-high resolution regional landslide risk assessment model LSA-M2 into the lightweight landslide risk assessment model LSA-M2D, retaining the core parameters of the residual correction module. It also rectifies the landslide identification model SEG-M into the lightweight landslide identification model SEG-MD, removing redundant attention heads while retaining multi-scale skip connections. When memory is less than 4GB, the residual correction module of the lightweight landslide risk assessment model LSA-M2D is disabled; when computation latency is greater than 5 seconds, the input data is dynamically downsampled to a 30-meter resolution to ensure real-time response. Static data is pre-stored on edge devices and updated weekly via cloud synchronization; dynamic data is acquired from local weather stations via the LoRaWAN protocol. The edge computing module publishes the coordinates of high-risk areas via the MQTT protocol. Upon receiving an MQTT request, the cloud computing module prioritizes scheduling idle nodes for detection and transmits the data in chunks via HTTP / 2 streaming. The transmission priority is: deformation early warning data > risk distribution map > historical data. When the network is down, the edge computing module caches computing resources locally for 72 hours, supporting fully offline operation; after the network is restored, the cloud load is reduced through differential synchronization.

9. A landslide prediction and early warning system based on remote sensing technology according to claim 1, characterized in that, The federated learning module enables collaborative training across different regions while protecting data privacy, specifically as follows: The edge computing module server uses Paillier homomorphic encryption for local gradients, and the ciphertext is directly aggregated in the cloud to prevent the leakage of the original data. The encrypted gradient is transmitted through threshold signature to prevent man-in-the-middle attacks. Weights are dynamically allocated based on data quality scores: Q i =αRecall i +βPrecision i +γCompleteness i In the formula, w i As the weight, Q i For quality scoring, α, β, and γ are weighting coefficients, which are dynamically adjusted according to the regional landslide monitoring needs. i Precision is the key factor for data recall. i For data accuracy, Completeness i For data integrity, Q i Low quality region (Q < 0.5) i <0.5) The weight is forcibly decayed to below 0.1 to suppress noise interference; Cross-regional characteristics Alignment: Before federated aggregation, maximum mean difference regularization is applied to the parameters of each edge model to minimize the difference in regional feature distribution. Elastic weight solidification: Shared layer parameters are solidified to retain global feature extraction capabilities; Local adaptation layers allow for edge fine-tuning, adapting to specific regions. Training process closed loop: After each round of federated training, the cloud computing module distributes the global model to the edge computing module and updates the local distillation model synchronously; the edge computing module detects the local data distribution shift through adversarial verification and dynamically adjusts the training sampling weights.

10. A landslide prediction and early warning system based on remote sensing technology according to claim 1, characterized in that, It also includes an iterative optimization module, which achieves self-iterative optimization of the system based on a dynamic optimization mechanism using a supervised sample library and incremental learning and catastrophic forgetting suppression. Dynamic optimization mechanism of semi-supervised sample libraries: Automatic calibration of classification threshold C-TH: from the prediction sample library PL i The 10% samples with the highest uncertainty are selected, manually labeled, and added to the training set. The remaining 90% of samples are then reclassified using the updated threshold, and another 10% are selected for manual review. If the review recall rate R... 10p If the value is ≥0.95, the remaining samples are automatically labeled; otherwise, iteration continues. After each iteration, based on the recall and precision curves of the labeled samples, a binary search method is used to determine the minimum threshold C-THmin, ensuring R0. 10p ≥0.95; when R is used for two consecutive rounds 10p When the value is ≥0.95, the remaining unlabeled samples are automatically assigned pseudo-labels and added to the training set by combining the terrain mutation rate and deformation rate. During the iteration process, 10% of the samples are labeled in the first round, 9% remain in the second round, and so on, until the total amount of manual labeling is ≤5%. Samples with an automatic label confidence score <0.7 are marked as "to be reviewed" and are randomly selected by experts for inspection. When the inspection error is >5%, the model is fully retrained. Incremental learning and catastrophic forgetting inhibition: The importance of model parameters is calculated using the Fisher information matrix, with the following formula: In the formula, F i Here is the Fisher matrix; θ i Let be the i-th parameter of the model, representing the historically optimal parameter; This is an old task dataset; For the expected value of the data, represent the expectation of the data. The average of all samples in the sample; Fixed shared layer parameters, only minor adjustments to the local adaptation layer; The loss function is: in, λ is the loss function; λ is the regularization coefficient, used to balance the learning weights of new and old tasks, with a value of 0.

5. The optimal parameter values ​​after training the old task; F i Fisher information for parameters, reflecting their importance; Adversarial example generation: Perturbations are added to the input data using the fast gradient sign method to simulate distribution shift; Where ∈ represents the perturbation intensity, controlling the adversarial sample x. 对抗 The offset range is 0.

05. This represents the gradient of the loss function with respect to the input; Distribution offset detection: Among them, D KL P represents the Kullback-Leibler divergence, quantifying the difference between the old and new distributions. 新 P represents the distribution of the new data; 旧 This represents the distribution of old data. When the distribution shifts, the historical sample sampling weight is increased by 1.

5. Regions with severe shifts trigger full model retraining.

11. A landslide prediction and early warning method based on remote sensing technology, characterized in that, Includes the following steps: S1. Using historical data and the STH-M induced characteristic threshold model, a rapid classification of landslide risk over a large area is completed, and the classification result R1 is obtained. S2. The regional landslide risk is assessed by combining the graded result R1 and the pre-trained medium- and low-resolution fusion assessment model LSA-M1, resulting in the assessment result R2. S3, the medium-to-high resolution regional landslide risk assessment model LSA-M2 is based on the regional landslide risk assessment result R2 to perform a fusion assessment of regional landslide risk and obtain the fine assessment result R3 of small-scale landslide risk; S4. The deformation threshold model DISP-M extracts areas with high risk for three consecutive days from the regional landslide risk assessment result R2, generates a boundary mask MASK-1, collects Sentinel-1 InSAR data of the MASK-1 area and preprocesses it, and makes an early warning based on the preprocessed Sentinel-1 InSAR data, outputting a deformation risk distribution map R4. S5. The landslide identification model SEG-M extracts all landslide boundaries within the region based on the variable risk distribution map R4, and obtains the landslide boundary extraction result R5. S6. Using the regional landslide risk assessment result R2 and the landslide identification model SEG-M, the landslide boundary is extracted again to obtain the landslide boundary extraction result R6. R6 is used as a supplement to R5. Duplicate areas are removed, and the result of R5 is used first for conflict areas to generate the final landslide boundary map.

Images