A method for early identification of landslide disasters based on Beidou unmanned aerial vehicle and remote sensing

Abstract

This invention discloses an early identification method for landslide disasters based on BeiDou UAV and remote sensing, belonging to the field of geological disaster monitoring technology. The method first collaboratively acquires and preprocesses multi-source data from SAR imagery, BeiDou GNSS, and UAVs; then performs temporal coupling fusion; subsequently, it uses a CNN network to extract and fuse multi-source features of SAR deformation intensity, BeiDou displacement dynamic trends, UAV surface rupture, and 3D terrain; the fused features are input into a pre-constructed ConvNeXt-ViT dual-path landslide early warning model, outputting the landslide risk level and deformation rate for the next 15 days; finally, a closed-loop early warning is executed within 10 minutes through an integrated hardware and software early warning system. This invention achieves comprehensive monitoring of landslides from macro to meso to micro levels, with strong cross-regional adaptability, significantly improving early warning accuracy and timeliness, with a false alarm rate of <10% and a missed alarm rate of <5%.

Description

Technical Field

[0001] This invention relates to a method for early identification of landslide disasters based on Beidou unmanned aerial vehicles and remote sensing, belonging to the field of landslide disaster monitoring and early warning technology. Background Technology

[0002] Currently, the main methods for identifying landslides include total station measurement, UAV remote sensing monitoring, BeiDou GNSS measurement, and synthetic aperture radar differential interferometry (InSAR) measurement.

[0003] Manual landslide monitoring involves personnel using equipment such as total stations and inclinometers to set up monitoring points in the landslide area. They then measure parameters such as point coordinates, crack width, and slope gradient, comparing data from different time periods to identify potential landslide hazards. This method is characterized by its direct operation, low equipment barrier to entry, and ability to conduct detailed measurements in key local areas. It can directly obtain macroscopic deformation characteristics of the landslide surface. However, manual monitoring is greatly limited by terrain conditions (such as steep slopes and cliffs), poses personnel safety risks, has low efficiency, long monitoring cycles, difficulty in achieving large-scale continuous dynamic monitoring, and the data is highly subjective and susceptible to human error.

[0004] Unmanned aerial vehicle (UAV) remote sensing monitoring utilizes UAVs equipped with optical cameras and LiDAR (Light Detection and Ranging) to fly over landslide risk areas. It generates digital orthophoto maps (DOM) and digital surface models (DSM) through photogrammetry, or acquires point cloud data using LiDAR to generate digital ground models (DEM). This allows for the extraction of features such as landslide crack distribution, boundary range, and surface undulations, constructing a three-dimensional virtual reality model. It features high spatial resolution (up to centimeter level) and strong localized detailed monitoring capabilities, quickly covering small to medium-sized landslide areas and visually presenting microscopic deformations on the landslide surface (such as minute cracks). However, UAV monitoring is significantly affected by weather conditions (heavy rain, dense fog, strong winds), making flight impossible in severe weather. Flight requires airspace permission and is subject to significant airspace control restrictions. The massive data volume leads to time-consuming preprocessing (such as point cloud filtering and image stitching), making it difficult to meet real-time monitoring needs. Furthermore, periodic flights (such as quarterly or monthly flights) incur high equipment wear and manpower costs, making it impossible to continuously capture the dynamic deformation process of landslides.

[0005] InSAR (Synthetic Aperture Radar Interferometry) technology utilizes multiple SAR images acquired by satellites (such as Gaofen-3). Through image registration, interferometric pair screening, differential interferogram generation, and phase unwrapping, the phase difference is combined with the SAR wavelength to convert it into surface deformation along the radar line of sight, enabling landslide area deformation monitoring. It features large-scale (single coverage of hundreds to thousands of square kilometers) and all-weather (unaffected by day / night, cloud cover, or fog) monitoring capabilities. It can capture low-frequency deformation trends of landslides at a regional scale and can continuously acquire surface deformation data over long periods. However, InSAR technology is susceptible to atmospheric delays (such as tropospheric and ionospheric disturbances) and topographic residuals, leading to errors in deformation calculations. Its low temporal resolution (typically monthly) prevents it from capturing short-term high-frequency landslide displacements (such as daily or hourly deformations). Furthermore, it lacks sufficient accuracy in identifying small-scale features (such as fine cracks and local faults within tens of meters), making it difficult to reflect the microscopic deformation details of landslides.

[0006] BeiDou GNSS (Global Navigation Satellite System) monitoring involves deploying BeiDou base stations in stable bedrock areas outside landslide zones and GNSS rover stations in high-risk landslide areas. High-precision receivers receive BeiDou satellite signals, and after data preprocessing (integrity checks, gross error removal, and error correction) and precise positioning calculations, three-dimensional displacement data of the monitoring points is obtained, allowing for analysis of the dynamic deformation patterns of the landslide body. It features millimeter-level positioning accuracy and high-frequency dynamic monitoring (reaching minute or even second-level accuracy), simultaneously acquiring horizontal and vertical displacement information of the landslide body. It is not limited by line-of-sight conditions and can continuously output displacement data over long periods. However, BeiDou GNSS monitoring has limited coverage, enabling only "point-based" monitoring (dependent on the location of the rover stations), and cannot reflect the overall deformation distribution of the landslide area. In complex terrains such as deep mountains, canyons, and dense forests, satellite signals are easily blocked, leading to decreased data integrity and positioning accuracy. Furthermore, the equipment purchase, installation, and maintenance costs for the base stations and multiple rover stations are high, making large-scale, dense deployment economically unfeasible.

[0007] In summary, single monitoring technologies have inherent limitations in terms of coverage, accuracy, timeliness, and cost, making it difficult to meet the comprehensive, high-precision, and real-time dynamic needs for early identification and warning of landslide disasters, encompassing "macro-trends, meso-level locations, and micro-level details." Therefore, there is an urgent need for a comprehensive solution that can effectively integrate multi-source spatiotemporal data, fully leverage the advantages of each source, and perform intelligent analysis and decision-making. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a solution to the limitations of existing single landslide monitoring technology in terms of spatiotemporal scale, accuracy dimension and timeliness. It can effectively obtain the evolution law of landslide identification indicators and improve the efficiency of early landslide identification.

[0009] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0010] This invention provides a method for early identification of landslide disasters based on BeiDou UAV and remote sensing. The method includes: acquiring multi-source monitoring data of the monitoring area, wherein the multi-source monitoring data includes at least synthetic aperture radar image data, global navigation satellite system displacement time series data, and UAV observation data; performing collaborative preprocessing on the multi-source monitoring data to obtain preprocessed multi-source monitoring data, wherein the preprocessed multi-source monitoring data includes: surface deformation information, millimeter-level three-dimensional displacement time series, and centimeter-level surface model and micro-fracture features; performing spatiotemporal benchmark unification and multi-frequency time series coupling fusion on the preprocessed multi-source monitoring data to construct a spatiotemporally aligned fused dataset; inputting the fused dataset into a pre-constructed ConvNeXt-ViT landslide early warning model for multi-source feature extraction and collaborative prediction, and outputting the landslide risk level and deformation rate prediction results for the monitoring area in the future preset time period.

[0011] The above setup achieves the following results: This method integrates three technologies—satellite remote sensing for large-scale monitoring (SAR), ground-based high-precision point monitoring (GNSS), and near-ground high-resolution detailed survey (UAV)—through a systematic process. First, targeted collaborative preprocessing addresses the heterogeneity of various data sources, producing high-quality, standardized information on deformation, displacement, and surface features. Second, by unifying spatiotemporal benchmarks and fusing multiple frequencies, it breaks down data barriers, constructing a unified dataset suitable for in-depth analysis. Finally, it utilizes an advanced ConvNeXt-ViT model for end-to-end intelligent prediction, enabling the automatic extraction of landslide evolution patterns from multi-source heterogeneous data. This method overcomes the limitations of single-source monitoring, solves the challenge of deep data fusion, and significantly improves the foresight and accuracy of landslide risk identification through artificial intelligence models.

[0012] Further steps include: collaborative preprocessing of the multi-source monitoring data, including: interferometric stacking of the synthetic aperture radar image data to obtain surface deformation information of the monitoring area; positioning calculation of the global navigation satellite system displacement time series data to obtain millimeter-level three-dimensional displacement time series of the landslide body; delineation of landslide risk areas based on the surface deformation information, and conducting UAV aerial surveys and data processing within the risk areas to obtain centimeter-level digital orthophotos, digital elevation models, and three-dimensional surface models, and extracting micro-fracture features from them.

[0013] The above setup achieves the following effect: This collaborative preprocessing scheme is not simply parallel processing, but rather forms an organic decision chain. InSAR processing is used to obtain a large-scale deformation field for macroscopic screening and preliminary risk area location. Based on this location result, GNSS is deployed at key points and subjected to high-precision calculations, providing a reliable quantitative benchmark. Simultaneously, it guides UAVs to conduct targeted, detailed observations of high-risk areas, acquiring details of surface fracturing. This collaborative logic of "area scanning guiding point monitoring, macroscopic deformation location for detailed exploration" makes the preprocessing processes of the three types of data mutually supportive and multiplies efficiency, providing a high-quality data foundation with breadth, accuracy, and detail for subsequent fusion.

[0014] Further steps include: interferometric stacking processing of the synthetic aperture radar image data, including: selecting single-view complex images that cover the monitoring area and meet preset time baseline and incident angle requirements; using the image with the highest coherence as the main image, registering, radiometrically calibrating, and filtering it; selecting multiple effective interferometric pairs based on short spatial baseline, short temporal baseline, and high coherence criteria, and generating differential interferograms; performing weighted stacking, phase unwrapping, terrain phase removal, and atmospheric error correction on each differential interferogram to obtain the surface deformation along the radar line of sight, and generating surface deformation information after verification with global navigation satellite system data.

[0015] The above setup achieves the following effects: This InSAR processing workflow maximizes the quality and phase stability of interferometric data through rigorous image selection (time baseline, incident angle), precise registration using high-coherence images as the main image, and interferometric pair optimization based on multiple criteria (short baseline, high coherence). The use of weighted stacking (weights correlated with coherence) effectively suppresses random noise and improves the reliability of deformation signals. Finally, cross-validation with GNSS data ensures the absolute accuracy of the InSAR inversion deformation results, transforming them from relative measurements into reliable quantitative deformation information, providing credible input for subsequent analysis.

[0016] Further steps include: performing positioning calculations on the displacement time-series data from the Global Navigation Satellite System, including: deploying a reference station in a stable area outside the landslide-affected zone and multiple rover stations in a high-risk landslide area; performing data integrity checks, gross error removal, format conversion, and preliminary error correction on the raw observation data received by the reference station and rover stations; using dual-frequency observations to eliminate ionospheric delay and using filtering algorithms to suppress multipath effects; performing positioning calculations based on a relative positioning model to output millimeter-level displacement time-series sequences of the landslide body along the east, north, and vertical directions.

[0017] The above setup achieves the following results: This GNSS data processing scheme, through a "stable internal, dynamic external" station deployment strategy (base station in stable areas + rover station in risk areas), creates conditions for high-precision relative positioning. A systematic data preprocessing process (integrity checks, gross error removal) and error correction (ionospheric, multipath) ensures the quality of the raw observation data. Finally, through relative positioning model calculations, common-mode errors can be eliminated, and the three-dimensional displacement changes of the landslide body itself can be accurately extracted, yielding a displacement sequence with millimeter-level accuracy and temporal continuity. This data not only serves as independent deformation monitoring results but, more importantly, provides absolute benchmark verification for InSAR deformation and a source of control point coordinates for UAV observations, acting as a "precision anchor" for the entire multi-source data system.

[0018] Further steps include: delineating landslide risk zones based on the surface deformation information, including: using preset deformation rate thresholds and cumulative displacement thresholds as a basis, and combining the contours of deformation gradient abrupt change zones and continuous high deformation zones, to delineate landslide risk zones through geographic information system buffer analysis; and deploying ground control points calibrated by the Global Navigation Satellite System within the risk zones.

[0019] The above setup achieves the following results: This scheme enables intelligent and quantitative conversion from InSAR deformation fields to UAV-surveyed detailed areas. By setting dual thresholds for deformation rate and cumulative displacement, and combining deformation gradient and spatial continuity analysis, it can scientifically and objectively identify the most likely unstable key areas (risk zones) from a large monitoring area, avoiding the subjectivity and omissions of manual interpretation. Deploying GNSS-calibrated ground control points within this area provides the necessary known control points for subsequent high-precision geometric correction of UAV imagery, ensuring that UAV-generated DOM, DEM, and other products have centimeter-level absolute positioning accuracy, enabling precise spatial overlay of data from different sources.

[0020] Further steps include: unifying the spatiotemporal reference and coupling and fusing multi-frequency time series data after preprocessing. This includes: unifying the coordinate system of all multi-source monitoring data to the target geodetic coordinate system based on preset transformation parameters; unifying the time reference of all multi-source monitoring data to the target time system; matching multi-source time series data with different observation frequencies to a unified time node through interpolation or aggregation methods; and performing correlation analysis and regression modeling on the multi-source data after unifying the spatiotemporal reference to complete the time series coupling and fusion and set dynamic early warning thresholds.

[0021] The above setup achieves the following effects: This fusion process resolves the fundamental obstacles to collaborative analysis of multi-source data. Coordinate system one eliminates systematic biases in spatial location, enabling precise overlay of SAR deformation maps, GNSS points, and UAV imagery. A unified time reference ensures all data are analyzed on the same time axis. Frequency matching resolves the temporal alignment issues between low-frequency SAR, high-frequency GNSS, and discontinuous UAV observations. Based on this physical alignment, correlation analysis and regression modeling can establish quantitative mathematical relationships between different monitoring indicators (such as areal deformation rate and point displacement acceleration) at the information level. The final dynamic warning threshold, due to the fusion of multi-source information and consideration of environmental factors (such as rainfall), is more scientific and reliable than static thresholds based on a single data source, effectively reducing false alarms and missed alarms.

[0022] Further settings: The correlation analysis and regression modeling include: calculating the Pearson correlation coefficient between time series of different source data, establishing a regression model that reflects the quantitative relationship between different monitoring indicators; validating the regression model using reserved data or manually measured data; and setting a dynamic early warning threshold that changes with environmental factors by combining the results of the regression model with the geological characteristics of the landslide.

[0023] The above setup achieves the following effects: This step elevates data fusion from simple spatial overlay to quantitative relationship modeling. By calculating the Pearson correlation coefficient, key issues such as "whether InSAR deformation acceleration is synchronized with GNSS displacement surges" can be quantitatively assessed, allowing for the selection of strongly correlated and effective monitoring indicators. The established regression model can quantitatively describe the synergistic changes among multiple indicators. Validation using reserved data ensures the model's generalization ability and reliability. Finally, by combining dynamic thresholds set based on geological environment (such as soil type and slope) and triggering factors (such as rainfall), the early warning standard is no longer a fixed value but an intelligent rule that adaptively adjusts according to the internal and external conditions of the landslide, significantly improving the adaptability and accuracy of the early warning system.

[0024] Further configuration: The step of inputting the fused dataset into the pre-built ConvNeXt-ViT landslide early warning model for multi-source feature extraction and collaborative prediction includes: extracting synthetic aperture radar deformation intensity features and UAV optical image surface texture features using a preset two-dimensional convolutional neural network; converting global navigation satellite system displacement time series data into two-dimensional maps and extracting displacement dynamic trend features using a convolutional neural network; converting UAV lidar point cloud data into voxel grids or elevation maps and extracting three-dimensional terrain morphology features using a three-dimensional convolutional neural network; stitching and cross-modal fusion of the extracted multi-source features and capturing temporal dynamic correlations through a spatiotemporal recurrent neural network; and outputting landslide risk level classification results and future deformation rate regression prediction results through a fully connected layer based on the fused spatiotemporal features.

[0025] The above setup achieves the following results: This scheme designs a customized feature extraction and fusion pipeline for multi-source heterogeneous data. For SAR imagery (2D images) and DOM (2D images), 2D CNNs are used to extract spatial pattern features; for GNSS sequences (1D time series), they are converted to 2D maps and then CNNs are used to mine their temporal evolution patterns; for LiDAR point clouds (3D space), 3D CNNs are used to capture their three-dimensional terrain morphology features. This "divide and conquer" strategy ensures the full extraction of the most effective features from various data types. Subsequent stitching, cross-modal fusion, and time series modeling enable deep interaction and temporal correlation analysis between different modal features, allowing the model to understand the spatiotemporal coupling relationship between "where deformation intensifies" and "where new cracks appear," thus making more comprehensive and accurate predictions.

[0026] Further configuration: The pre-constructed ConvNeXt-ViT landslide early warning model is a fusion architecture containing a two-dimensional convolutional neural network, a three-dimensional convolutional neural network, and a spatiotemporal recurrent neural network; The process of splicing and cross-modal fusion of the extracted multi-source features is specifically as follows: the feature vectors from different source data are spliced ​​into a high-dimensional fusion feature vector, and then cross-modal feature interaction and dimensionality compression are performed through at least one one-dimensional convolutional layer.

[0027] The above setup achieves the following effect: This model architecture clearly defines the specific technical means of feature fusion. Multi-source feature vectors are directly concatenated, preserving all original feature information. Subsequently, a one-dimensional convolutional layer is used for processing. Its core functions are: 1. To perform cross-channel (i.e., cross-modal) information interaction, allowing the model to learn, for example, the weight relationship between the "SAR deformation feature channel" and the "DOM crack feature channel," strengthening meaningful correlations and suppressing noise; 2. To compress and reduce the dimensionality of high-dimensional features, removing redundant information and extracting the most essential fusion representation. This "concatenation + interactive compression" fusion method is more powerful than simple concatenation or weighted averaging, generating fusion features with higher information density and greater discriminative power, providing better input for the final risk classification and rate regression.

[0028] Further configuration: The pre-constructed ConvNeXt-ViT landslide early warning model is an early warning model based on a hybrid architecture of visual Transformer and convolutional neural network; the input of the early warning model is a multi-channel tensor, which is composed of multiple channels combined from the optical image channel of the UAV observation data, the deformation field normalization channel of the synthetic aperture radar image data, and the displacement time series data slices of the global navigation satellite system; the early warning model extracts local detail features through the convolutional neural network branch, extracts global correlation features through the Transformer encoder branch, and fuses the local detail features and global correlation features through a cross-attention mechanism to comprehensively encode the surface morphology, deformation level, and multi-source observation consistency information, and finally completes the prediction.

[0029] The above setup achieves the following effect: This ConvNeXt-ViT hybrid model architecture fully leverages the respective strengths of CNN and Transformer. CNN branches (such as ConvNeXt blocks) excel at capturing local details and spatially invariant features in images, such as subtle crack edges and small deformation spots, which are crucial for identifying early signs of rupture. Transformer branches, through a self-attention mechanism, can model the long-range dependencies between all pixels (or feature blocks) within the entire monitoring area, thereby grasping the overall deformation pattern, spatial correlation patterns, and consistency between multi-source data channels of the landslide. By fusing local detail features with global contextual features through a cross-attention mechanism, the model can make decisions based on both local evidence and the global situation, achieving a more comprehensive and in-depth understanding of the landslide state, thus greatly improving the accuracy and robustness of predictions.

[0030] Further configuration: The early warning model is trained and optimized through a cross-regional adaptive training framework, which includes: using generative adversarial networks and histogram matching to unify the style and align the features of heterogeneous data in the target region; using a dual-branch structure of spatial convolutional networks and frequency-domain-based Transformer encoders to extract domain-invariant landslide features; using the maximum mean difference loss function to align the feature distributions of the source and target domains; using a small amount of labeled data from the target region to fine-tune the top layer of the model; and introducing a region-sensitive attention mechanism.

[0031] The above setup achieves the following results: This training framework systematically solves the core challenge of cross-regional application of AI models in the field of geological disasters—"domain bias." Style unification (GAN, histogram matching) reduces apparent differences caused by sensor variations, lighting, and seasonality. By designing a specialized dual-branch structure (spatial domain + frequency domain) and supplementing it with MMD loss constraints, the model is forced to learn the common core physical characteristics of landslides across various regions (such as tensile and shear mechanical modes), rather than irrelevant region-specific features. Finally, fine-tuning using a small amount of data from the target region and introducing regional attention is an efficient form of "small sample adaptation," enabling the model to quickly capture and focus on the unique geological and geomorphological characteristics of new regions. This combined strategy ensures the model has strong generalization capabilities, allowing for rapid deployment in new monitoring areas without the need for costly data relabeling and model training, making it highly practical.

[0032] Further configuration: The method also includes: based on the landslide risk level and deformation rate prediction results, executing a closed-loop early warning process through a deployed integrated hardware and software early warning system; the integrated hardware and software early warning system executing the closed-loop early warning process includes: a hardware architecture including a data acquisition layer, a data transmission layer, a data processing layer, and an early warning release layer; wherein, the data transmission layer adopts dual-link backup of mobile communication network and satellite communication; the software system includes: a data platform for multi-source data storage and standardized preprocessing; a model service module encapsulating the early warning model and providing inference services; a decision module for triggering early warning based on dual conditions of the landslide risk level and deformation rate; and a visualization module for multi-dimensional data visualization and historical early warning feedback; the closed-loop early warning process specifically involves: the data acquisition layer periodically collecting and uploading data, which is then pushed to the cloud processing layer for model inference after edge preprocessing; the decision module determining whether to trigger an early warning based on the inference results, and simultaneously releasing early warning information through multiple terminals of the early warning release layer within a preset time; and subsequently, feedback and system optimization are performed through the visualization module.

[0033] The above setup achieves the following effects: It integrates all the aforementioned algorithms and methods into a single, operationally viable system. Dual-link transmission ensures the reliability and continuity of data transmission in harsh field environments. Modular software design decouples and efficiently coordinates data management, intelligent analysis, decision-making, and result display. A clearly defined "collection-processing-decision-release-feedback" closed-loop process specifies standard operating procedures from data to early warning action, ensuring the timeliness of early warnings (e.g., release within 10 minutes). This integrated hardware and software design transforms the technical solution of this invention from an offline analysis method into an automated, real-time, and highly available online early warning service system, truly realizing the transformation of scientific research results into practical productivity for disaster prevention and mitigation, and enhancing emergency response capabilities.

[0034] As a further improvement of the present invention, the method further includes: initiating an early warning when the landslide risk level exceeds the level threshold or the future deformation rate exceeds the deformation threshold.

[0035] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0036] (1) This invention integrates InSAR, Beidou GNSS and UAV three-source monitoring technologies, giving full play to the advantages of InSAR large-scale all-weather deformation monitoring, Beidou GNSS millimeter-level high-frequency displacement capture and UAV fine-grained surface feature extraction, realizing all-round monitoring of landslide “macro trend - mid-point positioning - micro details”, significantly improving the comprehensiveness of landslide deformation identification and data reliability.

[0037] (2) By using the seven-parameter coordinate unification, BDT time synchronization and ConvLSTM time series modeling technology, the intrinsic relationship between landslide InSAR deformation rate, Beidou displacement trend and UAV crack widening was explored, providing a quantitative scientific basis for the dynamic evolution law analysis and accurate early warning of landslide disasters.

[0038] (3) The present invention can predict the dynamic change of the identification index in the process of landslide evolution, which helps to discover the potential risks of landslide disasters in a timely manner and provides time guarantee for taking effective prevention and control measures.

[0039] (4) Based on the SSFSC framework and the regional sensitive attention mechanism, this invention realizes the adaptive adaptation of the model in different geological (rock / soil) and climatic (rainy / dry) regions, effectively solving the problem of weak cross-regional generalization ability of traditional models, and providing a standardized technical solution for multi-scenario landslide monitoring. Attached Figure Description

[0040] Figure 1 This is a flowchart of the method of the present invention.

[0041] Figure 2This is a flowchart of the UAV photogrammetry technology of the present invention.

[0042] Figure 3 This is a landslide after point cloud cropping according to the present invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention. Example 1

[0044] like Figure 1 As shown, this embodiment provides a method for early identification of landslide disasters based on BeiDou unmanned aerial vehicles and remote sensing, including the following steps:

[0045] Step S1: Acquire and preprocess multi-source data of the monitoring area; the multi-source data includes synthetic aperture radar (SAR) image data, BeiDou GNSS signal data, and UAV data, and perform preprocessing on each of them;

[0046] Step S2: Unify the preprocessed BeiDou GNSS data, SAR image data and UAV data to the same coordinate system and the same time reference, and perform time-series coupling fusion to obtain data after reference unification;

[0047] Step S3: Based on the unified benchmark data, extract multi-source features respectively; the multi-source features include the deformation intensity distribution features of the SAR image data, the displacement dynamic trend features of the BeiDou GNSS data, and the surface fracture semantic features and three-dimensional terrain morphology features of the UAV data; fuse the extracted multi-source features to obtain the fused features;

[0048] Step S4: Input the fused features into the pre-constructed ConvNeXt-ViT landslide early warning model and output the landslide prediction results, which include the landslide risk level and future deformation rate.

[0049] In step S1, the preprocessing of SAR image data includes: image registration, image cropping, orbit correction, radiometric correction, and denoising of multiple acquired SAR images; using Stacking-InSAR, effective interferometric pairs (coherence coefficient ≥ 0.4) are selected based on short spatial baselines (≤ 500m) and short temporal baselines (≤ 30 days), differential interferograms are generated and weighted stacked to obtain an initial stacked interferogram; the initial stacked interferogram is phase unwrapped, and after eliminating residual errors from topography and atmosphere, the phase difference is converted into deformation along the radar line of sight to obtain surface deformation information.

[0050] In step S1, the preprocessing of the BeiDou GNSS signal data includes: deploying a BeiDou reference station in the stable bedrock area outside the landslide area, and deploying 3-5 GNSS rover stations in the high-risk landslide area (evenly distributed along the main crack direction of the landslide, with at least one located in the rear tensile crack zone and one located in the front shear zone); performing data integrity checks, gross error removal, format conversion, preliminary error correction, and time synchronization on the received raw signals; and using real-time dynamic positioning (RTK) mode (positioning frequency 1Hz) to remove abnormal data by calculating precise orbit and clock errors, dual-frequency ionospheric correction, and multi-path filtering, and using real-time dynamic positioning (RTK) mode (positioning frequency 1Hz) to complete high-precision positioning calculations and obtain millimeter-level deformation and displacement data of the landslide.

[0051] In step S1, the preprocessing of the UAV data includes: delineating the landslide risk zone based on the surface deformation information; using a UAV equipped with an optical camera and LiDAR equipment to conduct observations with the risk zone as the core flight range; generating a digital orthophoto map (DOM) with centimeter-level resolution through the optical camera; acquiring laser point cloud data through the LiDAR equipment, and generating a digital ground model (DEM) and a digital surface model (DSM) through point cloud filtering; and extracting landslide surface rupture characteristics, crack distribution patterns, boundary ranges, and three-dimensional spatial elements based on the DOM, DEM, and DSM to construct a three-dimensional virtual reality model of the landslide.

[0052] In step S2, unifying to the same coordinate system specifically involves: using the BeiDou reference station as a reference, and employing the seven-parameter method to unify BeiDou rover data, SAR image data, and UAV data to the CGCS2000 coordinate system; unifying to the same time reference specifically involves: converting the time reference of all data from Coordinated Universal Time (UTC) to BeiDou system time (BDT). The time-series coupling and fusion includes: using a monthly time granularity, binding BeiDou high-frequency data, InSAR low-frequency data, and UAV quarterly data to the same time axis to achieve frequency matching; calculating the quantitative correlation between multi-source data using Pearson correlation coefficients and regression models, and setting dynamic early warning thresholds after verification.

[0053] In step S3, the extraction of multi-source features specifically includes: using 2DCNN to extract SAR image deformation intensity distribution features and UAV DOM surface rupture semantic features; converting BeiDou GNSS time-series displacement data into a "time-displacement" two-dimensional map and using CNN to extract displacement dynamic trend features; converting LiDAR point clouds into voxel grids or elevation maps and using 3DCNN to extract three-dimensional terrain morphology features; the multi-source feature fusion involves concatenating the extracted feature vectors and inputting them into a fusion convolutional layer for cross-modal association learning.

[0054] The construction and training method of the ConvNeXt-ViT landslide early warning model includes: preprocessing UAV DOM, InSAR deformation field, and BeiDou time-series data into a 9-channel tensor (3-channel DOM + 1-channel InSAR + 5-channel BeiDou time-series slices) and inputting it into the model; extracting local features of cracks and faults through local feature branches containing 3 layers of ConvNeXt blocks; simultaneously capturing the global spatial correlation of landslides through the ViT branch of a 6-layer Transformer encoder; fusing local and global features into a unified feature vector through a 4-head cross-attention mechanism; training the model using a hybrid loss function (0.6×cross-entropy loss + 0.4×MAE loss), outputting landslide risk level (0-3) classification and deformation rate regression results for the next 15 days. During training, pre-trained weights from the remote sensing dataset are loaded, and a phased unfreezing strategy is adopted, combined with data augmentation and pseudo-label enhancement to improve model accuracy; the final early warning results are optimized through geological prior rule filtering (e.g., removing high-risk samples with slope ≤15°) and multi-model ensemble voting.

[0055] Specifically, acquiring and preprocessing SAR image data includes:

[0056] Acquire multiple SAR image data from the Gaofen-3 satellite within the monitoring area;

[0057] The acquired SAR images were subjected to image registration, image cropping, orbit correction, radiometric correction, and denoising. My suggestion is to remove... Figure 3 As shown, Figure 3 This is a schematic diagram of a landslide after point cloud cropping according to the present invention.

[0058] The interferometric stacking method is used to screen effective interferometric pairs based on short spatial baselines and short temporal baselines, generate differential interferograms, and perform weighted stacking to obtain the initial stacked interferogram;

[0059] After phase unwrapping of the initial stacked interferogram and eliminating residual errors from topography and atmosphere, the phase difference is converted into deformation along the radar line of sight to obtain surface deformation information.

[0060] Specifically, the range of the short spatial baseline is limited to ≤500m, and the range of the short temporal baseline is limited to ≤30 days; when screening effective interferometric pairs, the coherence coefficient is ≥0.4.

[0061] Specifically, in step S1, acquiring and preprocessing BeiDou GNSS signal data includes:

[0062] Deploy BeiDou reference stations in stable bedrock areas outside the landslide area, and deploy 3-5 GNSS mobile stations in high-risk landslide areas to receive raw BeiDou satellite signals;

[0063] Perform data integrity checks, gross error removal, format conversion, preliminary error correction, and time synchronization on the raw BeiDou satellite signals received by the base station and rover station;

[0064] By calculating precise orbit and clock error, dual-frequency ionospheric correction, and multi-path filtering to remove abnormal data, high-precision positioning calculations are completed, and millimeter-level BeiDou GNSS time-series displacement data of landslides are obtained.

[0065] Specifically, the 3-5 GNSS rover stations are evenly distributed along the direction of the main crack of the landslide, with at least one rover station deployed in the tensile crack zone at the rear edge of the landslide and at least one rover station deployed in the shear zone at the front edge of the landslide; the high-precision positioning calculation adopts the real-time dynamic positioning (RTK) mode, and the positioning frequency is set to 1Hz.

[0066] Specifically, in step S1, the UAV observation data is acquired and preprocessed, including:

[0067] Landslide risk zones were delineated based on the aforementioned surface deformation information;

[0068] A drone equipped with an optical camera and LiDAR device was used to conduct observations within the risk area as its core flight range.

[0069] The optical camera generates a digital orthophoto map (DOM) with centimeter-level resolution.

[0070] The LiDAR device acquires laser point cloud data, and the point cloud is filtered to generate a digital ground model (DEM) and a digital surface model (DSM).

[0071] Based on the DOM, DEM and DSM, the surface rupture characteristics, crack distribution patterns, boundary range and three-dimensional spatial elements of the landslide are extracted to construct a three-dimensional virtual reality model of the landslide.

[0072] Specifically, in step S2, the methods for unifying to the same coordinate system include:

[0073] Using the BeiDou reference station as a benchmark, the seven-parameter method is adopted to unify BeiDou rover data, SAR image data, and UAV data into the CGCS2000 coordinate system;

[0074] Methods for unifying to the same time base include:

[0075] Convert the time base of all data from Coordinated Universal Time (UTC) to BeiDou system time (BDT).

[0076] Specifically, in step S2, temporal coupling fusion is performed, including:

[0077] Using a monthly time granularity, BeiDou GNSS data, SAR image data, and UAV data are aligned to the same time axis to achieve frequency matching;

[0078] Quantitative associations among multi-source data were calculated using Pearson correlation coefficients and regression models.

[0079] The reliability of the correlation is verified by reserving data and manual testing. A report containing a monthly data comparison table, correlation coefficients and prediction models is output. Dynamic early warning thresholds are set based on the correlation results to complete the time-series coupling and fusion of multi-source data.

[0080] Specifically, in step S3, multi-source features are extracted, including:

[0081] 2DCNN was used to extract deformation intensity distribution and surface rupture semantic features from SAR images and digital orthophoto maps (DOM).

[0082] After converting the BeiDou GNSS time-series displacement data into a two-dimensional "time-displacement" map, CNN was used to extract the dynamic trend features of the displacement.

[0083] After converting the Digital Earth Model (DEM) and Digital Surface Model (DSM) into voxel grids or elevation maps, 3DCNN is used to extract three-dimensional terrain morphology features, and a fixed-dimensional feature vector is output uniformly.

[0084] Multi-source CNN features are concatenated and then input into a fusion convolutional layer to learn the cross-modal correlation of "InSAR deformation-BeiDou trend-UAV crack-LiDAR terrain".

[0085] Spatiotemporal CNNs are used to model the fused multi-source temporal features, enabling end-to-end collaborative prediction of multi-source heterogeneous data through CNNs.

[0086] Specifically, the construction and training method of the ConvNeXt-ViT landslide early warning model includes:

[0087] The UAV DOM, InSAR deformation field, and BeiDou time series data are preprocessed into a 9-channel tensor and input into the model. The 9 channels include 3 channels of DOM, 1 channel of InSAR deformation field, and 5 channels of BeiDou time series slice.

[0088] Local features of cracks and misalignments are extracted by local feature branches containing 3 layers of ConvNeXt blocks;

[0089] Simultaneously, the global spatial correlation of the landslide is captured through the ViT branch of the 6-layer Transformer encoder;

[0090] The local and global features are fused into a unified feature vector through a 4-head cross-attention mechanism;

[0091] The model is trained using a hybrid loss function, and the output results are a landslide risk level classification and a deformation rate regression for the next 15 days.

[0092] The training method also includes: loading pre-trained weights from the remote sensing dataset and training using a phased unfreezing strategy; improving model accuracy by combining data augmentation and pseudo-label enhancement; and optimizing the final early warning results through geological prior rule filtering and multi-model ensemble voting.

[0093] An early warning will be activated when the landslide risk level exceeds the level threshold or the future deformation rate exceeds the deformation threshold.

[0094] Example 2:

[0095] This embodiment provides an early AI identification technology for landslide disasters based on BeiDou unmanned aerial vehicles and remote sensing, which includes the following steps:

[0096] (1) Select L1 level single-view complex images from the interferometric wide-swath (IW) mode of the Gaofen-3 satellite (C-band), which must cover the monitoring area and a 5km buffer zone, with a time baseline of ≤15 days and an incident angle of 30°-45°. Remove unqualified data with a loss of lock ≥5% or strong noise >1%. Use images with high coherence as the main images, and perform fine registration using the SIFT+RANSAC algorithm (residual ≤0.5 pixels). Crop the images to 2560×2560 pixels by extending 500m outward from the monitoring boundary. Load the precise orbit file to correct the orbit (positioning error ≤15m). After normalizing the radiometrics by Gamma-MAP filtering, convert the gray values ​​to backscattering coefficients σ using calibration parameters. 0 An improved Lee filter was used for noise reduction.

[0097] (2) Organize the preprocessed SAR images and simultaneously acquire the precise orbit files of the Gaofen-3 satellite.

[0098] And high-precision DEM. Using the stacking-inSAR method, 30-40 effective interferometric pairs are selected based on short spatial baseline (vertical baseline |B⊥|≤100m), short temporal baseline (≤15 days), and high coherence (γ≥0.3), which can be achieved with the help of SARscape software.

[0099] For each valid pair, the initial interferogram is obtained by multiplying the primary and secondary images by their complex conjugates. The flat phase calculated according to the formula is then subtracted to generate a differential interferogram. The interferogram is then weighted according to the formula wi=|γi|² (the first...). i The weights of the amplitude interference pairs, where i : No. i The index of the amplitude interference pair (differential interferogram), i=1,2,…,N; i : No. i The complex coherence coefficient of the amplitude interference pair; γi: the first iThe coherence coefficient moduli (between 0 and 1) of the amplitude interferogram pairs are weighted and stacked, and the initial stacked interferogram is obtained by using the average phase formula. The phase is unwrapped using the minimum cost flow method, and areas with residuals > 0.5 rad are removed; the topographic phase calculated by DEM is subtracted, atmospheric errors are eliminated by spatiotemporal filtering, and finally, the phase is calculated according to d = λ・φ. corr / (4π)(λ=5.6cm)(d:Surface deformation (displacement) along the radar line of sight (LOS), λ:SAR radar operating wavelength. Here, λ=5.6 cm is given, corresponding to the C-band (Gaofen-3), φ corr The corrected phase generally refers to the "deformation phase" obtained after unwrapping, subtracting the topographic phase calculated by DEM, and reducing atmospheric errors and residual noise through spatiotemporal filtering. The LOS-direction deformation is converted and verified with BeiDou GNSS (deviation ≤ 2mm) to obtain the surface deformation information.

[0100] (3) A Beidou reference station shall be set up in the stable bedrock area outside the landslide-affected area. The site selection shall meet the following requirements: there shall be no obstructions within a 15° elevation angle around it, and it shall be far away from vibration interference sources. The reference station shall be equipped with a high-precision GNSS receiver that supports the Beidou-3 system, and the static positioning accuracy shall be ≤2mm.

[0101] Deploy 3-5 GNSS rover stations in high-risk landslide areas, with a spacing of 500-1000m between the rover stations to cover key deformation areas of the landslide. The rover station receiver model should be consistent with that of the base station to ensure uniform data acquisition parameters.

[0102] Data integrity checks were performed on the raw BeiDou satellite signal data received by the base station and rover. Data packet loss was investigated through the receiver log file, with a packet loss rate of less than 0.5%. Missing data segments were marked, and abnormal observations were removed using the 3σ criterion. The raw data in UBX format output by the receiver was converted to the industry-standard RINEX 3.03 format. Satellite clock bias was compensated based on BeiDou broadcast ephemeris, and the initial value of tropospheric delay was estimated and corrected using the Saastamoinen model. The data timestamps of the base station and rover were adjusted with BeiDou system time as the reference to ensure that the time difference between the two is ≤1ms.

[0103] The system loads the BeiDou precise orbit and clock error products (sampling interval 15min) provided by the international GNSS service to replace the broadcast ephemeris to reduce orbit errors. It uses dual-frequency observation data to construct an ionospheric delay error model to eliminate the influence of the first-order ionosphere. It uses a median filtering algorithm (filter window size set to 5 epochs) to filter the abnormal data caused by multipath effects. It completes the high-precision positioning calculation of the base station and the rover based on the RTK or PPK relative positioning model. Finally, it outputs the three-dimensional displacement data of the landslide body along the east, north, and vertical directions, with displacement accuracy reaching the millimeter level. The data output format is a minute-level displacement time series, which yields the millimeter-level deformation and displacement data of the landslide.

[0104] (4) Figure 2 As shown, based on the surface deformation information obtained from Stacking-InSAR, with "deformation rate > 5 mm / month and cumulative displacement > 20 mm" as thresholds, and combined with the contours of deformation gradient abrupt change areas and continuous high deformation areas, landslide risk areas are delineated through ArcGIS buffer (500 m) and overlay analysis, covering potential diffusion areas and key areas such as the leading and trailing edges.

[0105] Within the risk zone, deploy 3-5 BeiDou GNSS-calibrated centimeter-level ground-based monitoring points (GCPs) covering the four corner points, crack inflection points, and boundary feature points, with an adjacent spacing of ≤300m. Perform static observations of the GCPs for ≥2 hours, and combine this with calculations from the base station to obtain CGCS2000 coordinates, achieving a horizontal accuracy of ≤2cm and an elevation accuracy of ≤3cm, providing a benchmark for UAV data correction.

[0106] An industrial-grade UAV equipped with a 20-megapixel camera and LiDAR was used for observation. Flight altitude was calculated based on focal length, with ≥80% forward overlap and ≥60% lateral overlap. The LiDAR point cloud density was ≥50 points / m², and the scanning frequency was 100kHz. RTK positioning was enabled, and flight was conducted during clear weather with wind speeds <5m / s. Data was processed using Pix4Dmapper: optical images were stitched together using the SIFT algorithm and combined with GCP correction to generate a 5cm resolution DOM (planar accuracy ≤5cm); the LiDAR point cloud was denoised to generate a 5cm grid DEM / DSM. Cracks were extracted in ArcGIS using a grayscale gradient threshold (15-20) and their extent was defined using InSAR boundaries. Surfer was used for DEM cross-section analysis to obtain factors such as fault height and slope. Finally, ContextCapture was used to construct a 3D model, based on the DSM and using the DOM as texture, with a texture resolution of 5cm and geometric deviation ≤3cm, visually displaying crack trends and spatial morphology to support stability analysis and early warning.

[0107] (5) Using the BeiDou base station as the reference, select at least three high-level control points in the known CGCS2000 coordinate system. Calculate the seven parameters using the least squares method to convert the original WGS84 coordinates of the BeiDou rover to CGCS2000 coordinates. SAR images are converted to WGS84 via radar geometric model and then substituted into the seven parameters for conversion. UAV DOM / DEM are combined and corrected using GCP to achieve coordinate unification. In terms of time reference, batches of SAR and UAV data are converted to BDT with an 8-hour time difference. Verify BeiDou data to ensure time deviation ≤1ms and unify the timestamp format. Match the frequency using the last day of each month as the node: calculate the monthly average displacement using BeiDou (removing data with a missing rate >10%), obtain the monthly deformation value using InSAR weighted by coherence, and complete the monthly value using linear interpolation for quarterly UAV data.

[0108] Data correlation was analyzed using Pearson coefficient (target |r|≥0.75), and a regression model (R²≥0.85) was constructed (the input X of the regression model includes multi-source indicators such as LOS cumulative deformation obtained from InSAR inversion, GNSS vertical displacement, and horizontal displacement; the output Y of the regression model is the reference displacement, which is the cumulative displacement obtained from the most accurate BeiDou GNSS). The model was verified using 20% ​​reserved data (MAE≤2mm) and manual measurements (deviation≤5mm). Then, a classification threshold was set in combination with the geological characteristics of landslides. When the rainfall during the rainy season is >100mm / month, the high-risk threshold was lowered, and finally, the temporal coupling and fusion of multi-source data was completed.

[0109] (6) When using 2DCNN to extract the deformation intensity distribution features of SAR images and the semantic features of surface ruptures in UAV DOM, the input data needs to be standardized and preprocessed first: the deformation intensity map of the SAR image is normalized to the [0,1] interval by Min-Max and cropped into 256×256 pixel slices. The UAV DOM retains the surface texture at a resolution of 5cm and is also normalized and cropped to the same size. The 2DCNN network adopts a 3-layer convolutional structure. The first layer uses a 3×3 convolutional kernel (64 channels) to capture the deformation gradient of the SAR image and the edges of fine cracks in the DOM, and enhances the nonlinear expression through the GELU activation function. The second layer uses a 5×5 convolutional kernel (128 channels) to extract the spatial clustering features of the high deformation area of ​​the SAR (such as the outline of the continuous deformation zone) and the connectivity features of cracks in the DOM (such as the topological relationship between the main crack and the branch crack). The third layer uses a 1×1 convolutional kernel (256 channels) to compress the feature dimension, outputting two types of 256-dimensional feature vectors, corresponding to the spatial distribution pattern of deformation intensity and the semantic attributes of surface rupture (such as crack length, orientation, and density). When converting BeiDou GNSS time-series displacement data into a "time-displacement" two-dimensional map, the horizontal axis represents the time series (30 days), and the vertical axis represents the displacement value (normalized to [-1,1]). The displacement changes are converted into a 256×256 pixel heatmap through grayscale mapping. The CNN feature extraction for this map adopts a 4-layer convolutional structure: the first two layers of 3×3 convolutional kernels (64→128 channels) capture short-term displacement fluctuations (such as single-day sudden changes), the third layer of 7×7 convolutional kernels (256 channels) identifies medium-term trends, and the fourth layer of 1×1 convolutional kernels compresses the data into 256-dimensional feature vectors, focusing on extracting key nodes in the displacement dynamics (such as the time of the first acceleration and the time when the maximum displacement value occurs) and trend types (linear growth and step-like leaps). When processing LiDAR point clouds, the 3D point cloud (coordinate range X / Y / Z in CGCS2000 coordinate system) is first converted into a 32×32×32 voxel grid (voxel resolution 0.5m×0.5m×0.5m) through voxelization, with each voxel recording the point cloud density and average elevation; or projected as a 256×256 pixel elevation map (Z value normalized to [0,1]). The 3DCNN uses a 5-layer convolutional structure. The first three layers use 3×3×3 convolutional kernels (64→128→256 channels) to extract the 3D terrain undulation features in the voxel grid (such as the steep slope at the rear edge of the landslide and the bulge at the front edge). The last two layers use 1×1×1 convolutional kernels to compress the feature vector into a 256-dimensional feature vector, focusing on capturing the rate of change of terrain slope, the 3D misalignment height of cracks (vertical difference), and the overall spatial morphology of the landslide body (such as whether it is distributed in a stepped manner).

[0110] The feature vectors of SAR (256-dimensional), DOM (256-dimensional), BeiDou (256-dimensional), and LiDAR (256-dimensional) are concatenated into a 1024-dimensional fused feature, which is then input into a fusion convolutional layer. Cross-modal feature interaction (such as associating the spatial correspondence between "high deformation area" and "dense crack area") is achieved through two 1×1 convolutional kernels (512→256 channels), while compressing the dimension to 256. Subsequently, a spatiotemporal CNN (ConvLSTM) module is connected, with three time steps (corresponding to feature data for three consecutive months), a hidden layer dimension of 256, and a gating mechanism to capture the dynamic correlation of temporal features (such as whether the acceleration of BeiDou displacement is accompanied by the expansion of SAR deformation range and the continuous widening of cracks). Finally, the end-to-end prediction results are output through a fully connected layer—including the deformation rate (continuous value) and risk level (0-3 classification) for the next 15 days, realizing deep collaboration of multi-source heterogeneous data at the feature layer and decision layer.

[0111] (7) By converting the heterogeneous data style of the target area through CycleGAN, pseudo data with consistent style is generated to eliminate systematic bias of the equipment for the imaging differences of different sensors (such as SAR, UAV camera, LiDAR). Simultaneously, histogram matching is used to adjust the gray distribution of the target area data so that it is aligned with the brightness and contrast features of the source domain data. Attention mechanism is used to focus on the core area of ​​the landslide (cracks, faults, deformation zones) to enhance the matching accuracy of these key textures and ensure that the structural features of the core area are not distorted after conversion.

[0112] The spatial branch uses a 3-layer CNN (3×3 / 5×5 convolution kernel, number of channels 64→128→256) to capture local common features of landslides through a sliding window, such as the linear morphology of cracks, the vertical height difference of faults, and the spatial clustering pattern of high deformation areas.

[0113] The frequency branch first decomposes the data into low-frequency (overall trend) and high-frequency (detail noise) components using wavelet transform. Then, a 6-layer Transformer encoder (16×16 Patch, 8-head attention) captures the global distribution pattern while filtering out region-specific noise (such as local vegetation interference and sensor random noise). The dual-branch features are fused through 4-head cross-attention to establish a correlation mapping between "local details and global distribution."

[0114] The maximum mean difference (MMD) loss function is introduced to calculate and minimize the distance between the feature distributions of the source domain and the target domain, which encourages the model to learn domain-invariant features (such as the “rear edge tensile cracking-middle shearing-front edge bulging” structural pattern common to all landslides).

[0115] During the model adaptation phase, the bottom feature extraction layer and the middle fusion layer are frozen (preserving the general features learned from the source domain), while only the top discriminant layer is unfrozen. Training is performed using 10%-20% of the labeled data from the target region, with a region-sensitive attention mechanism introduced to focus on learning the unique characteristics of the target region (such as vertical cracks in loess landslides and bedding displacement in rock landslides). Finally, the model is tested in three or more different regions (such as the southwestern mountainous areas, the Loess Plateau, and the southeastern hills). By controlling the F1 score fluctuation to ≤10% and the core feature (crack boundary, deformation center) recognition deviation to ≤3 pixels (corresponding to a 15cm accuracy for UAV DOM), the model's cross-regional adaptability is ensured, achieving stable recognition of landslide features in different geological environments.

[0116] (8) The ConvNeXt-ViT landslide early warning model was adopted, and multi-source data preprocessing was completed in advance to construct a 9-channel tensor input: the 3-channel UAV DOM was directly used as the first 3 channels. The InSAR deformation field was normalized to [0,1] by Min-Max to form the 4th channel. The Beidou time-series displacement data was divided into 5 subsequences of 6 days / slices according to the time window. The mean of each slice was calculated and normalized to form the 5th to 9th channels. All channel data were uniformly resized to 224×224 pixels and then normalized by Z-score to ensure that the input data distribution was consistent. The model feature extraction stage adopts a "local + global" dual-path architecture: a 3-layer ConvNeXt block constitutes the local feature extraction branch, with each layer using a 7×7 depthwise separable convolution (channel counts successively 64→128→256), coupled with the GELU activation function and LayerNorm layers, focusing on capturing local details such as crack edges in the DOM and clustering of high-deformation areas in InSAR; a 6-layer Transformer encoder forms the global feature extraction branch, first dividing the 9-channel tensor into 16×16 patches (196 in total), converting them into 768-dimensional vectors through a linear embedding layer, and then extracting cross-channel global spatial correlations through an 8-head self-attention mechanism and a Feed-Forward network. The dual-branch features are fused through 4-head cross-attention, mapping the local 256-dimensional features and the global 768-dimensional features to the same dimension and concatenating them to output a 256-dimensional fused feature vector. This vector comprehensively encodes three types of information in the monitoring area: (1) the surface morphology and texture features of the UAV DOM, such as cracks, landslide embankments and accumulation boundaries; (2) the displacement level, deformation gradient and accelerated evolution features in the past 30 days observed by InSAR and BeiDou; and (3) the consistency and spatial-temporal coupling relationship between morphology-deformation and multi-source observations. The 256-dimensional feature serves as a scene-level landslide state representation and is input into the subsequent classification head to complete the landslide risk level prediction.

[0117] The model training uses a hybrid loss function (0.6× cross-entropy loss + 0.4×MAE loss): cross-entropy loss is used to optimize the classification task of landslide risk level (0-3), and MAE loss is used to improve the regression accuracy of deformation rate (continuous value) in the next 15 days.

[0118] During training, pre-trained weights for remote sensing image classification (e.g., pre-trained on the SEN12MS dataset) are loaded. A phased unfreezing strategy is adopted—only the top fully connected layer is trained in the first 5 rounds, the top 2 layers of the Transformer encoder are unfrozen in rounds 6-10, and all layers are unfrozen from round 11 onwards to avoid overfitting. At the same time, data augmentation (random flipping, rotation, and Gaussian noise addition) and pseudo-label augmentation (after predicting unlabeled samples, pseudo-labels with a confidence of ≥0.9 are selected to participate in training) are introduced to further improve the model's generalization ability.

[0119] During the model optimization phase, geological prior rules were incorporated for filtering: samples predicted as high-risk but located in stable bedrock areas (slope < 15°) were marked as false alarms. A multi-model ensemble voting system was constructed using ConvNeXt-ViT, SCDUNet++, and SCGC-Net, determining the final risk level according to the "majority rule" principle. The output included landslide risk levels (Level 0: No risk; Level 1: Low risk; Level 2: Medium risk; Level 3: High risk) and the average daily deformation rate (mm / day) for the next 15 days. Validation on the test set showed a classification accuracy ≥ 92% and a deformation rate prediction MAE ≤ 1.2 mm / day, meeting the accuracy requirements for landslide early warning.

[0120] (9) The hardware architecture adopts a four-layer architecture of "acquisition-transmission-processing-publishing": The data acquisition layer is equipped with Beidou reference stations, industrial-grade UAVs, automatic weather stations and SAR receiving terminals; the transmission layer adopts 4G / 5G + Tiantong-1 satellite dual links to ensure uninterrupted data in mountainous areas without network coverage; the processing layer deploys edge computing boxes and GPU servers (NVIDIA A100, supporting real-time inference of ConvNeXt-ViT models); and the publishing layer is equipped with sound and light warning piles, outdoor LED screens and mobile APP. The software system is constructed with four core modules: The data middle platform adopts PostgreSQL + PostGIS database to uniformly store multi-source data (Beidou time series, InSAR deformation, UAV DOM / DEM, meteorological data, etc.). The system automatically performs standardized preprocessing (coordinate transformation to CGCS2000, time synchronization to BDT); the model service module encapsulates the ConvNeXt-ViT early warning model as a RESTful API, supporting single real-time inference (response time ≤ 2s) and batch data processing (average processing of 100GB of data per day); the early warning decision module sets dual trigger conditions, initiating an early warning when the risk level is ≥ 2 or the deformation rate is > 15mm / month, while loading geological prior rules (such as automatically eliminating false alarms in stable areas with a slope < 15°); the visualization feedback module is based on the WebGIS platform, overlaying and displaying multi-source data layers (deformation heat map, crack distribution, early warning range) and historical early warning logs, supporting spatiotemporal queries and trend analysis.

[0121] The system executes a closed-loop process of "data acquisition → edge preprocessing → cloud model inference → decision judgment → multi-terminal early warning within 10 minutes → feedback review": the acquisition layer uploads data once an hour, the edge computing box pushes the data to the cloud after initial cleaning, the GPU server calls the model inference to generate risk level and deformation rate prediction, the decision module verifies and triggers an early warning, each issuing terminal responds synchronously, and the accuracy of the early warning is analyzed afterward through the feedback module.

[0122] The optimization process employs a dynamic threshold strategy (lowering the deformation rate warning threshold to 10mm / month when rainfall exceeds 100mm / month in the rainy season, and raising it to 20mm / month in the dry season), and uses multi-model weighted voting (ConvNeXt-ViT weight 0.5, SCDUNet++ and SCGC-Net weights 0.25 each) to improve prediction robustness. On the hardware side, rain and lightning protection covers are added, and backup power supplies are configured to cope with extreme weather. On the software side, LSTM interpolation is used to complete missing data. The user interface is simplified, and three user permission categories—administrators, monitoring personnel, and the general public—are designed to adapt to different usage scenarios, ultimately achieving a landslide early warning system that combines high accuracy, high reliability, and ease of use.

[0123] In the method of this embodiment,

[0124] Method starting point: Acquire raw multi-source data (SAR imagery, BeiDou GNSS signals, UAV observation data) of the monitoring area.

[0125] Method endpoint: Output landslide risk level and deformation rate prediction, and complete the closed-loop early warning release.

[0126] Core innovations: multi-source data collaborative acquisition, unified spatiotemporal fusion, cross-regional adaptive AI model, and integrated hardware and software early warning system.

[0127] like Figure 1 As shown, the detailed method steps of the present invention include:

[0128] S1: Multi-source data acquisition and preprocessing stage

[0129] The goal of this phase is to acquire and preprocess SAR imagery, BeiDou GNSS data, and UAV data to provide high-quality input for subsequent fusion analysis.

[0130] S1.1SAR Image Data Acquisition and Preprocessing

[0131] Main component: SAR data processing system (such as SARscape software).

[0132] action:

[0133] 1. Acquisition: Acquire L1 level single-view complex images of the Gaofen-3 satellite (C-band) in interferometric wide swath mode (IW) covering the monitoring area and a 5km buffer zone, with a time baseline of ≤30 days and an incident angle of 30°-45°.

[0134] 2. Screening: Remove unqualified images with a data loss rate ≥5% or strong noise >1%.

[0135] 3. Registration: The image with the highest coherence is used as the main image, and the SIFT+RANSAC algorithm is used for fine registration, with a registration residual of ≤0.5 pixels.

[0136] 4. Cropping: Expand the monitoring boundary by 500m and crop the image to 2560×2560 pixels.

[0137] 5. Calibration: Load the precise track file for track calibration (positioning error ≤ 15m); use Gamma-MAP filtering for radiometric normalization, and convert the grayscale values ​​into backscattering coefficients σ using calibration parameters. 0 .

[0138] 6. Noise Reduction: An improved Lee filtering algorithm is used for noise reduction.

[0139] Object: Raw SAR image.

[0140] Result: High-quality SAR images were obtained after preprocessing.

[0141] S1.2 BeiDou GNSS Data Acquisition and Preprocessing

[0142] Main components: High-precision GNSS receiver (base station and rover), data preprocessing module.

[0143] action:

[0144] 1. Deployment:

[0145] One Beidou reference station is deployed in the stable bedrock area outside the landslide area (site selection requirements: 15° elevation angle without obstruction, far away from vibration source), with a static positioning accuracy of ≤2mm.

[0146] In high-risk landslide areas (preferably areas showing large deformation rates or recent significant acceleration in InSAR and existing GNSS monitoring, or areas with unfavorable geological conditions such as developed surface fissures, collapse scarps, bedding weak interlayers, and abundant groundwater), 3-5 GNSS mobile stations should be deployed evenly along the direction of the main landslide fissure, with at least one station located in the rear tensile fracture zone and one station located in the front shear zone, with a station spacing of 500-1000m.

[0147] 2. Data Acquisition: The base station and the rover simultaneously receive BeiDou satellite signals.

[0148] 3. Pretreatment:

[0149] Integrity check: Check the data loss rate, which should be less than 0.5%, and mark any missing data segments.

[0150] Outlier removal: Outlier observations are removed using the 3σ criterion.

[0151] Format conversion: Convert raw UBX format data to RINEX 3.03 format.

[0152] Preliminary correction: Satellite clock bias was compensated based on broadcast ephemeris; the initial value of tropospheric delay was estimated and corrected using the Saastamoinen model.

[0153] Time synchronization: Using the BeiDou system time as a reference, adjust the data timestamp to ensure that the time deviation between the base station and the rover is ≤1ms.

[0154] 4. Precise positioning calculation:

[0155] Load the BeiDou precision orbit and clock difference products provided by IGS (sampling interval 15min).

[0156] An ionospheric delay error model was constructed using dual-frequency observation data to eliminate the influence of the first-order ionosphere.

[0157] The median filtering algorithm (with a window size of 5 epochs) is used to filter out outlier data caused by multipath effects.

[0158] Based on RTK or PPK relative positioning models, high-precision positioning calculations are completed, and three-dimensional displacement data of the landslide body in the east, north, and vertical directions are output with millimeter-level accuracy. The data is presented as a minute-level displacement time series.

[0159] Object: Raw signals from BeiDou satellites.

[0160] Results: Clean, compatible, and synchronized millimeter-level deformation and displacement data of landslides were obtained.

[0161] S1.3 UAV Data Acquisition and Preprocessing

[0162] Main components: Industrial-grade drone (equipped with a 20-megapixel optical camera and LiDAR), and data processing software (such as Pix4Dmapper).

[0163] action:

[0164] 1. Risk zone delineation: Based on the InSAR deformation information obtained from S1.1, landslide risk zones are delineated using the thresholds of "deformation rate > 5 mm / month and cumulative displacement > 20 mm" and combined with the deformation gradient abrupt change zone, through ArcGIS buffer (500 m) analysis.

[0165] 2. GCP Deployment: Deploy 3-5 BeiDou GNSS-calibrated centimeter-level ground control points (GCPs) within the risk area, covering the four corner points, crack inflection points, and boundary feature points, with an adjacent spacing of ≤300m. Perform static observation of the GCPs for ≥2 hours to obtain CGCS2000 coordinates (planar accuracy ≤2cm, elevation accuracy ≤3cm).

[0166] 3. Flight observation: Under clear weather and wind speed <5m / s conditions, the UAV flies along a preset route (heading overlap ≥80%, lateral overlap ≥60%), LiDAR point cloud density ≥50 points / ㎡, scanning frequency 100kHz, and RTK positioning is enabled.

[0167] 4. Data Processing:

[0168] DOM generation: Optical images are stitched together using the SIFT algorithm and geometrically corrected using GCP to generate a 5cm resolution digital orthophoto map (DOM) with a planar accuracy of ≤5cm.

[0169] DEM / DSM generation: After denoising LiDAR point clouds, 5cm grid digital terrain model (DEM) and digital surface model (DSM) are generated.

[0170] 5. Feature extraction: In ArcGIS, crack distribution is extracted using a grayscale gradient threshold (15-20); the landslide range is determined by combining InSAR boundaries; and DEM cross-section analysis is performed using Surfer to obtain information such as fault height and slope.

[0171] 6. 3D Modeling: ContextCapture was used to construct a 3D virtual reality model of the landslide, based on DSM and using DOM as texture, with a texture resolution of 5cm and a geometric deviation of ≤3cm.

[0172] Target: Landslide risk area.

[0173] Results: DOM, DEM, DSM and three-dimensional landslide model were obtained, including micro-fracture features.

[0174] S2: Multi-source data spatiotemporal unification and fusion stage

[0175] The goal of this phase is to unify the coordinate system and time reference of multi-source data and achieve temporal coupling and fusion.

[0176] S2.1 Coordinate System 1

[0177] Main body: Coordinate transformation module.

[0178] Action: Using the BeiDou base station as a reference, select at least three high-level control points in the known CGCS2000 coordinate system and calculate the seven parameters using the least squares method. Then, unify the original WGS84 coordinates of the BeiDou rover, SAR imagery (converted to WGS84 via radar geometric model), and UAV DOM / DEM data to the CGCS2000 coordinate system using the seven-parameter transformation.

[0179] Targets: BeiDou mobile station data, SAR imagery data, and UAV data.

[0180] Result: All data coordinate systems are consistent with CGCS2000.

[0181] S2.2 Time Base Unification

[0182] Main component: Time synchronization module.

[0183] Action: Convert the time base of all data from Coordinated Universal Time (UTC) to BeiDou System Time (BDT). Specifically, add an 8-hour time difference to SAR and UAV data, verify BeiDou data to ensure time deviation ≤1ms, and standardize the timestamp format.

[0184] Object: Timestamps of all source data.

[0185] Result: All data time bases were unified to BDT.

[0186] S2.3 Frequency Matching and Timing Coupling Fusion

[0187] Main component: Data fusion engine.

[0188] action:

[0189] 1. Frequency matching: Using the last day of each month as the time node, bind BeiDou high-frequency data (calculate the monthly average displacement and remove data with a missing rate >10%), InSAR low-frequency data (calculate the monthly deformation value by coherence weighting) and UAV quarterly data (use linear interpolation to complete the monthly value) to the same time axis.

[0190] 2. Correlation analysis: Plot multi-source time series curves and space-time heatmaps; use Pearson correlation coefficient (target |r|≥0.75) to analyze the quantitative correlation between "BeiDou displacement growth rate" and "UAV crack widening"; establish a regression model (requires R²≥0.85).

[0191] 3. Verification and threshold setting: Verify the reliability of the correlation using 20% ​​of the reserved data (MAE ≤ 2mm) and manually measured data (deviation ≤ 5mm); based on the correlation results and landslide geological characteristics, set dynamic early warning thresholds (e.g., Beidou displacement growth rate > 15mm / month corresponds to the crack reaching the warning value after 1 month; when the rainfall during the rainy season is > 100mm / month, the threshold is lowered to 10mm / month).

[0192] Object: Multi-source time series data after unifying the spatiotemporal reference.

[0193] Results: The temporal coupling and fusion of multi-source data was completed, and the correlation model and dynamic early warning threshold were output.

[0194] S3: Multi-source feature extraction and collaborative prediction stage

[0195] The goal of this phase is to use deep learning models to extract multi-source features and perform end-to-end collaborative prediction of landslide risks.

[0196] S3.1 Multi-source CNN Feature Extraction

[0197] Main body: Feature extraction module (2DCNN, 3DCNN).

[0198] action:

[0199] 1. SAR Feature Extraction: The InSAR deformation intensity map is normalized to [0,1] using Min-Max and cropped to 256×256 pixels. A 3-layer 2DCNN (3×3 / 5×5 / 1×1 convolutional kernels, 64→128→256 channels, GELU activation function) is used to extract the deformation intensity distribution features, outputting a 256-dimensional feature vector.

[0200] 2. UAV DOM Feature Extraction: The DOM (5cm resolution) is normalized and cropped to 256×256 pixels. The same 2DCNN structure as SAR is used to extract semantic features of surface ruptures (such as crack length, direction, and density), and a 256-dimensional feature vector is output.

[0201] 3. BeiDou GNSS Feature Extraction: The BeiDou time-series displacement data is converted into a two-dimensional heatmap of "time-displacement" (256×256 pixels, displacement values ​​normalized to [-1,1]). A 4-layer CNN (convolutional kernel 3×3 / 3×3 / 7×7 / 1×1, number of channels 64→128→256→256) is used to extract the dynamic trend features of displacement, and output a 256-dimensional feature vector.

[0202] 4. LiDAR Feature Extraction: The LiDAR point cloud is converted into a 32×32×32 voxel grid (0.5m³ resolution) or a 256×256 pixel elevation map. A 5-layer 3DCNN (3×3×3 convolutional kernels, 64→128→256 channels, followed by 1×1×1 convolutional compression) is used to extract 3D terrain morphology features, outputting a 256-dimensional feature vector.

[0203] Object: Preprocessed source data.

[0204] Results: One 256-dimensional feature vector each of SAR, DOM, BeiDou, and LiDAR was obtained.

[0205] S3.2 Feature Fusion and Spatiotemporal Prediction

[0206] Main components: Fusion convolutional layers and spatiotemporal CNN (ConvLSTM).

[0207] action:

[0208] 1. Feature concatenation: The four 256-dimensional feature vectors obtained in S3.1 are concatenated into a 1024-dimensional fused feature vector.

[0209] 2. Cross-modal fusion: The stitched features are input into the fusion convolutional layer (2 layers of 1×1 convolution, number of channels 512→256) to learn the cross-modal correlation of "InSAR deformation-BeiDou trend-UAV crack-LiDAR terrain" and output 256-dimensional fusion features.

[0210] 3. Temporal modeling: Input the fused features into the ConvLSTM module (set 3 time steps, corresponding to 3 consecutive months of data, hidden layer dimension 256) to capture the dynamic correlation of time series.

[0211] 4. Collaborative prediction: The end-to-end prediction results are output through the fully connected layer, including the deformation rate (continuous value) and landslide risk level (0-3 classification) for the next 15 days.

[0212] Object: Multi-source CNN feature vectors.

[0213] Results: Preliminary predictions of landslide risk level and deformation rate were obtained.

[0214] S4: Model Cross-Regional Adaptation and Optimization Stage

[0215] The goal of this phase is to enhance the adaptability of the model to different regions through the SSFSC framework and to use advanced models for accurate early warning.

[0216] S4.1 Domain Adaptation Based on the SSFSC Framework

[0217] Main body: SSFSC framework (including CycleGAN, CNN, Transformer, etc.).

[0218] action:

[0219] 1. Style Consistency: CycleGAN is used to transform the style of heterogeneous data in the target area to eliminate sensor differences; histogram matching and attention mechanisms are combined to align the texture of the core area of ​​the landslide (cracks, faults).

[0220] 2. Dual-branch feature extraction:

[0221] Spatial branching: A 3-layer CNN (3×3 / 5×5 convolutional kernels, 64→128→256 channels) was used to extract local common features of landslides.

[0222] Frequency branch: The data is first subjected to wavelet transform, and then a 6-layer Transformer encoder (16×16Patch, 8-head attention) is used to capture the global distribution pattern and filter out regional noise.

[0223] 3. Feature fusion and alignment: Dual-branch features are fused through 4-head cross-attention; source and target domain features are aligned using the maximum mean difference (MMD) loss function, and domain-invariant features are learned.

[0224] 4. Fine-tuning and Testing: Freeze the bottom and middle layers of the framework, and only fine-tune the top discriminant layer; train using 10%-20% of the labeled data of the target region, and introduce a region-sensitive attention mechanism. Test in 3 or more regions with different characteristics (such as the southwestern mountainous area, the Loess Plateau, and the southeastern hilly area) to ensure that the F1 score fluctuation is ≤10% and the recognition deviation of core features (crack boundaries, deformation centers) is ≤3 pixels.

[0225] Object: Multi-source data of the target region.

[0226] Results: A landslide identification model with enhanced cross-regional adaptability was obtained.

[0227] S4.2 Training and Optimization of ConvNeXt-ViT Landslide Early Warning Model

[0228] Main component: Model training system (such as GPU server).

[0229] action:

[0230] 1. Data preprocessing: The UAV DOM (3 channels), InSAR deformation field (1 channel, normalized), and BeiDou time series data (5 channels, averaged and normalized by 6-day slices) are constructed into a 9-channel tensor, uniformly resized to 224×224 pixels, and Z-score normalized.

[0231] 2. Model Architecture:

[0232] Local feature branches: 3-layer ConvNeXt blocks (7×7 depthwise separable convolution, number of channels 64→128→256, GELU activation, LayerNorm).

[0233] Global feature branch: 6-layer Transformer encoder (16×16Patch, 8-head self-attention, embedding dimension 768).

[0234] Feature fusion: Local (256-dimensional) and global (768-dimensional) features are fused into a 256-dimensional feature vector through 4-head cross attention.

[0235] 3. Model Training:

[0236] Loss function: Mixed loss (0.6 × cross-entropy loss + 0.4 × MAE loss).

[0237] Training strategy: Load remote sensing dataset (such as SEN12MS) pre-train weights; unfreeze in stages (rounds 1-5: train only the output layer; rounds 6-10: unfreeze the top 2 layers of Transformer; round 11 onwards: unfreeze all layers); reduce the learning rate from 1e-4 to 1e-5.

[0238] Data augmentation: Random flipping, cropping, brightness perturbation, and adding Gaussian noise are employed.

[0239] Pseudo-label augmentation: For predictions of unlabeled samples, pseudo-labels with a confidence level > 0.9 are selected and added to the training.

[0240] 4. Model optimization:

[0241] Geological prior filtering: Remove samples that are predicted to be high-risk but are located in stable bedrock areas (slope ≤ 15°).

[0242] Multi-model ensemble: ConvNeXt-ViT (weight 0.5), SCDUNet++ (weight 0.25), and SCGC-Net (weight 0.25) are used for weighted voting to determine the final risk level.

[0243] Object: 9-channel multi-source data tensor.

[0244] Results: An optimized landslide early warning model was obtained, which outputs the risk level (0-3) and the deformation rate (mm / day) for the next 15 days. The classification accuracy is required to be ≥92% and the deformation rate prediction MAE is ≤1.2mm / day.

[0245] S5: Landslide Early Warning System Establishment and Implementation Phase

[0246] The goal of this phase is to build an integrated hardware and software early warning system and execute a closed-loop early warning process.

[0247] S5.1 Hardware Architecture Setup

[0248] Subject: System deployment personnel.

[0249] action:

[0250] 1. Data Acquisition Layer: Deploy BeiDou reference stations (accuracy ≤ 2mm), industrial-grade drones (optical cameras + LiDAR), automatic weather stations, and SAR receiving terminals.

[0251] 2. Transmission Layer: Employs dual-link transmission using 4G / 5G + Tiantong-1 satellite. When 4G / 5G signals are interrupted, the system automatically switches to the BeiDou short message satellite link, with a single data transmission latency of ≤60s.

[0252] 3. Processing layer: Deploy edge computing boxes (preprocessing time for BeiDou data ≤ 5 minutes) and GPU servers (such as NVIDIA A100).

[0253] 4. Release layer: Configure sound and light warning piles, outdoor LED screens, and mobile APP.

[0254] Object: Hardware device.

[0255] Result: The hardware architecture deployment of "collection-transmission-processing-publishing" was completed.

[0256] S5.2 Software System Construction

[0257] action:

[0258] 1. Data Platform: Based on PostgreSQL+PostGIS database, it uniformly stores multi-source data and automatically completes standardized preprocessing (coordinate transformation to CGCS2000, time synchronization to BDT).

[0259] 2. Model Service Module: The ConvNeXt-ViT model trained in S4.2 is encapsulated as a RESTful API, supporting real-time inference (response time ≤ 2s) and batch processing (100GB of data processed per day).

[0260] 3. Early warning decision module: Set dual trigger conditions (risk level ≥ 2 or deformation rate > 15 mm / month); load geological prior rules (such as automatically eliminating false alarms in stable areas with slope ≤ 15°).

[0261] 4. Visual feedback module: Based on the WebGIS platform, it overlays and displays deformation heat maps, crack distribution, early warning range and historical early warning logs, and supports spatiotemporal query and trend analysis.

[0262] Object: Software module.

[0263] Result: The software system development and integration were completed.

[0264] S5.3 Closed-Loop Early Warning Process Execution

[0265] Main component: Early warning system.

[0266] Action: Execute a closed-loop process of "data acquisition → edge preprocessing → cloud model inference → decision-making → multi-terminal early warning within 10 minutes → feedback and review":

[0267] 1. Data Acquisition: The acquisition layer uploads data once per hour.

[0268] 2. Edge preprocessing: The edge computing box completes the initial data cleaning.

[0269] 3. Cloud-based model inference: The GPU server calls the model API to generate risk level and deformation rate predictions.

[0270] 4. Decision-making: The early warning decision module verifies the results. If the triggering conditions are met, an early warning is activated.

[0271] 5. Warning Issuance: Warning information will be issued simultaneously within 10 minutes through sound and light warning posts, LED screens, and mobile APP.

[0272] 6. Feedback and Review: Analyze the accuracy of early warnings through the visual feedback module for system optimization.

[0273] Object: Real-time monitoring data.

[0274] Results: Early identification and accurate warning of landslide disasters were achieved, and closed-loop management was completed.

[0275] Method endpoint

[0276] The system outputs landslide risk levels (Level 0: No risk; Level 1: Low risk; Level 2: Medium risk; Level 3: High risk) and a 15-day deformation rate prediction, and disseminates early warning information through multiple terminals. The system continuously optimizes through dynamic thresholds and multi-model voting to ensure a false alarm rate of <10% and a false negative rate of <5%. Example 3

[0277] This embodiment aims to describe in detail the hardware and software system architecture for implementing the method of the present invention. The landslide disaster early identification system adopts a layered design, including a data acquisition layer, a data transmission layer, a data processing layer, and an early warning release layer connected sequentially via wired and wireless communication links.

[0278] 1. Data Acquisition Layer

[0279] This layer is responsible for the automated collection of multi-source data and is the sensing edge of the system.

[0280] The BeiDou / GNSS monitoring subsystem includes a BeiDou base station (using a high-precision receiver, with a static positioning accuracy ≤2mm) deployed in the stable bedrock area outside the landslide zone and 3-5 GNSS rover stations deployed in the high-risk landslide area. The base station and rover stations form a monitoring network via a wireless network to continuously collect millimeter-level displacement data.

[0281] The UAV remote sensing subsystem includes an industrial-grade UAV platform and its onboard optical camera and LiDAR equipment. Based on flight missions issued by the data processing layer, this subsystem automatically performs flight operations within a designated risk area.

[0282] SAR data receiving terminal: used to receive and buffer raw image data from SAR satellites such as Gaofen-3.

[0283] Auxiliary weather stations (optional but recommended): collect on-site environmental parameters such as rainfall and temperature to provide auxiliary information for early warning decision-making.

[0284] 2. Data transmission layer

[0285] This layer is responsible for reliably and in real-time transmitting the data from the acquisition layer to the processing layer.

[0286] Core component: A dual-link transmission gateway consisting of a 4G / 5G communication module and a satellite communication module (specifically, a BeiDou short message communication module or a TianTong-1 satellite module).

[0287] Workflow: The system prioritizes high-speed, low-cost 4G / 5G networks for data transmission by default. When the system detects that the 4G / 5G signal strength is below a preset threshold or is completely interrupted, the link management unit will automatically and seamlessly switch to the satellite communication link to ensure that data can still be transmitted in harsh environments without terrestrial network coverage, with a single data transmission latency of ≤60 seconds.

[0288] 3. Data Processing Layer

[0289] This layer is the "brain" of the system, responsible for data aggregation, processing, model inference, and intelligent decision-making. It consists of hardware infrastructure and upper-layer software systems.

[0290] Hardware infrastructure:

[0291] Edge computing box: Deployed near the landslide monitoring site. Its main function is to perform real-time preprocessing of raw data collected by the BeiDou / GNSS subsystem, including data integrity checks, gross error removal, and format conversion, and strictly control the preprocessing time to within 5 minutes, greatly reducing the pressure on cloud transmission and processing.

[0292] GPU servers: Deployed in cloud data centers. They are the core computing power units, specifically designed to run computationally intensive artificial intelligence models, namely the ConvNeXt-ViT landslide early warning model, to complete feature extraction, fusion, and risk prediction of multi-source data.

[0293] Software system:

[0294] Data Platform Module: Built on a spatial database. It acts as a unified data hub, receiving and storing all multi-source heterogeneous data from the transport layer. This module incorporates preprocessing logic, automatically performing coordinate unification (to CGCS2000) and time synchronization (to BDT), providing clean and standardized data services for upper-layer applications.

[0295] Model Service Module: Adopting a microservice architecture, this module encapsulates the trained ConvNeXt-ViT early warning model into a RESTful API. It receives standardized data pushed from the data platform, calls the underlying GPU server for real-time inference (response time ≤ 2 seconds) or batch data processing, and passes the prediction results (risk level, deformation rate) to the early warning decision module.

[0296] Early Warning Decision Module: This module loads preset dual trigger conditions (risk level ≥ 2 or deformation rate > 15 mm / month). Upon receiving the prediction results from the model service, it automatically judges based on these conditions and verifies the results by combining them with geological prior rules (such as automatically filtering false alarms in areas with slope < 15°), and finally generates an instruction on whether to issue an early warning.

[0297] Visual feedback module: Developed based on WebGIS technology. It provides administrators with an interactive interface that dynamically overlays and displays InSAR deformation heatmaps, UAV orthophotos, crack distribution, real-time warning areas, and historical warning records. Simultaneously, it provides a warning feedback entry point for post-event analysis of warning accuracy, achieving closed-loop review.

[0298] 4. Early Warning Issuance Layer

[0299] This layer is responsible for quickly delivering early warning information to target users in various forms.

[0300] On-site audible and visual early warning posts: These are deployed at key intersections within the landslide-affected area. Upon receiving an early warning command, they emit a strong audible and visual alarm signal, directly alerting personnel on site.

[0301] Outdoor LED displays: Deployed in public places such as communities and roadsides, they scroll through text and graphics to display warning levels, affected areas, and evacuation tips.

[0302] Mobile App: Designed for administrators and residents within the region. It pushes precise early warning notifications, risk maps, and emergency guidelines to ensure information reaches individuals directly.

[0303] System workflow (closed loop):

[0304] The system automatically executes the following closed-loop process according to a preset period (e.g., every hour) or trigger conditions:

[0305] Data acquisition → data transmission → edge preprocessing → cloud model inference → decision-making → multi-terminal early warning within 10 minutes → feedback and review.

[0306] This system embodiment, through deep integration and collaborative work of software and hardware, solidifies the method and process of the present invention into an efficient, reliable, and operational entity system, realizing the full automation and intelligence of the landslide disaster from "perception-cognition-decision-action".

[0307] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0308] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0309] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0310] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0311] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for early identification of landslide disaster based on Beidou unmanned aerial vehicle and remote sensing, characterized in that, include: Acquire multi-source monitoring data for the monitoring area, including at least synthetic aperture radar image data, global navigation satellite system displacement time series data, and UAV observation data; The multi-source monitoring data is subjected to collaborative preprocessing to obtain preprocessed multi-source monitoring data, which includes: surface deformation information, millimeter-level three-dimensional displacement time series, and centimeter-level surface model and micro-fracture characteristics. The preprocessed multi-source monitoring data is unified in spatiotemporal reference and coupled and fused with multi-frequency time series to construct a spatiotemporally aligned fused dataset. The fused dataset is input into the pre-built ConvNeXt-ViT landslide early warning model for multi-source feature extraction and collaborative prediction, and the landslide risk level and deformation rate prediction results for the monitored area in the future preset period are output. The ConvNeXt-ViT landslide early warning model uses a 9-channel tensor as its input. This 9-channel tensor is constructed as follows: the first 3 channels are the 3 channels of the UAV digital orthophoto DOM; the synthetic aperture radar (InSAR) deformation field is normalized to the [0,1] interval using Min-Max and used as the 4th channel; the millimeter-level three-dimensional displacement time series of the landslide body is divided into 5 slices according to the time window, with each slice corresponding to 6 days of displacement data. The mean displacement of each slice is calculated and normalized to form the 5th to 9th channels; all channel data are uniformly sized to 224×224 pixels and are Z-score normalized. The ConvNeXt-ViT landslide early warning model includes: The local feature extraction branch consists of three stacked ConvNeXt blocks, each ConvNeXt block being paired with a GELU activation function and a layer normalization layerNorm; the local feature extraction branch outputs a 256-dimensional local feature vector. The global feature extraction branch consists of a 6-layer Transformer encoder, which first divides the input 9-channel 224×224 tensor into 16×16 image patches, for a total of 196 image patches; each image patch is mapped to a 768-dimensional embedding vector through a linear embedding layer; the global feature extraction branch outputs a 768-dimensional global feature vector. The feature fusion module employs a 4-head cross-attention mechanism, which maps the 256-dimensional local feature vector output from the local feature extraction branch to the same dimension and then concatenates them. The concatenated feature vector is then compressed into a 256-dimensional fused feature vector through a fully connected layer. This 256-dimensional fused feature vector is used to map and obtain the deformation rate prediction result for a future preset time period. The classification head, which is connected to the output of the feature fusion module, is used to obtain the landslide risk level of the monitoring area based on the 256-dimensional fused feature vector.

2. The method according to claim 1, characterized in that, The multi-source monitoring data undergoes collaborative preprocessing, including: Interferometric stacking processing is performed on the synthetic aperture radar image data to obtain surface deformation information of the monitored area; The displacement time series data of the Global Navigation Satellite System are used to perform positioning calculations to obtain the millimeter-level three-dimensional displacement time series sequence of the landslide body; Based on the surface deformation information, landslide risk areas are delineated, and UAV aerial surveys and data processing are carried out in the risk areas to obtain centimeter-level digital orthophotos, digital elevation models, and three-dimensional surface models, from which micro-fracture features are extracted.

3. The method according to claim 2, characterized in that, The synthetic aperture radar image data is subjected to interferometric stacking processing, including: Select single-view multiple images that cover the monitoring area and meet the preset time baseline and incident angle requirements; The image with the highest coherence is used as the main image, and it is registered, radiometrically calibrated and filtered. Based on short spatial baselines, short temporal baselines, and high coherence criteria, multiple effective interferometric pairs are screened and differential interferograms are generated. The differential interferograms are weighted, stacked, unwrapped, have their terrain phase removed, and atmospheric error corrected to obtain the surface deformation along the radar line of sight. After verification with data from the Global Navigation Satellite System, surface deformation information is generated.

4. The method according to claim 2, characterized in that, Positioning calculation is performed on the displacement time series data of the Global Navigation Satellite System, including: Base stations were set up in stable areas outside the landslide-affected area, and multiple mobile stations were set up in high-risk landslide areas. The raw observation data received by the base station and rover station are subjected to data integrity checks, gross error removal, format conversion, and preliminary error correction. Dual-frequency observations were used to eliminate ionospheric delay, and filtering algorithms were employed to suppress multipath effects. The relative positioning model is used to calculate the positioning and output the millimeter-level displacement time series of the landslide body along the east, north and vertical directions.

5. The method according to claim 2, characterized in that, Landslide risk zones are delineated based on the aforementioned surface deformation information, including: Based on preset deformation rate thresholds and cumulative displacement thresholds, and combined with the contours of deformation gradient abrupt change zones and continuous high deformation zones, landslide risk zones are delineated through geographic information system buffer analysis. Ground control points calibrated by the Global Navigation Satellite System will be deployed within the risk area.

6. The method according to claim 1, characterized in that, The preprocessed multi-source monitoring data undergoes spatiotemporal benchmark unification and multi-frequency time-series coupling fusion, including: The coordinate systems of all multi-source monitoring data are unified to the target geodetic coordinate system based on preset transformation parameters; Unify the time base of all multi-source monitoring data to the target time system; Multi-source time-series data with different observation frequencies are matched to a unified time node through interpolation or aggregation methods. Correlation analysis and regression modeling are performed on multi-source data after unifying the spatiotemporal benchmark, and time-series coupling fusion is completed and dynamic early warning thresholds are set.

7. The method according to claim 6, characterized in that, The association analysis and regression modeling include: Calculate the Pearson correlation coefficient between time series data from different sources and establish a regression model that reflects the quantitative relationship between different monitoring indicators; The regression model is validated using reserved data or manually measured data; Based on the results of the regression model and the geological characteristics of landslides, a dynamic early warning threshold that varies with environmental factors is set.

8. The method according to claim 1, characterized in that, The fused dataset is input into the pre-built ConvNeXt-ViT landslide early warning model for multi-source feature extraction and collaborative prediction, and also includes: The deformation intensity features of synthetic aperture radar and the surface texture features of UAV optical images are extracted using a pre-defined two-dimensional convolutional neural network, respectively. The displacement time series data of the Global Navigation Satellite System is converted into a two-dimensional map, and the dynamic trend features of displacement are extracted using a convolutional neural network; Convert UAV lidar point cloud data into voxel grids or elevation maps, and use 3D convolutional neural networks to extract 3D terrain morphology features; The extracted multi-source features are concatenated and fused across modalities, and temporal dynamic correlations are captured through a spatiotemporal recurrent neural network. Based on the fused spatiotemporal characteristics, the landslide risk level classification results and future deformation rate regression prediction results are output through a fully connected layer.

Images