Landslide susceptibility assessment method based on time-series insar

Abstract

This invention discloses a landslide susceptibility assessment method based on time-series InSAR, comprising: obtaining the surface deformation rate by time-series inversion of SAR images of the target monitoring area, and rendering it as an RGB color image; inputting the color image into a trained landslide hazard identification model to identify landslide hazard areas; if a landslide hazard area is detected, continuously monitoring the deformation of that area; calculating the total information content composed of multi-source influence factors for each grid in the target monitoring area; wherein, the multi-source influence factors include the deformation rate and several static influence factors; inputting the total information content of the multi-source influence factors of each grid in the target monitoring area into the landslide susceptibility assessment model to obtain the susceptibility level corresponding to each grid in the target monitoring area. This invention can significantly improve the efficiency of landslide hazard identification and the accuracy of landslide susceptibility assessment, and is applicable to the monitoring and assessment of geological disasters such as landslides in mountainous areas, urban subsidence, and surface deformation in mining areas.

Description

Technical Field

[0001] This invention belongs to the field of disaster identification and assessment, specifically relating to a landslide susceptibility assessment method based on time-series InSAR. Background Technology

[0002] Existing landslide monitoring methods (such as levels, total stations, GNSS, crack gauges, etc.) can capture surface deformation characteristics, but they have significant shortcomings in complex terrain conditions: on the one hand, traditional equipment is limited in terms of measurement accuracy, deployment density, and data acquisition convenience, making it difficult to achieve large-scale and efficient monitoring; on the other hand, field deployment and maintenance are labor-intensive and risky, greatly increasing the burden on operational safety; in addition, the current data processing mode, which still relies on manual visual interpretation, can hardly meet the actual needs of "automated and intelligent" landslide monitoring.

[0003] Landslide disasters exhibit multi-scale coupling characteristics, closely related to factors such as topography, geological conditions, climate, and human activities. Traditional susceptibility assessment systems largely rely on static environmental factors (geological, hydrological, etc.) and pay less attention to dynamic factors such as surface deformation and earthquakes, often leading to erroneous predictions of landslide susceptibility in certain areas. Furthermore, the accuracy of landslide susceptibility models heavily depends on historical landslide cataloging data. The historical landslide databases used for model training have significant deficiencies—domestic historical landslide data cataloging is incomplete, especially in mountainous areas where there are obvious cataloging gaps. Relying solely on traditional field survey methods (typically lagging by 3-5 years) makes it difficult to achieve annual data updates. These dual limitations in data foundation and modeling methods severely restrict the spatiotemporal accuracy of regional landslide risk early warning, failing to meet the requirements for accurate and timely landslide prediction. Summary of the Invention

[0004] This invention provides a landslide susceptibility assessment method based on time-series InSAR, which can effectively monitor landslide deformation and identify landslide disasters at an early stage.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0006] A landslide susceptibility assessment method based on time-series InSAR includes:

[0007] The time series of SAR images for the target monitoring area in the current time period is obtained, and the satellite line-of-sight deformation rate of the corresponding area is obtained by inversion. The deformation rate is then rendered and converted into an RGB color image.

[0008] The obtained color images are input into the trained landslide hazard identification model to identify the landslide hazard areas;

[0009] If a landslide hazard area is detected, all-weather time-series InSAR deformation monitoring will be carried out on the landslide hazard area, and an early warning mechanism will be triggered when surface deformation anomalies occur.

[0010] For each grid in the target monitoring area, the attribute values ​​of multi-source impact factors in the historical landslide catalog are obtained, and the total information content of the multi-source impact factors is calculated using an information content model; wherein, the multi-source impact factors include deformation rate and several static impact factors related to geology and environment;

[0011] The total information of multi-source influencing factors of each grid in the target monitoring area is input into the trained landslide susceptibility assessment model to obtain the susceptibility level of each grid in the target monitoring area.

[0012] Furthermore, the landslide hazard identification model is obtained by training a pre-built deep convolutional neural network structure;

[0013] The pre-built deep convolutional neural network structure includes: an input layer, a feature extraction module, a feature fusion module, and a detection layer;

[0014] The feature extraction module uses multi-layer convolution operations to extract features step by step in order to construct a multi-layer feature representation;

[0015] The feature fusion module first uses an optimized PANet architecture to integrate the extracted multi-level features, and then uses a multi-scale adaptive noise reduction module with an integrated attention mechanism to filter effective features.

[0016] The detection layer is used to generate the bounding box and confidence level of the landslide area based on the selected valid features.

[0017] Furthermore, the optimized PANet architecture adds a fourth detection branch to the original three detection branches of the Neck structure based on the YOLOv5 model to capture features at the small target scale. Simultaneously, the features extracted by the original three detection branches of the Neck structure are mapped to the image space of the fourth detection branch and fused with the features extracted by the fourth detection branch to generate a shallow gradient flow. Deep fusion of the multi-level feature representations extracted by the feature extraction module is achieved through a bidirectional propagation path that transmits semantic information of large targets from top to bottom and detailed features of small targets from bottom to top. Channel-dimensional splicing is employed to enhance the interaction and information complementarity between cross-level features.

[0018] Furthermore, the multi-scale adaptive noise reduction module integrating the attention mechanism includes:

[0019] First, the input feature map is divided into four independent branches;

[0020] Then, each branch is first filtered by a 1×1 convolutional network, and then three of the branches are processed by convolutional kernels of different scales to calculate the feature values ​​at each scale. Here, idendity represents the input feature map of the multi-scale adaptive noise reduction module.

[0021] The feature maps obtained from the four branches are then weighted by the channel attention mechanism.

[0022] Finally, the results from each branch are concatenated and then output through a 1×1 convolutional network for dimensionality reduction.

[0023] Furthermore, the aforementioned static influencing factors related to geology and environment include four types: topography, geological structure, human engineering activities, and natural environment; among them, the static influencing factors related to topography include: elevation, slope, aspect, curvature, plan curvature, profile curvature, topographic relief, surface roughness, and topographic humidity index.

[0024] Static influencing factors of geological structures include: geological age and distance from faults;

[0025] Static influencing factors of human engineering activities include: distance from roads and land use;

[0026] Static influencing factors of the natural environment include: vegetation normalization index, rainfall, and distance from water systems.

[0027] Furthermore, before constructing the multi-source influencing factors, the Pearson correlation coefficient analysis method is used to remove strongly correlated static influencing factors, and then the remaining static influencing factors are combined with the deformation rate to construct the multi-source influencing factor set.

[0028] Furthermore, the historical landslide catalog is constructed from the actual landslide areas and all landslide hazard areas detected and identified by the landslide hazard identification model, and is updated regularly based on new identification results.

[0029] Furthermore, the total information content of the statistical multi-source influence factors includes:

[0030] Each influencing factor is classified, and the relative historical landslide density under different classifications of each influencing factor is calculated:

[0031] ;

[0032] In the formula, Let k be the relative historical landslide density in the j-th grade interval; Let K be the number of historical landslides in the j-th grade interval; N be the total number of historical landslides in the target monitoring area; S represents the number of grid cells in the j-th grade interval for the influence factor k; S represents the total number of grid cells in the target monitoring area.

[0033] The probability of landslide occurrence is assessed by calculating the landslide information value corresponding to each grade interval of different influencing factors using the density characteristics of historical landslides. The calculation formula is as follows:

[0034] ;

[0035] In the formula, Let k be the information content of the influencing factor k in the j-th grade interval, and after normalization, it represents the probability of landslide disaster occurring in that interval;

[0036] Calculate the total information content of all impact factors, i.e., the total information content of multi-source impact factors:

[0037] ;

[0038] Where m represents the number of factors influencing landslide disasters. This represents the total information content of the multi-source influencing factors.

[0039] Furthermore, the landslide susceptibility assessment model is obtained by training and optimizing a random forest ensemble learning model based on decision trees.

[0040] Furthermore, when training and optimizing the decision tree-based ensemble learning model, Bayesian optimization is introduced to automatically adjust the model's hyperparameters.

[0041] Beneficial effects:

[0042] This invention provides a landslide susceptibility assessment method based on temporal InSAR. It automatically identifies landslide hazard areas and extracts landslide centroids through a trained landslide hazard identification model. Combining the national historical landslide database and multi-source heterogeneous environmental data, it utilizes information content and a random forest model to conduct landslide susceptibility assessment, coupled with Bayesian optimization to significantly improve the accuracy of the assessment. This invention achieves three aspects: monitoring landslide deformation rate, identifying landslide hazard points, and predicting landslide occurrence probability. It is applicable to early warning of landslide disasters in various complex terrain areas, demonstrating significant innovation and broad application prospects. Attached Figure Description

[0043] Figure 1 This is a flowchart of an automated time-series InSAR and landslide susceptibility assessment algorithm according to an embodiment of the present invention;

[0044] Figure 2 This is a flowchart illustrating the automated collection and standardized processing of InSAR data according to an embodiment of the present invention.

[0045] Figure 3This is a flowchart illustrating the hazard identification and interpretation process based on a deep convolutional neural network structure, as described in an embodiment of the present invention.

[0046] Figure 4 This is a structural diagram of the adaptive noise reduction module described in an embodiment of the present invention;

[0047] Figure 5 This is a flowchart illustrating the quantitative processing and landslide susceptibility assessment of the multi-source heterogeneous data evaluation system according to an embodiment of the present invention.

[0048] Figure 6 This is the result of automated processing of SBAS-InSAR deformation rate in a certain irrigation area according to an embodiment of the present invention;

[0049] Figure 7 This is the result of landslide susceptibility assessment of a certain irrigation district based on its learning model, according to an embodiment of the present invention. Detailed Implementation

[0050] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0051] This embodiment provides a landslide susceptibility assessment method based on temporal InSAR, referencing... Figure 1 As shown, it includes the following steps:

[0052] Step 1: Obtain the SAR image time series of the target monitoring area in the current time period, and invert the line of sight (LOS) deformation rate of the corresponding area. Then, render and convert it into an RGB color image for image recognition model processing.

[0053] In this embodiment, high-resolution SAR imagery data of the study area, along with corresponding precise orbit files and high-precision DEM digital elevation models, are acquired using the minimum satellite revisit period as the foundational data for inversion analysis. Based on Small Baseline Subsets InSAR (SBAS-InSAR) technology, the process involves generating a spatiotemporal baseline network, generating interferograms, phase unwrapping, solving for deformation rates, and geocoding to invert long-term wide-area deformation rate information of the study area, obtaining millimeter-level surface deformation results in a geographic coordinate system. (Reference) Figure 2 As shown, the automated collection and standardization of InSAR data includes:

[0054] Step 1.1: After defining the rectangular latitude and longitude range of the target area, the script performs a centroid search and acquires and displays SAR (Synthetic Aperture Radar) images of the target area. The first time this step is executed, the user needs to manually select the SAR image location covering the target area. Subsequently, the script automatically acquires SAR images at regular intervals of 12 days / 24 days.

[0055] Step 1.2: Automatically collect the corresponding POD Precise Orbit Ephemerides based on the "SAR image results of the target area".

[0056] Step 1.3: Using the user-defined "target area rectangular latitude and longitude range" as a reference, fill the entire latitude and longitude range by 2° and round it to obtain DEM digital elevation model data in a wide area, ensuring that its range completely covers the SAR image area.

[0057] Step 1.4: Repeat Step 1.1 to perform the target area centroid search again, ensuring the complete collection of long-term SAR images for the target area. Subsequently, the SBAS spatiotemporal baseline network is initially displayed with a default "spatial baseline 350m, temporal baseline 96d" and an image pair connectivity txt file is output. The user needs to intervene and adjust the parameters the first time this step is executed to ensure reliable SAR image pair connectivity and prevent any disconnections. After completion, the script will transfer the user-optimized "spatial baseline, temporal baseline" parameters for application in the standardization process.

[0058] Step 1.5: After completing the above steps, perform SAR image preprocessing with default parameters to generate an interferogram co-registration stack, including: Burst extraction, SLC co-registration, Burst stitching, interferogram generation, Goldstein interferogram filtering and denoising, and interferogram phase unwrapping.

[0059] Step 1.6: Perform SBAS-InSAR normalization processing based on the interferogram co-matching stack to obtain the deformation rate results in geographic coordinate system in raster format. Upon first execution of this step, the user will receive a txt text file storing SBAS-InSAR related computational parameters, including: absolute data read path, reference point latitude and longitude, low-coherence interferogram, and other information.

[0060] Step 1.7: In the GIS environment, the deformation rate raster obtained in Step 1.6 is color-rendered to obtain RGB three-channel data.

[0061] To improve the efficiency of subsequent landslide hazard identification, the extracted wide-area deformation rate results are automatically divided into blocks of 2048 pixels × 2048 pixels, which serve as the data basis for landslide hazard identification in step 2.

[0062] Step 2: Input the obtained color image into the trained landslide hazard identification model to detect and identify the landslide hazard areas.

[0063] The landslide hazard identification model described in this embodiment is obtained by training a pre-built deep convolutional neural network structure. The pre-built deep convolutional neural network structure refers to... Figure 3 As shown, it includes: an input layer, a feature extraction module, a feature fusion module, and a detection layer.

[0064] The input layer constructs a multi-channel input matrix for the image (3×H×W) after deformation rate color rendering.

[0065] The feature extraction module (Backbone) uses multi-layer convolution operations to extract features step by step, from low-level local features (edges, textures) to high-level abstract features (landslide morphology, deformation areas), gradually constructing a multi-level feature representation.

[0066] The feature fusion module (Neck) first uses an optimized PANet architecture to integrate the extracted multi-level features, and then uses a multi-scale adaptive noise reduction module (PANRM) that integrates attention mechanisms to filter effective features.

[0067] In satellite remote sensing images, the detected targets are extremely small, often accompanied by unclear textures and color distortion. Therefore, the optimized PANet architecture described in this embodiment adds a detection branch to the Neck structure based on the YOLOv5 model to capture small-scale targets. Simultaneously, it maps feature values ​​from other layers to the current layer's image space and fuses them with the current layer's features, enriching the feature information and generating a more reliable shallow gradient flow. A bidirectional propagation path, transmitting semantic information of large targets from top to bottom and detailed features of small targets from bottom to top, is used to deeply fuse the multi-level feature representations extracted by the feature extraction module. A 3-channel dimension (stacked according to width, height, and channel dimensions) concat operation is employed to enhance the interaction and information complementarity between cross-level features.

[0068] The multi-scale adaptive noise reduction module that integrates attention mechanisms, refer to Figure 4As shown, the process includes: First, the input feature map is divided into four independent branches; then, each branch is filtered by a 1×1 convolutional network, followed by three branches being processed by convolutional kernels of different scales (3×3, 5×5, 7×7) to calculate feature values ​​at each scale, where idendity represents the original input feature map; the feature maps obtained from the four branches are then weighted by an channel attention mechanism (ECA) to determine the high-value information to be focused on in the feature maps; finally, the results obtained from each branch are concatenated by channel and then output through a 1×1 convolutional network for dimensionality reduction. This multi-scale adaptive denoising module filters useful features and suppresses redundant noise features.

[0069] The detection layer is used to generate bounding boxes and confidence scores for landslide areas based on the selected valid features. The detection layer of the deep convolutional neural network structure simultaneously predicts whether each grid cell of the input deep convolutional neural network image contains a potential landslide risk.

[0070] The landslide hazard identification model used in this embodiment was trained using an iterative learning mechanism. The NECK layer was improved by adding a branch to focus on the features of small targets under high-resolution conditions, thus enhancing the predictive ability for small targets. When the confidence level of the model's automatic landslide hazard identification results is <60%, manual verification under an optical remote sensing base map is further considered to determine whether a landslide hazard exists.

[0071] Step 3: If a landslide hazard area is detected, all-weather temporal InSAR deformation monitoring is conducted on that area, and an early warning is triggered when surface deformation anomalies occur. Specifically, the system continuously acquires surface deformation rate data for the hazard area and calculates the real-time deformation rate during the monitoring period. When the monitored indicators exceed a preset warning threshold, an early landslide warning is triggered. The warning triggering condition can be expressed as:

[0072] ;

[0073] in, The cumulative deformation change or deformation rate during the monitoring period, where T is a preset warning threshold (e.g., |10| mm / yr or |20| mm / yr). When the above conditions are met, the system automatically initiates the warning response process, promptly alerting relevant management departments to take preventative measures.

[0074] Step 4: For each grid cell in the target monitoring area, obtain its multi-source influence factor attribute values ​​from the historical landslide catalog, and calculate the total information content of the multi-source influence factors using an information content model. The multi-source influence factors include deformation rate and several static influence factors related to geology and the environment. (Reference) Figure 5 As shown.

[0075] (1) Establish multi-source influencing factors.

[0076] This invention uses a landslide disaster database as training samples for a landslide susceptibility assessment model. This database is based on actual landslide disasters from the "National Geological Disaster Point Data" for a specific year, and combines this with landslide hazard areas detected and identified by a convolutional neural network-based landslide hazard model to construct and dynamically update a "historical landslide catalog."

[0077] Secondly, taking into account the geographical location and natural environmental characteristics of the target area, and ensuring the objectivity and availability of data, we fully considered the systematic, hierarchical and non-redundant nature of landslide influencing factors, and collected multi-source geological and environmental data of the target study area, covering four major categories of static influencing factors: topography (elevation, slope, aspect, curvature, plane curvature, profile curvature, topographic relief, surface roughness, topographic humidity index), geological structure (geological age, distance from fault), human engineering activities (distance from road, land use), and natural environment (normalized vegetation index NDVI, rainfall, distance from water system).

[0078] In the GIS environment, based on the "historical landslide catalog" (point vector) data, the above-mentioned influencing factors are processed by "multi-value extraction to historical landslide catalog points" to obtain the factor attribute values ​​corresponding to each historical landslide point, which are used as input features for training the landslide susceptibility model.

[0079] As a preferred technical solution of the present invention, a Pearson correlation coefficient function is constructed to quantify the correlation coefficients of the above-mentioned static influence factors and remove strongly correlated static influence factors. Based on the judgment matrix, factors with higher average correlation coefficients to other static influencing factors were eliminated, ultimately establishing the static influencing factors participating in the landslide susceptibility model training. The formula for calculating the Pearson correlation coefficient is as follows:

[0080] ;

[0081] in, and These are the attribute values ​​(such as slope value and elevation value) of two different influencing factors for the i-th landslide point in the historical landslide catalog. and These represent the average values ​​of all landslide points corresponding to the influencing factors; n represents the number of landslides in the historical landslide catalog (historical landslide catalog data = landslide hazard identification and interpretation data + historical actual landslide disaster point data).

[0082] (2) Quantitative processing of multi-source influence factors.

[0083] The established multi-source influencing factors have differences in spatial resolution and data structure. All influencing factors need to be standardized in the GIS environment (including resampling, projection transformation and vector rasterization) to ensure that the data is in raster format and that elements such as cell size (30m), resolution and projection coordinate system are consistent.

[0084] The above influencing factors are multi-source heterogeneous data, mostly in continuous raster format, making them difficult to use directly for model analysis. Before assessing landslide susceptibility, each influencing factor needs to be graded and processed into discrete raster data to accurately reflect the nonlinear relationship between influencing factors and landslide occurrence, and to intuitively reveal the patterns between each influencing factor and landslide distribution. The processing method is as follows: in a GIS environment, different influencing factor raster data are reclassified; line vector data (such as road vectors, water system vectors, and fault vectors) need to undergo Euclidean distance calculation and be converted into continuous raster data before reclassification. Generally, the natural breakpoint method is used for reclassification, and the number of graded intervals is 5.

[0085] After obtaining the grading intervals of all influencing factors, a quantitative analysis of the contribution of influencing factors to landslides was conducted using a 30m×30m raster pixel as the smallest evaluation unit.

[0086] First, calculate the relative historical landslide density of each influencing factor to reflect its sensitivity to landslide occurrence:

[0087] ;

[0088] In the formula, Let k be the relative historical landslide density in the j-th grade interval; Let K be the number of historical landslides in the j-th grade interval; N be the total number of historical landslides in the target monitoring area; denoted as the number of graticules in the j-th grade interval for the influence factor k; S is the total number of graticules in the target monitoring area.

[0089] Then, based on information theory, an information content model is constructed. Using the density characteristics of historical landslides, the landslide information content values ​​corresponding to different influencing factors at each grade level are calculated to assess the probability of landslide occurrence. The calculation formula is as follows:

[0090] ;

[0091] In the formula, Let be the information content of the influencing factor k in the j-th grade interval, and after normalization, it represents the probability of landslide disaster occurring in this interval.

[0092] Finally, calculate the total information content of all impact factors, that is, the total information content of multi-source impact factors:

[0093] ;

[0094] Where m represents the number of factors influencing landslide disasters. This represents the total information content of the multi-source influencing factors. The higher the total information content value, the greater the probability of a landslide disaster occurring in that grid.

[0095] This embodiment uses the natural breakpoint method to classify the levels of the influencing factors. Different grid attribute values ​​belong to different intervals. For example, the slope is divided into the following 5 intervals: 0–5°, 5–15°, 15–30°, 30–45°, and >45°.

[0096] Suppose a pixel has the following characteristics: its slope is 12°, belonging to the second-level interval (5–15°), with an information content of 0.4; its NDVI is 0.65, belonging to the third-level interval, with an information content of 0.6; and its rainfall is 1200 mm, belonging to the first-level interval, with an information content of 0.2. Then, the total information content of this pixel = 0.4 + 0.6 + 0.2 = 1.2 (assuming only three influencing factors: slope, NDVI, and rainfall).

[0097] Step 5: Input the total information of multi-source influencing factors of each grid in the target monitoring area into the landslide susceptibility assessment model to obtain the susceptibility level of each grid in the target monitoring area. The landslide susceptibility assessment model is obtained by training and optimizing an ensemble learning random forest (RF) model based on decision trees.

[0098] The training data input for the landslide susceptibility assessment model is as follows: positive samples are historical landslide catalog data (including historical landslide databases and landslide hazards identified and interpreted by convolutional neural networks); negative samples are randomly generated non-landslide points, avoiding known landslide buffer zones and water system units during random sampling, forming a complete training dataset containing both landslide and non-landslide samples. Positive and negative samples are divided into a training set (70%) and a validation set (30%) to maintain spatial distribution balance, and the fields of both positive and negative samples should contain the total information content attribute values ​​of m landslide influencing factors.

[0099] The hyperparameters of the landslide susceptibility assessment model are defined as follows: the larger the number of decision trees (n_estimators), the more stable the model, but the computational cost increases; the maximum depth of the trees (max_depth) suppresses overfitting and captures complex nonlinear relationships; the minimum number of samples required for node splitting (min_samples_split) indicates the minimum number of samples required for internal node splitting.

[0100] Training an ensemble learning random forest model based on decision trees: During training, the model learns the complex nonlinear relationship between influencing factors and landslide hazards through human interaction.

[0101] Bayesian optimization: Bayes-SearchCV is introduced to automatically adjust the hyperparameters of decision tree-based ensemble learning random forest models, optimize model parameters, accelerate the hyperparameter tuning process, and improve training efficiency.

[0102] Performance evaluation of the trained landslide susceptibility assessment model: To ensure the stability of the model, cross-validation and AUC (Area Under Curve) were used to evaluate the model performance.

[0103] Visualization of evaluation results: Based on the random forest model optimized by Bayesian hyperparameters, raster images of landslide occurrence probability in the study area were extracted. In GIS, the landslide occurrence probability was reclassified into 5 intervals using the natural breakpoint method, resulting in a landslide susceptibility zoning map, which includes five susceptibility levels: extremely low susceptibility, low susceptibility, moderate susceptibility, high susceptibility, and extremely high susceptibility.

[0104] Iterative optimization: The deformation rate obtained in step 2 is used as a dynamic influencing factor. Together with several other static influencing factors, the total information content is calculated. Then, the total information content of all grids in the target monitoring area is input into the landslide susceptibility evaluation model according to the grid arrangement format. The landslide susceptibility of all grids in the target monitoring area is evaluated, and the timeliness of the landslide susceptibility evaluation is dynamically improved, providing probability information of landslides occurring in the study area in the near future.

[0105] Example:

[0106] This case study focuses on the Quanmutang Irrigation District, located in the central area of ​​the "Hengshao Arid Corridor" in Hunan Province (111°10′ to 112°10′ east longitude, 26°30′ to 27°30′ north latitude), which is divided into 20 irrigation areas. The terrain is mainly hilly and mountainous, with a high elevation in the southwest and a low elevation in the northeast. The highest elevation is 1498 m, and the lowest is 56 m. The Quanmutang Irrigation District spans the Zishui and Xiangjiang River basins, encompassing almost all landform types, including erosion-deposition, dissolution, denudation, and erosional landforms. The tectonic features are complex and interwoven, mostly exhibiting NE-trending tectonic features, manifested as faults and folds. Because most open channels, aqueducts, and other water conservancy projects are built along mountainsides, the construction of related supporting facilities has resulted in long-term disturbance of the surrounding mountain slopes, reducing slope stability and increasing the risk of landslides.

[0107] The SAR image data used for deformation rate inversion in this example is the Sentinel-1A Single Look Complex (SLC) image provided by the Copernicus Browser. This image contains the raw amplitude and phase information of the radar echo signals received during satellite acquisition, and has a high spatial resolution (5 m × 20 m), which can meet the accuracy requirements for surface deformation monitoring and inversion in the Quanmutang Irrigation District.

[0108] S1: InSAR Data Acquisition. The latitude and longitude of the Quanmutang Irrigation District were filled with 2° and rounded to the nearest integer to collect DEM (Digital Elevation Model) data for the study area. Sentinel-1A SLC-level SAR imagery and precise orbit files were automatically acquired every 12 or 24 days via script. In this example, 80 images were acquired. Taking a single image as an example, its naming format is as follows:

[0109] S1A_IW_SLC__1SDV_20240814T103501_20240814T103528_055208_06BAB0_4EC5.zip

[0110] The corresponding POD precise orbit determination ephemeris data is as follows:

[0111] S1A_OPER_AUX_POEORB_OPOD_20240903T070636_V20240813T225942_20240815T005942.EOF

[0112] S2: After completing the above InSAR data collection, preprocess the data based on the SBAS-InSAR technology principle, including: Burst extraction, SLC co-registration, Burst stitching, interferogram generation, interferogram filtering and denoising, and interferogram phase unwrapping.

[0113] SBAS-InSAR technology employs a multi-master image matching strategy, freely combining SAR images to form multiple small baseline subsets that meet spatiotemporal baseline thresholds. Images meeting these conditions are paired and used as the basis for interferometry, ensuring that connected image pairs have short spatiotemporal baselines. This effectively reduces temporal decorrelation, improves interferogram quality, significantly increases the number of coherent pixels, and removes flat and topographic phases from spatially discrete points. Finally, by performing phase unwrapping and deformation calculations on each small baseline subset, the time series results for each subset are obtained. Singular Value Decomposition (SVD) is then used for joint solution to obtain the time series and average deformation rate of the surface deformation.

[0114] A spatiotemporal baseline network was generated. Based on 80 SLC images of the study area, a threshold control method was used, setting "spatial baseline = 350m, temporal baseline = 96d". Images meeting the threshold conditions were linked and combined, resulting in 321 image link pairs. The maximum spatial baseline was 339.06m, the minimum spatial baseline was 10.93m, the maximum temporal baseline was 96d, and the minimum temporal baseline was 12d. The average number of link pairs per SAR image was 8, ensuring interferogram quality, minimizing the effects of decoherence, and providing sufficient image pair redundancy.

[0115] Differential interferometry processing. To improve computational efficiency, the range-to-azimuth ratio during interferometry processing is set to 5:1, ensuring consistent spatial resolution in the two corresponding directions. This reduces noise levels and improves the reliability of the interferometric results. Coherence coefficient maps and interferograms are generated; the pixels in this step are approximately squares with a side length of 25m. The phase information of any pixel x in a given interferogram is:

[0116] ;

[0117] Where λ is the wavelength of the band used in the SAR image; and These are the cumulative deformations in the radar line-of-sight (LOS) direction at times ta and tb, respectively; This represents the phase difference between two moments caused by atmospheric delay effects, which is not removed by default in this invention; Randomly generated noise phase differences can be reduced through filtering; The residual topographic phase in differential interferogram i can be removed by simulation using external digital elevation model (DEM) data, ultimately preserving the deformation phase information.

[0118] Interferogram phase unwrapping. This example uses the SNAPHU (Stanford Method for Permanent Scatterers in Radar Interferometry) phase unwrapping package and parallel processing to achieve unwrapping. To reduce the influence of noise phase, the Goldstein filtering method is used to smooth the interferogram with a filter strength of 0.5 to ensure the accuracy and reliability of the unwrapping results.

[0119] Deformation reference point selection. The interferometric phase obtained by InSAR measurement is a relative observable. For any pixel, the phase of any unwrapped interferogram is relative. Therefore, it is necessary to reference all interferograms to a common reference pixel. This invention can automatically select pixels with high spatial coherence (≥0.85) or select reference points based on user-defined stable points.

[0120] Geocoding of Deformation Results. Since the above deformation information was calculated in the SAR coordinate system, in order to obtain deformation information that matches the geographic coordinate system, it is necessary to convert the deformation inversion results from the SAR coordinate system to the geographic coordinate system, obtain the encoded deformation rate and vector deformation points, and remove invalid values ​​outside the image range to generate the final surface deformation rate map and cumulative deformation map of the Quanmutang Irrigation Area. Figure 6 The result is the automated processing result of SBAS-InSAR deformation rate in a certain irrigation area.

[0121] S3: Delineating the landslide range based on InSAR deformation rate using the YOLOv5 model architecture.

[0122] The identified landslides are linked with the actual historical landslide database to construct a catalog of historical landslides in the study area (360 in this embodiment).

[0123] S4: Data Collection and Preprocessing for the Evaluation Factor System: Considering the unique geographical location and natural environmental characteristics of the irrigation area, and ensuring data objectivity and availability, the InSAR deformation rate (a dynamic influencing factor) and 14 static influencing factors were selected. The static influencing factors include four main categories: topography (elevation, slope, aspect, plane curvature, profile curvature, topographic relief, surface roughness, topographic humidity index), geological structure (geological age, distance from faults), human engineering activities (distance from roads, land use), and natural environment (NDVI, rainfall, distance from water systems). These factors essentially cover the main driving factors of landslides. The data was standardized and converted to a 30m×30m raster format, ultimately generating a raster dataset of 3498 rows × 3160 columns.

[0124] In a GIS environment, based on historical landslide cataloging (point vector) data, the attribute values ​​of the aforementioned influencing factors are extracted to landslide points using a multi-value extraction tool to form a training sample set. The information content of landslide influencing factors is calculated: based on information theory, for different influencing factors, the relative historical landslide density is used to represent the distribution intensity of landslides within each factor interval, and the total information content of the pixels is summed.

[0125] S5: Using historical landslide catalog data as training samples, and combining a random negative sample sampling strategy (avoiding landslide and water buffer zones), a complete training dataset containing both landslide and non-landslide samples is formed. Attribute values ​​of 14 static influencing factors are extracted from the GIS environment and divided into a training set (70%) and a test set (30%). A decision tree-based ensemble learning model, Random Forest (RF), is constructed to handle the classification and prediction task in landslide susceptibility assessment. Landslide prediction results are obtained through Bayes-SearchCV optimization. Susceptibility levels are calculated, and based on the natural breakpoint method, the area is divided into five susceptibility levels: extremely low, low, medium, high, and extremely high, generating a susceptibility zoning map. Figure 7 This is the result of landslide susceptibility assessment in a certain irrigation area based on its learning model.

[0126] Model accuracy was validated by using cross-validation and the AUC (Area Under Curve) value of the ROC curve to evaluate model performance.

[0127] Iterative optimization of the landslide susceptibility assessment model: Based on the deformation rate results obtained by SBAS-InSAR processing, its information content value is calculated, a static-dynamic influence factor is constructed, and coupled with the RF landslide susceptibility assessment model to reflect the dynamic process of landslide occurrence, provide information on the probability of recent occurrence of landslide hazard points, and effectively improve the model performance.

[0128] S6: Intelligent Early Warning System for Potential Hazard Points. Based on deep learning model identification results and supplemented by InSAR technology for real-time monitoring, it continuously monitors areas of significant deformation in the Quanmutang Irrigation District and triggers early warnings when abnormal surface deformation occurs. The warning conditions are as follows:

[0129] (In this embodiment, T=10mm / yr);

[0130] When InSAR data shows changes in the total amount of information about surface deformation When the threshold T is exceeded, an early warning is triggered: the warning information is promptly issued to the local government and residents, and dangerous slope areas of the reservoir are marked to reduce the risk of landslides blocking the river or potential landslides.

[0131] These specific implementation methods are applicable to various geological disaster types, including landslides in mountainous areas, urban subsidence, and deformation in mining areas, and have broad application prospects in disaster prevention, mitigation, and risk management.

[0132] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, such changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A landslide susceptibility assessment method based on time-series InSAR, characterized in that, include: The time series of SAR images for the target monitoring area in the current time period is obtained, and the satellite line-of-sight deformation rate of the corresponding area is obtained by inversion. The deformation rate is then rendered and converted into an RGB color image. The obtained color images are input into the trained landslide hazard identification model to identify the landslide hazard areas; If a landslide hazard area is detected, all-weather time-series InSAR deformation monitoring will be carried out on the landslide hazard area, and an early warning mechanism will be triggered when surface deformation anomalies occur. For each grid in the target monitoring area, the attribute values ​​of multi-source impact factors in the historical landslide catalog are obtained, and the total information content of the multi-source impact factors is calculated using an information content model. The multi-source impact factors include dynamic impact factors of deformation rate and several static impact factors related to geology and environment. The historical landslide catalog is constructed from historically actual landslide areas and all landslide hazard areas detected and identified by the landslide hazard identification model, and is updated regularly based on new identification results. The total information of multi-source influencing factors of each grid in the target monitoring area is input into the trained landslide susceptibility assessment model to obtain the susceptibility level of each grid in the target monitoring area. The landslide hazard identification model is obtained by training a pre-built deep convolutional neural network structure. The pre-built deep convolutional neural network structure includes: an input layer, a feature extraction module, a feature fusion module, and a detection layer; The feature fusion module first uses an optimized PANet architecture to integrate the extracted multi-level features, and then uses a multi-scale adaptive noise reduction module with an integrated attention mechanism to filter effective features. The optimized PANet architecture adds a fourth detection branch to the original three detection branches of the Neck structure based on the YOLOv5 model. This fourth branch is used to capture the features of small targets. Simultaneously, the features extracted by the original three detection branches of the Neck structure are mapped to the image space of the fourth detection branch and fused with the features extracted by the fourth detection branch to generate a shallow gradient flow. A bidirectional propagation path, transmitting semantic information of large targets from top to bottom and detailed features of small targets from bottom to top, is used to deeply fuse the multi-level feature representations extracted by the feature extraction module. Channel-dimensional splicing is employed to enhance the interaction and information complementarity between cross-level features. The multi-scale adaptive noise reduction module integrating attention mechanisms includes: First, the input feature map is divided into four independent branches; Then, each branch is first filtered by a 1×1 convolutional network, and then three of the branches are processed by deep convolutional kernels of different scales to calculate the feature values ​​at each scale. The feature maps obtained from the four branches are then weighted by the channel attention mechanism. Finally, the results from each branch are concatenated and then output through a 1×1 convolutional network for dimensionality reduction.

2. The landslide susceptibility assessment method based on time-series InSAR according to claim 1, characterized in that, The feature extraction module uses multi-layer convolution operations to extract features step by step in order to construct a multi-layer feature representation; The detection layer is used to generate the bounding box and confidence level of the landslide area based on the selected valid features.

3. The landslide susceptibility assessment method based on time-series InSAR according to claim 1, characterized in that, The aforementioned static influencing factors related to geology and the environment include four types: topography, geological structure, human engineering activities, and natural environment; among them, Static influencing factors related to topography and geomorphology include: elevation, slope, aspect, curvature, plan curvature, profile curvature, topographic relief, surface roughness, and topographic humidity index. Static influencing factors of geological structures include: geological age and distance from faults; Static influencing factors of human engineering activities include: distance from roads and land use; Static influencing factors of the natural environment include: vegetation normalization index, rainfall, and distance from water systems.

4. The landslide susceptibility assessment method based on time-series InSAR according to claim 3, characterized in that, Before constructing the multi-source influence factors, the Pearson correlation coefficient analysis method was used to remove strongly correlated static influence factors, and then the remaining static influence factors were combined with the deformation rate dynamic influence factors to construct the multi-source influence factor set.

5. The landslide susceptibility assessment method based on time-series InSAR according to claim 1, characterized in that, The calculation of the total information content of the multi-source impact factor includes: Each influencing factor is classified, and the relative historical landslide density under different classifications of each influencing factor is calculated: ； In the formula, Let k be the relative historical landslide density in the j-th grade interval; Let K be the number of historical landslides in the j-th grade interval; N be the total number of historical landslides in the target monitoring area; S represents the number of grid cells in the j-th grade interval for the influence factor k; S represents the total number of grid cells in the target monitoring area. The probability of landslide occurrence is assessed by calculating the landslide information value corresponding to each grade interval of different influencing factors using the density characteristics of historical landslides. The calculation formula is as follows: ； In the formula, Let k be the information content of the influencing factor k in the j-th grade interval, and after normalization, it represents the probability of landslide disaster occurring in that interval; Calculate the total information content of all impact factors, i.e., the total information content of multi-source impact factors: ； Where m represents the number of factors influencing landslide disasters. This represents the total information content of the multi-source influencing factors.

6. The landslide susceptibility assessment method based on time-series InSAR according to claim 1, characterized in that, The landslide susceptibility assessment model is obtained by training and optimizing a random forest ensemble learning model based on decision trees.

7. The landslide susceptibility assessment method based on time-series InSAR according to claim 6, characterized in that, When training and optimizing decision tree-based ensemble learning models, Bayesian optimization is introduced to automatically adjust the model's hyperparameters.

Images