Landslide hidden danger point identification and classification method and system based on InSAR

Abstract

The invention discloses a landslide hidden danger point identification and classification method and system based on InSAR, and the method comprises the following steps: processing regional InSAR data based on a StaMPS-SBAS time sequence InSAR processing method of a fusion distributed scatterer target, and obtaining a regional deformation rate diagram; acquiring an optical image map consistent with the time of the regional deformation rate map of the monitored region, and producing a regional deformation rate map fused with the regional optical image; integrating the regional deformation rate diagram and the regional deformation rate diagram fusing the regional optical influence, making a data set for training, and inputting the data set into a deep learning model to train a landslide classification model; and inputting the InSAR deformation rate diagram to be interpreted into the landslide classification model, and outputting the landslide position and landslide classification of the InSAR deformation rate diagram by the landslide classification model. According to the method, the high-quality regional deformation rate diagram can be obtained, the automation degree is high, the landslide position in the region can be automatically identified, and the landslide can be classified.

Description

Technical Field

[0001] The invention relates to a method for identifying and classifying landslide hazard points, and in particular to an InSAR-based method and system for identifying and classifying landslide hazard points. Background Art

[0002] Landslide is a common natural disaster caused by natural factors or human activities, mainly occurring in mountainous areas and river valleys. Using InSAR deformation rate maps to identify and classify landslide points in a region is very common in landslide hazard monitoring. For the identification and classification of landslide hazard points, under existing technical conditions, the use of InSAR deformation maps to complete the identification and classification of landslide hazard points still mainly relies on manual interpretation. Due to the wide coverage of InSAR deformation maps, manual interpretation is time-consuming and labor-intensive, and relies on the interpretation personnel's work experience, which is prone to misjudgment and omission. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a method and system for identifying and classifying landslide hazard points based on InSAR, which can automatically identify and classify landslide hazard points in a target area with high identification accuracy.

[0004] Technical solution: The InSAR-based landslide hazard point identification and classification method of the present invention comprises the following steps:

[0005] (1) The regional InSAR data are processed based on the StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets to obtain the regional deformation rate map. The quality of the regional deformation rate map can be improved by processing the InSAR data based on the StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets. The high-quality regional deformation rate map is the prerequisite for the accurate identification and classification of the entire landslide hazard point.

[0006] (2) Collecting optical images that are consistent with the regional deformation rate map of the monitored area in time, and producing a regional deformation rate map that integrates the optical images of the area;

[0007] (3) Integrate the regional deformation rate map and the regional deformation rate map with regional optical influence to create a training data set, and input the data set into the deep learning model to train the landslide classification model;

[0008] Landslides are scattered and there is no need to extract boundaries. The landslide classification model uses the regional deformation rate map as input and performs target detection. Landslides are complex and need to extract boundaries. The landslide classification model uses the regional deformation rate map fused with the regional deformation rate map as input and performs instance segmentation.

[0009] Annotate the deformation rate map and the regional deformation rate map of the fused regional optical image to obtain annotated instance segmentation module training set and target detection module training set;

[0010] The instance segmentation module training set and the target detection module training set are input into the deep learning model training to obtain the landslide classification model;

[0011] (4) The InSAR deformation rate map to be interpreted is input into the landslide classification model, and the landslide classification model outputs the landslide location and landslide classification of the InSAR deformation rate map. When the landslide situation of the InSAR deformation rate map is complex, the landslide classification model also extracts the deformation area boundary.

[0012] Furthermore, the specific steps of the StaMPS-SBAS time-series InSAR processing method based on fusion of distributed scatterer targets include:

[0013] Preprocess the regional InSAR data to obtain the differential interferometric phase set;

[0014] Selecting ground points with similar or consistent scattering characteristics in the differential interference phase set as homogeneous points to obtain a set of homogeneous points;

[0015] Set a homogeneous point number threshold, extract homogeneous points with a homogeneous point number greater than the threshold as DS points, extract SDFP points of the slow decoherence filter phase, and merge the SDFP points with the DS points;

[0016] Based on the 3-D phase unwrapping algorithm, the entangled phase of the SDFP / DS fusion point set is unwrapped, the DEM error and SCLA / AOE are estimated and removed according to the phase stability, and the atmospheric delay phase is estimated and removed through spatiotemporal filtering, and finally the deformation rate map is obtained.

[0017] Furthermore, the specific steps of preprocessing the regional InSAR data to obtain the differential interferometric phase set are as follows:

[0018] By adopting enhanced spectral diversity method, the single-view complex images of InSAR data are registered and the time series of intensity images of single-view complex images are extracted.

[0019] The previously acquired SRTMDEM data are registered with the selected super master image and simulated as an interferometric phase;

[0020] Under the setting of the temporal and spatial baseline threshold, a differential interferometric combination is constructed to perform interferometric operations on the registered SLC image sequence.

[0021] The terrain phase simulated by the DEM data is removed to obtain the differential interferometric phase set.

[0022] Furthermore, the specific steps to obtain a set of homogeneous points are as follows:

[0023] The homogeneous point set based on likelihood ratio test is used as the initial value to refine the reference center pixel;

[0024] Small intervals and high-quality observations are used to iteratively control type II errors and minimize the number of heterogeneous points in a set of homogeneous points.

[0025] The specific steps for extracting DS points and SDFP points are as follows:

[0026] Performing homogeneous filtering on the differential interferometric phase set according to the homogeneous point set, and obtaining the differential interferometric phase set after homogeneous phase optimization;

[0027] A homogeneous point number threshold is set according to the homogeneous point set, and points with homogeneous point numbers greater than the threshold are initially selected as DS points, and then spectral filtering in the range / azimuth direction is performed to obtain the DS preliminary point set DSC;

[0028] After homogeneous phase optimization, the differential interferometric phase set is filtered in range and azimuth directions, and the SDFP points are preliminarily screened according to the amplitude difference threshold method to obtain the SDFP preliminary point set SDFPC.

[0029] Perform BP-LP combined filtering on SDFPC and DSC respectively to estimate the temporal coherence of each point in SDFPC and DSC;

[0030] Repeat the above SDFP point screening method to select all SDFP points from SDFPC and screen DS points from DSC.

[0031] Furthermore, the specific steps of inputting the data set into the deep learning model to train the landslide classification model are:

[0032] Divide the data set into training set, validation set and test set according to the preset ratio;

[0033] Use the training set to train a deep learning model with the self-attention mechanism as the core, and record the model parameters when the loss function value of the validation set reaches the minimum;

[0034] The trained deep learning model is used to predict the data of the test set to verify the accuracy of the deep learning model, and the deep learning model whose accuracy meets the preset requirements is retained as the landslide classification model.

[0035] Furthermore, the target detection module is built based on the open source framework Faster-RCNN and trained using a regional deformation rate map.

[0036] Furthermore, the instance segmentation module is constructed with a deep learning model based on the self-attention mechanism, and uses a deformation rate map of fused optical images for training.

[0037] Based on the same inventive concept, the present invention also provides a landslide hazard point identification and classification system based on InSAR, comprising:

[0038] Input module, used to input InSAR data;

[0039] The conversion processing module is used to receive the InSAR data from the input module and process it using the StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets to obtain a regional deformation rate map;

[0040] A fusion module is used to obtain an optical image map that is consistent with the regional deformation rate map of the monitored area in time, and to produce a regional deformation rate map of the optical image of the fused area;

[0041] A data set acquisition module, used to integrate the regional deformation rate map and the regional deformation rate map integrating the regional optical influence, and to produce a data set for training;

[0042] The training module uses the dataset to train the deep learning model and obtain the landslide classification model;

[0043] The recognition and classification module decodes the InSAR deformation rate map based on the landslide classification model and outputs the landslide location and landslide classification of the InSAR deformation rate map.

[0044] Furthermore, the data sets acquired by the data set acquisition module include the following two types:

[0045] Landslides are scattered and there is no need to extract boundaries. The landslide classification model uses the regional deformation rate map as input and performs target detection. The deformation rate map is annotated to obtain the annotated instance segmentation module training set.

[0046] The landslide situation is complex and the boundary needs to be extracted. The landslide classification model takes the regional deformation rate map of the fused regional deformation rate map as input and performs instance segmentation. The regional deformation rate map of the fused regional optical image is annotated to obtain the annotated target detection module training set.

[0047] The method of using the data set to train the deep learning model to obtain the landslide classification model includes: dividing the data set into a training set, a validation set, and a test set according to a preset ratio;

[0048] Use the training set to train a deep learning model with the self-attention mechanism as the core, and record the model parameters when the loss function value of the validation set reaches the minimum;

[0049] The trained deep learning model is used to predict the data of the test set to verify the accuracy of the deep learning model, and the deep learning model whose accuracy meets the preset requirements is retained as the landslide classification model.

[0050] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: the present invention can obtain high-quality regional deformation rate maps, which is conducive to improving the overall accuracy of identifying landslide hazard points; the present invention has a high degree of automation, can automatically identify the location of landslides in the region, and can extract regional boundaries as needed, and can classify landslides according to the activity state and intensity of landslide deformation. The present invention is based on high-quality regional deformation rate maps and optical image maps, and uses a landslide classification model based on deep learning to apply to landslide detection, replacing manual automatic detection with high work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The data processing flow of the method of the present invention.

[0052] Figure 2 This is a test comparison diagram of Example 1 of the present invention.

[0053] Figure 3 This is a framework diagram of the target detection module of Example 1 of the present invention.

[0054] Figure 4 It is a framework diagram of the instance segmentation module of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0056] Example 1

[0057] As shown in the figure, the landslide hazard point identification and classification method based on InSAR of the present invention comprises the following steps:

[0058] (1) The StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets processes regional InSAR data to obtain regional deformation rate maps. The quality of regional deformation rate maps can be improved by processing InSAR data based on the StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets. Distributed scatterers refer to the existence of multiple scatterers in the resolution unit of a pixel, and the backscattering signal of no single object dominates the pixel signal. The phase values ​​of these scatterers are randomly distributed within the period of (-π, π). Scatterers with this characteristic are called distributed scatterers. The StaMPS-SBAS (StaMPSSmallBaselineSubset) time-series InSAR processing method refers to a software tool for surface deformation monitoring, which is based on the StaMPS (Stanford Method for Persistent Scatterers) algorithm and combined with the SBAS (SmallBaselineSubset) method. StaMPS-SBAS can extract surface deformation information from synthetic aperture radar interferometry (InSAR) data for studying crustal movement, seismic activity, groundwater extraction, etc.

[0059] Specifically, the specific steps of the StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets include the following four steps:

[0060] (11) Preprocess the regional InSAR data to obtain the differential interferometric phase set.

[0061] By adopting the enhanced spectral diversity method, the single-look complex image of InSAR data is registered, and the time series of the intensity image of the single-look complex image is extracted. Among them, the enhanced spectral diversity method refers to a technology used in wireless communication systems, which increases the diversity of signals by using multiple different spectrum resources in transmission, thereby reducing the possibility of signal interference and improving the capacity and coverage of the system. It combines signal enhancement and spectrum diversity technology to improve the robustness and efficiency of the system while ensuring signal quality. Single-look complex image (SLC) is a form of SAR (synthetic aperture radar) image. The single-look complex image is the original SAR image after processing the entire Doppler bandwidth and correcting it with the latest various satellite parameters (auxiliary data). Each pixel is represented by a complex number, including real and imaginary parts. The modulus of the complex number is the intensity of the SAR image, and the argument of the complex number is the phase of the SAR image. The intensity image refers to the image obtained by calculating the modulus of each pixel of the single-look complex image.

[0062] The previously acquired SRTM DEM data is registered with the selected super master image and simulated as an interferometric phase. The SRTM DEM data refers to the terrain elevation data obtained by the US Shuttle Radar Topography Mission (SRTM) and is presented in the form of a digital elevation model (DEM); the SRTM DEM data exists in the form of raster data, which divides the earth's surface into grid cells of equal size, each containing an elevation value. The super master image refers to an image selected from all interferogram images for registration with other remaining interferogram images.

[0063] Under the setting of the spatiotemporal baseline threshold, a differential interferometric combination is constructed to perform interferometric operations on the registered SLC image sequence.

[0064] The terrain phase simulated by the DEM data is removed to obtain the differential interferometric phase set.

[0065] (12) Ground points with similar or consistent scattering characteristics in the differential interference phase set are selected as homogeneous points to obtain a set of homogeneous points.

[0066] First, the homogeneous point set based on the likelihood ratio test is used as the initial value to refine the reference center pixel, and then small intervals and high-quality observations are used to iteratively control the type II error to minimize the number of heterogeneous points in the homogeneous point set. Among them, the likelihood ratio test is a statistical hypothesis testing method, which is mainly used to compare the goodness of fit of two different models. Its basic principle is to compare the likelihood function values ​​of the two models. The likelihood function measures the degree of fit between the model and the data. The larger its value, the better the fit between the model and the data. Type II error (also called β error) refers to the probability of incorrectly accepting the null hypothesis when the null hypothesis is not true. "Using small intervals and high-quality observations to iteratively control type II error" refers to controlling the division of small intervals and reducing type II errors during the iterative processing of high-quality observation data.

[0067] Specifically, the initial window size is controlled, and the initial homogeneous point set is selected only from the pixels directly and adjacent to the reference center pixel, that is, a 5×5 estimation window is defined to obtain the initial homogeneous point set Ω init .

[0068] The initial set Ω init The intensity of all pixels within the reference center pixel is averaged to obtain the estimated value According to formula (1), all homogeneous points in a larger estimation window (15×15) can be selected.

[0069]

[0070] Where P{·} is the probability, L is the number of views, and N is the total number of time series samples. and are the time series means of the reference central pixel and the pixel to be estimated in the window, respectively.

[0071] Alternatively, in the case of small and medium samples, such as a short time series SAR data set, the Gaussian assumption cannot approximate the true distribution, so according to the characteristic that the SAR data intensity obeys the exponential distribution, the limitation of the Gaussian assumption is removed by accumulating and obeying the gamma distribution G to obtain a more accurate interval.

[0072]

[0073] The exponential distribution has a mean and standard deviation of θ. i ,i=1,...,N, the cumulative sum follows the gamma distribution, that is, Its N times sample mean It also obeys the gamma distribution, make According to the characteristics of the gamma function, X obeys the standard gamma distribution with a scale parameter of N, X~G(N,1).

[0074] α quantile g of the gamma distribution α / 2；N and the sample size N are known, the confidence level interval can be fixed once it is given:

[0075]

[0076] Gamma distribution is a positively skewed distribution, and the interval values ​​are asymmetric, so the tail is taken into account. When selecting homogeneous points, the sample mean in formula (3) is and θ are the time series average values ​​of the pixel intensity to be estimated in the large window mentioned in formula (1). and reference center pixel estimation (5×5 initial estimation window, initial set Ω init The time series mean of all pixel intensities in ), and the number of samples N is the number of time series samples.

[0077] (13) Set a homogeneous point number threshold, extract homogeneous points with a homogeneous point number greater than the threshold as DS points, extract the SDFP points of the slow decoherence filter phase, and merge the SDFP points with the DS points.

[0078] The differential interferometer phase set is homogeneously filtered according to the homogeneous point set to obtain the differential interferometer phase set after homogeneous phase optimization.

[0079] A homogeneous point number threshold is set according to the homogeneous point set, and points with homogeneous point numbers greater than the threshold are preliminarily selected as DS points. Spectral filtering in the range / azimuth direction is performed to obtain the DS preliminary point set DSC.

[0080] The differential interferometric phase set after homogeneous phase optimization is filtered in range and azimuth directions, and the SDFP points are preliminarily screened according to the amplitude difference threshold method to obtain the SDFP preliminary point set SDFPC.

[0081] BP-LP combined filtering is performed on SDFPC and DSC respectively to estimate the temporal coherence of each point in SDFPC and DSC. BP-LP combined filtering refers to a filtering method that combines the characteristics of band-pass (BP) filtering and low-pass (LP) filtering to extract and retain useful phase information while suppressing noise and unnecessary frequency components; BP-LP combined filtering first uses band-pass filtering to extract phase information within the frequency range of interest, and then uses low-pass filtering to smooth the extracted phase to further reduce the impact of noise.

[0082] Repeat the above SDFP point screening method to select all SDFP points from SDFPC and screen DS points from DSC.

[0083] Finally, the SDFP points and DS points are merged.

[0084] Based on the 3-D phase unwrapping algorithm, the entangled phase of the SDFP / DS fusion point set is unwrapped, the DEM error, SCLA / AOE are estimated and removed based on the phase stability, and the atmospheric delay phase is estimated and removed through spatiotemporal filtering, and finally the deformation rate map is obtained. Among them, the 3-D phase unwrapping algorithm refers to an algorithm that adds the time dimension (that is, multiple images of the same object) on the basis of 2D phase unwrapping based on a single image, and also incorporates the time dimension data into its mathematical model. DEM refers to the abbreviation of Digital Elevation Model, which is a digital geographic data model used to describe the elevation information of the earth's surface. SCLA (Space Charge Limited Accumulation) refers to the limited accumulation of space charge, which involves the charge accumulation process of the radar receiver, which affects the strength and quality of the signal, thereby indirectly affecting the accuracy of InSAR interferometry, and needs to be reduced or removed during data processing. AOE (Atmospheric Offset Error) refers to the atmospheric offset error. Changes in atmospheric conditions (such as temperature, humidity, and air pressure) can affect the propagation path and phase of radar signals, thereby introducing errors into interferometric measurements, which need to be removed. Temporal and spatial filtering refers to filtering the interferogram in time (for the same pixel in different images) and space (for different pixels in the same image).

[0085] In order to verify that the regional deformation rate map obtained by the method of the present invention has high quality, this embodiment uses the StaMPS-SBAS method, the StaMPS method fused with DS targets, the SBAS method, and the StaMPS-SBAS method fused with DS targets to process two different landslides respectively, and compares the regional deformation rate maps obtained by different methods, such as Figure 2 As shown, (a) to (d) are the Lagu Village landslide, and (g) to (h) are the Dahua Village landslide. Figure 2It can be seen from the comparison of the regional deformation rate graphs in (g) that for landslide deformation measurement, the number of observation points of the traditional StaMPS-SBAS method is very limited, and it is impossible to obtain effective data of the landslide body; in the StaMPS method with DS target fusion, DSC is subjected to homogeneous point phase filtering, and the phase stability of DS points is improved, which can fill the holes in the StaMPS observation results to a certain extent, but it is still impossible to obtain effective data of the entire landslide body; the more traditional SBAS method can monitor the landslide body, but for the Dahua Village landslide in (g), there are still a large number of hole values, and it is impossible to provide an accurate range in the subsequent automatic identification; and the StaMPS-SBAS method with DS target fusion proposed in the present invention can clearly and continuously show the landslide body deformation rate results. Therefore, in this embodiment, the StaMPS-SBAS method with DS target fusion is very suitable for the refined deformation measurement of small-scale, low-coherence complex scenes such as single landslides or small-area landslide clusters.

[0086] (2) Collect optical images that are consistent with the regional deformation rate map of the monitored area, and produce a regional deformation rate map that integrates the regional optical images. The regional optical image map is obtained by an optical imaging satellite, and the collection time must be consistent with the regional deformation rate map. Figure 1 The format is TIFF.

[0087] Specifically, the regional deformation rate map within the range is matched with the optical image map to ensure that their corresponding timestamps are consistent, and on this basis, the ground features and surface feature lines with the same name are kept consistent, so that the deformation rate map and the optical image map are accurately aligned; finally, the deformation rate map is used as an additional channel to generate a regional deformation rate map that integrates the optical image of the area.

[0088] (3) Integrate the regional deformation rate map and the regional deformation rate map fused with regional optical effects to create a data set for training, and input the data set into the deep learning model to train the landslide classification model; the landslide classification model needs to train two types of regional deformation rate maps and regional deformation rate maps fused with regional optical images during the training phase. When using the model, the corresponding data input can be selected according to needs.

[0089] Specifically, the specific steps of inputting the data set into the deep learning model to train the landslide classification model are: dividing the data set into a training set, a validation set, and a test set according to a preset ratio;

[0090] Use the training set to train a deep learning model with the self-attention mechanism as the core, and record the model parameters when the loss function value of the validation set reaches the minimum;

[0091] The trained deep learning model is used to predict the data of the test set to verify the accuracy of the deep learning model, and the deep learning model whose accuracy meets the preset requirements is retained as the landslide classification model.

[0092] According to the actual situation of landslide, landslide can be divided into the following two types:

[0093] (a) Landslides are scattered and there is no need to extract boundaries. The landslide classification model takes the regional deformation rate map as input and performs target detection.

[0094] (b) The landslide situation is complex and the boundary needs to be extracted. The landslide classification model takes the regional deformation rate map fused with the regional deformation rate map as input and performs instance segmentation.

[0095] (32) Label the deformation rate map and the regional deformation rate map of the fused regional optical image to obtain a labeled instance segmentation module training set and a target detection module training set.

[0096] Specifically, the basic category labels are set in advance, and the deformation rate map of the optical image of the fused area to be labeled is imported into the Labelme software; the boundary of the deformation area of ​​the hidden danger point is drawn according to the manual interpretation and field investigation results, and the category is assigned to complete the production of the instance segmentation module training set; then, the target box is drawn only on the regional deformation rate map and the category is assigned to complete the production of the target detection module training set.

[0097] (33) The instance segmentation module training set and the target detection module training set are input into the deep learning model training to obtain the landslide classification model.

[0098] Specifically, the target detection module is built based on the open source framework Faster-RCNN and is trained using a regional deformation rate map. The model parameters when the loss function is minimized are recorded, and the test set data is predicted to verify the model accuracy, and finally the model that meets the accuracy requirements is saved. Preferably, a four-step alternating iterative training method (4-Step Alternating Training) or an end-to-end approximate joint training method (Approximate joint training) is used. The four-step alternating iterative training method (4-Step Alternating Training) is a training strategy used in FasterR-CNN, including training RPN, training FastR-CNN, tuning RPN, and tuning FastR-CNN. Through this four-step alternating iterative training method, FasterR-CNN can continuously optimize the performance of RPN and RCNN on the basis of shared convolutional features, thereby achieving high-precision target detection. End-to-end approximate joint training is a deep learning method whose core concept is to closely combine multiple parts of the system (for example, acoustic model and language model) for training to optimize the performance of the entire system. In end-to-end training, each model or module is no longer considered as an independent unit, but as a whole, and the overall performance is maximized by adjusting the parameters of the entire system. The key to this method is that it considers the output of the entire system, not just the output of a single model, so it can optimize system performance more comprehensively.

[0099] Furthermore, the Faster-RCNN network used in this embodiment has a specific framework diagram as shown in Figure 3 As shown in the figure, the convolution layer is implemented by combining convolution + ReLU activation function + pooling. Specifically, 13 convolution layers, 13 ReLU activation function layers and 4 pooling layers are used to extract the feature map of the image. The RPN network layer generates the target frame based on the feature map and transmits it to the ROI pooling layer for comprehensive processing, and then transmits the result to the fully connected layer for classification and target frame refinement.

[0100] Specifically, the instance segmentation module is built with a deep learning model with self-attention mechanism as the core, and the deformation rate map of fused optical images is used for training.

[0101] Furthermore, as shown in the specific framework diagram 4, the deep learning model with the self-attention mechanism as the core will be built based on DANet. In the deformation rate map of the optical image of the fusion area, the deformation rate map will be regarded as an additional channel of the fusion map. The deep learning model with the self-attention mechanism as the core first uses the dilated residual network (DRN) to extract the features in the deformation rate map, and then passes through the convolution layer and passes it into the position attention mechanism module and the channel attention concentration module with the self-attention mechanism as the core, respectively, to obtain the connection between the features at different positions and the interdependence between channels. Finally, the output results of the two attention modules are element-wise summed to complete the feature fusion, and a fully connected layer is used to convert the position and channel information into the final result. Among them, the dilated residual network (DRN) is a special neural network structure that aims to optimize image processing and recognition performance by combining dilated convolution and residual connection. Among them, dilated convolution, also known as hollow convolution or atrous convolution, is a key component of the dilated residual network. Compared with traditional convolution operations, dilated convolution introduces holes between convolution kernels, allowing the convolution kernel to skip some pixels during the convolution process. This design allows dilated convolution to increase the receptive field while keeping the number of parameters and computational complexity unchanged, allowing the network to capture a wider range of contextual information. Secondly, residual connection is the core idea of ​​residual network. In deep neural networks, as the number of network layers increases, the gradient vanishing and representation bottleneck problems will gradually become prominent, resulting in a decrease in network performance. Residual connection introduces a "short circuit" mechanism to connect the input directly to the output, making it easier for the network to optimize the residual, that is, the difference between the input and the output, during learning. This design not only helps to alleviate the gradient vanishing problem, but also improves the training speed and stability of the network. The dilated residual network combines dilated convolution and residual connection, using dilated convolution to capture a wider range of contextual information, while maintaining the network training efficiency through residual connection. This structure enables the dilated residual network to perform well in tasks such as image segmentation, object detection, and image classification, especially when deep features of images need to be extracted.

[0102] (4) The InSAR deformation rate map to be interpreted is input into the landslide classification model, and the landslide classification model outputs the landslide location and landslide classification of the InSAR deformation rate map. When the landslide situation of the InSAR deformation rate map is complex, the landslide classification model also extracts the deformation area boundary. Among them, the landslide classification is mainly classified according to the scale of the landslide, and is divided into small landslides, medium landslides, large landslides, and extra-large landslides; preferably, in actual application, according to the use requirements, the regional deformation rate information obtained from the InSAR deformation rate map can also be used to classify the landslide into creeping landslides, slow landslides, medium-speed landslides, and high-speed landslides according to the sliding speed of the landslide.

[0103] Example 2

[0104] The InSAR-based landslide hazard point identification and classification system of the present invention comprises an input module, a conversion processing module, a fusion module, a data set acquisition module, a training module and an identification and classification module.

[0105] Specifically, the input module is used to input InSAR data.

[0106] The conversion processing module can receive the InSAR data of the input module and process it using the StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets to obtain the regional deformation rate map.

[0107] Specifically, the single-look complex images of InSAR data are firstly registered by adopting the enhanced spectral diversity method, and the time series of the intensity images of the single-look complex images are extracted; the SRTMDEM data are registered with the super master image and simulated as the interference phase; under the setting of the spatiotemporal baseline threshold, a differential interferometric combination is constructed, and the interference operation is performed on the registered SLC image sequence; the terrain phase simulated by the DEM data is removed to obtain the differential interferometric phase set.

[0108] Then, ground points with similar or consistent scattering characteristics in the differential interferometer phase set are selected as homogeneous points to obtain a homogeneous point set. The homogeneous point set based on the likelihood ratio test is used as the initial value to refine the reference center pixel, and then small intervals and high-quality observations are used to iteratively control the type II error to minimize the number of heterogeneous points in the homogeneous point set.

[0109] Set the threshold of homogeneous points, extract homogeneous points with a number greater than the threshold as DS points, extract the SDFP points of the slowly decoherent filtering phase, and merge the SDFP points with the DS points. Perform homogeneous filtering on the differential interferometric phase set according to the homogeneous point set to obtain the differential interferometric phase set after homogeneous phase optimization; set the threshold of homogeneous points according to the homogeneous point set, preliminarily select the points with a number greater than the threshold as DS points, and perform range / azimuth spectrum filtering to obtain the DS preliminary point set DSC; perform range / azimuth spectrum filtering on the differential interferometric phase set after homogeneous phase optimization, preliminarily screen the SDFP points according to the amplitude difference threshold method, and obtain the SDFP preliminary point set SDFPC; perform BP-LP combined filtering on SDFPC and DSC respectively, and estimate the temporal coherence of each point in SDFPC and DSC; repeat the above SDFP point screening method to select all SDFP points from SDFPC and screen DS points from DSC.

[0110] Based on the 3-D phase unwrapping algorithm, the entangled phase of the SDFP / DS fusion point set is unwrapped, the DEM error and SCLA / AOE are estimated and removed according to the phase stability, and the atmospheric delay phase is estimated and removed through spatiotemporal filtering, and finally the deformation rate map is obtained.

[0111] The fusion module can collect optical image maps that are consistent in time with the regional deformation rate map of the monitored area, and produce a regional deformation rate map of the fused area optical images.

[0112] Specifically, the regional deformation rate map within the range is matched with the optical image map to ensure that their corresponding timestamps are consistent, and on this basis, the ground features and surface feature lines with the same name are kept consistent, so that the deformation rate map and the optical image map are accurately aligned; finally, the deformation rate map is used as an additional channel to generate a regional deformation rate map that integrates the optical image of the area.

[0113] The data set acquisition module can integrate the regional deformation rate map and the regional deformation rate map of the fusion area optical effect to produce a data set for training. Specifically, according to the actual situation of the landslide, the landslide is divided into the following two types:

[0114] (a) Landslides are scattered and there is no need to extract boundaries. The landslide classification model takes the regional deformation rate map as input and performs target detection. The deformation rate map is annotated to obtain the annotated instance segmentation module training set.

[0115] (b) The landslide situation is complex and the boundary needs to be extracted. The landslide classification model takes the regional deformation rate map of the fused regional deformation rate map as input and performs instance segmentation. The regional deformation rate map of the fused regional optical image is labeled to obtain a labeled target detection module training set.

[0116] Set the basic category labels in advance, and import the deformation rate map of the optical image of the fused area to be labeled on the Labelme software; draw the boundary of the deformation area of ​​the hidden danger point according to the manual interpretation and field investigation results, and assign categories to complete the production of the instance segmentation module training set; then, draw the target box only on the regional deformation rate map and assign categories to complete the production of the target detection module training set.

[0117] The training module can use the training set to train the deep learning model and obtain the landslide classification model. Specifically, the data set is divided into a training set, a validation set, and a test set according to a preset ratio; the training set is used to train the deep learning model with the self-attention mechanism as the core, and the model parameters when the loss function value of the validation set reaches the minimum are recorded; the trained deep learning model is used to predict the data of the test set to verify the accuracy of the deep learning model, and the deep learning model whose accuracy meets the preset requirements is retained as the landslide classification model.

[0118] The recognition and classification module decodes the InSAR deformation rate map based on the landslide classification model and outputs the landslide location and landslide classification of the InSAR deformation rate map. Specifically, when the landslide situation of the input InSAR deformation rate map is complex, the landslide classification model also extracts the deformation area boundary.

Claims

1. A landslide hazard point identification and classification method based on InSAR, characterized by: The following steps are included: (1) The regional InSAR data were processed based on the StaMPS-SBAS time-series InSAR processing method that integrates distributed scatterer targets to obtain regional deformation rate maps; (2) Collecting optical images that are consistent with the regional deformation rate map of the monitored area, and producing a regional deformation rate map that integrates the optical images of the area; (3) Integrate the regional deformation rate map and the regional deformation rate map with regional optical influence to create a training data set, and input the data set into the deep learning model to train the landslide classification model; Landslides are scattered and there is no need to extract boundaries. The landslide classification model uses the regional deformation rate map as input and performs target detection. Landslides are complex and need to extract boundaries. The landslide classification model uses the regional deformation rate map fused with the regional deformation rate map as input and performs instance segmentation. Annotate the deformation rate map and the regional deformation rate map of the fused regional optical image to obtain annotated instance segmentation module training set and target detection module training set; The instance segmentation module training set and the target detection module training set are input into the deep learning model training to obtain the landslide classification model; (4) The InSAR deformation rate map to be interpreted is input into the landslide classification model, and the landslide classification model outputs the landslide location and landslide classification of the InSAR deformation rate map. When the landslide situation of the InSAR deformation rate map is complex, the landslide classification model also extracts the deformation area boundary.

2. The InSAR-based landslide hazard point identification and classification method according to claim 1, characterized in that: The specific steps of the StaMPS-SBAS time-series InSAR processing method based on fusion of distributed scatterer targets described in step (1) include: Preprocess the regional InSAR data to obtain the differential interferometric phase set; Selecting ground points with similar or consistent scattering characteristics in the differential interference phase set as homogeneous points to obtain a set of homogeneous points; Set a homogeneous point number threshold, extract homogeneous points with a homogeneous point number greater than the threshold as DS points, extract SDFP points of the slow decoherence filter phase, and merge the SDFP points with the DS points; Based on the 3-D phase unwrapping algorithm, the entangled phase of the SDFP / DS fusion point set is unwrapped, the DEM error and SCLA / AOE are estimated and removed according to the phase stability, and the atmospheric delay phase is estimated and removed through spatiotemporal filtering, and finally the deformation rate map is obtained.

3. The InSAR-based landslide hazard point identification and classification method according to claim 2 is characterized in that: The specific steps for preprocessing regional InSAR data and obtaining the differential interferometric phase set are as follows: By adopting enhanced spectral diversity method, the single-view complex images of InSAR data are registered and the time series of intensity images of single-view complex images are extracted. The previously acquired SRTMDEM data are registered with the selected super master image and simulated as an interferometric phase; Under the setting of the temporal and spatial baseline threshold, a differential interferometric combination is constructed to perform interferometric operations on the registered SLC image sequence. The terrain phase simulated by the DEM data is removed to obtain the differential interferometric phase set.

4. The InSAR-based landslide hazard point identification and classification method according to claim 2 is characterized in that: The specific steps to obtain a set of homogeneous points are as follows: The homogeneous point set based on the likelihood ratio test is used as the initial value to refine the reference center pixel; Small intervals and high-quality observations are used to iteratively control type II errors and minimize the number of heterogeneous points in a set of homogeneous points.

5. The method for identifying and classifying landslide potential risk points based on InSAR according to claim 2, characterized in that: The specific steps for extracting DS points and SDFP points are as follows: Performing homogeneous filtering on the differential interferometric phase set according to the homogeneous point set, and obtaining the differential interferometric phase set after homogeneous phase optimization; A homogeneous point number threshold is set according to the homogeneous point set, and points with homogeneous point numbers greater than the threshold are initially selected as DS points, and then spectral filtering in the range / azimuth direction is performed to obtain the DS preliminary point set DSC; After homogeneous phase optimization, the differential interferometric phase set is filtered in range and azimuth directions, and the SDFP points are preliminarily screened according to the amplitude difference threshold method to obtain the SDFP preliminary point set SDFPC. Perform BP-LP combined filtering on SDFPC and DSC respectively to estimate the temporal coherence of each point in SDFPC and DSC; Repeat the above SDFP point screening method to select all SDFP points from SDFPC and screen DS points from DSC.

6. The method for identifying and classifying landslide potential risk points based on InSAR according to claim 1, characterized in that: The specific steps of inputting the data set into the deep learning model to train the landslide classification model in step (3) are as follows: Divide the data set into training set, validation set and test set according to the preset ratio; Use the training set to train a deep learning model with the self-attention mechanism as the core, and record the model parameters when the loss function value of the validation set reaches the minimum; The trained deep learning model is used to predict the data of the test set to verify the accuracy of the deep learning model, and the deep learning model whose accuracy meets the preset requirements is retained as the landslide classification model.

7. The InSAR-based landslide hazard point identification and classification method according to claim 6 is characterized in that: The object detection module is built based on the open source framework Faster-RCNN and is trained using regional deformation rate maps.

8. The InSAR-based landslide hazard point identification and classification method according to claim 6 is characterized by: The instance segmentation module is constructed with a deep learning model based on the self-attention mechanism, and uses a deformation rate map of fused optical images for training.

9. A landslide hazard point identification and classification system based on InSAR, characterized by: include, Input module, used to input InSAR data; The conversion processing module is used to receive the InSAR data from the input module and process it using the StaMPS-SBAS time-series InSAR processing method based on the fusion of distributed scatterer targets to obtain a regional deformation rate map; A fusion module is used to obtain an optical image map that is consistent with the regional deformation rate map of the monitored area in time, and to produce a regional deformation rate map of the optical image of the fused area; A data set acquisition module, used to integrate the regional deformation rate map and the regional deformation rate map integrating the regional optical influence, and to produce a data set for training; The training module uses the dataset to train the deep learning model and obtain the landslide classification model; The recognition and classification module decodes the InSAR deformation rate map based on the landslide classification model and outputs the landslide location and landslide classification of the InSAR deformation rate map.

10. The InSAR-based landslide hazard point identification and classification system according to claim 9, characterized in that: The data sets acquired by the data set acquisition module include the following two types: Landslides are scattered and there is no need to extract boundaries. The landslide classification model uses the regional deformation rate map as input and performs target detection. The deformation rate map is annotated to obtain the annotated instance segmentation module training set. Landslides are complex and need to be extracted from the boundary. The landslide classification model takes the regional deformation rate map fused with the regional deformation rate map as input and performs instance segmentation; Annotate the regional deformation rate map of the optical image of the fusion area to obtain the labeled target detection module training set; The method of using the data set to train the deep learning model to obtain the landslide classification model includes: dividing the data set into a training set, a validation set, and a test set according to a preset ratio; Use the training set to train a deep learning model with the self-attention mechanism as the core, and record the model parameters when the loss function value of the validation set reaches the minimum; The trained deep learning model is used to predict the data of the test set to verify the accuracy of the deep learning model, and the deep learning model whose accuracy meets the preset requirements is retained as the landslide classification model.

Images