A method and system for automatically detecting a surface landslide subsidence risk area

Abstract

The patent application provides a kind of surface landslide subsidence risk area automatic detection method and system, comprising: the satellite data of target area is collected;Based on satellite data, using small baseline set aperture radar interferometry technique, obtain the ground deformation rate graph of target area;Based on ground deformation rate graph, through pyramid scene analysis network, identify the surface landslide subsidence risk area in target area;The present application adopts small baseline set aperture radar interferometry technique is to satellite data is interferometric analysis, therefore can overcome interference signal, through the improved pyramid scene analysis network of the present application not only can the ground deformation rate extracted by small baseline set aperture radar interferometry technique be automatically identified, also can reduce the amount of calculation in the process of identifying ground subsidence risk area, therefore not only save the expense of each time please expert analysis, also help to improve the efficiency and reliability of large-scale landslide monitoring.

Description

TECHNICAL FIELD

[0001] The patent application belongs to the technical field of ground surface settlement detection, and particularly relates to a ground surface landslide settlement risk area automatic detection method and system. BACKGROUND

[0002] In a high-voltage power transmission system, a power transmission tower as one of the core infrastructures, its safe and stable operation state plays an important role in promoting the prosperity and development of regional economy. The power transmission tower is widely distributed in the mountainous area with complex terrain and variable geological conditions. These areas are mostly located on the rear edge of the slope or on the ridge, and face extremely complex geological structure and environmental challenges. In view of the frequent occurrence of geological disasters such as settlement and landslide, it poses a serious risk and potential threat to the long-term and effective operation of the power transmission network. Therefore, the systematic settlement hidden danger investigation and early identification work for the area along the power grid is not only important for the safe and stable operation of the power grid, but also has far-reaching academic and practical significance for preventing natural disasters from damaging key infrastructures and maintaining stable regional power supply.

[0003] In the field of settlement hidden danger identification, the traditional field reconnaissance assisted by classic measuring means such as level, global positioning system (GPS), and crack meter, although has a long history, its application scope is limited to a small area, and there are significant limitations in efficiency, cost and environmental adaptability, especially when dealing with high-position, long-range or hidden landslide settlement. With the development of optical remote sensing technology and interferometric synthetic aperture radar (InSAR) technology, these means are gradually becoming important tools for landslide and settlement hidden danger identification. Therefore, the field of settlement and landslide hidden danger identification is experiencing a transformation from traditional to modern high-tech means.

[0004] The interferometric synthetic aperture radar (InSAR) may introduce interference signals under natural environmental factors such as vegetation cover, atmospheric disturbance and complex terrain conditions, thereby affecting the accuracy of the detection results, and the monitoring capability is limited in specific situations. Moreover, the results of the interferometric synthetic aperture radar are usually analyzed by visual interpretation and manual editing method, which is low in efficiency and high in cost. SUMMARY

[0005] In order to overcome the deficiencies of the prior art, the patent application proposes a ground surface landslide settlement risk area automatic detection method, which comprises:

[0006] Satellite data of a target area is collected;

[0007] Based on the satellite data, a ground surface deformation rate map of the target area is obtained by using a small baseline set interferometric synthetic aperture radar (InSAR) technology.

[0008] based on the ground surface deformation rate map, a landslide subsidence risk area in the target area is identified by a pre-constructed improved pyramid scene analysis network;

[0009] The improved pyramid scene analysis network comprises an improved inverted residual layer, a multi-pyramid pooling structure, and a hollow pyramid pooling structure, and the improved inverted residual layer is a deep separable convolution structure based on point-by-point convolution combined with deep convolution.

[0010] Preferably, the construction of the improved pyramid scene analysis network comprises:

[0011] Based on the historical satellite data of the target area, a historical ground surface deformation rate map and a deformation label of the historical ground surface deformation rate map are obtained.

[0012] The historical ground surface deformation rate map and the deformation label of the historical ground surface deformation rate map are used as a sample set.

[0013] The sample set is divided into a training sample set and a verification sample set.

[0014] Based on the training sample set, the improved inverted residual layer, the multi-pyramid pooling structure, and the hollow pyramid pooling structure in the pyramid scene analysis network are iteratively trained to obtain a trained pyramid scene analysis network.

[0015] Based on the verification sample set, the trained pyramid scene analysis network is corrected to complete the construction of the pyramid scene analysis network.

[0016] Preferably, the historical satellite data of the target area is used to obtain the historical ground surface deformation rate map and the deformation label of the historical ground surface deformation rate map, which comprises:

[0017] Based on the historical satellite data of the target area, a small-baseline set aperture radar interferometry technology is used to obtain a historical ground surface deformation rate map.

[0018] The historical ground surface deformation rate map is subjected to deformation identification to obtain the deformation label of the historical ground surface deformation rate map.

[0019] Preferably, based on the training sample set, the improved inverted residual layer, the multi-pyramid pooling structure, and the hollow pyramid pooling structure in the pyramid scene analysis network are iteratively trained to obtain a trained pyramid scene analysis network, which comprises:

[0020] The historical ground surface deformation rate map in the training sample set is inputted, efficient feature compression is performed on the improved inverted residual layer, and a deformation rate feature vector is obtained.

[0021] Based on the deformation rate feature vector, multi-scale image feature information is obtained through a multi-pyramid pooling structure and a hollow-pyramid pooling structure respectively, and a deformation rate fusion feature map is generated according to the multi-scale image feature information;

[0022] The deformation rate fusion feature map is up-sampled by using a bilinear interpolation method to obtain a predicted deformation label of the historical ground surface deformation rate map.

[0023] The predicted deformation label of the historical ground surface deformation rate map is compared with a deformation label corresponding to a historical ground surface deformation rate map in the training sample set, and the weights of the improved inverted residual layer, the multi-pyramid pooling structure and the hollow-pyramid pooling structure are updated according to the comparison result, and the number of iterations is iterated until the number of training iterations reaches a pre-set number of training iterations, and the training of the pyramid scene analysis network is completed.

[0024] Preferably, the point-by-point convolution includes a first point-by-point convolution and a second point-by-point convolution; and the historical ground surface deformation rate map in the training sample set is inputted to obtain a deformation rate feature vector through efficient feature compression of the improved inverted residual layer, including:

[0025] Based on the historical ground surface deformation rate map in the training sample set, the first point-by-point convolution is performed to obtain a first point-by-point convolution feature map.

[0026] Based on the first point-by-point convolution feature map, a deep convolution is performed to obtain a deep convolution feature map; and based on the deep convolution feature map, the second point-by-point convolution is performed to obtain a deformation rate feature vector.

[0027] Preferably, based on the satellite data, a small-baseline set aperture radar interferometry technology is used to obtain the ground surface deformation rate map of the target region, including:

[0028] A plurality of synthetic aperture radar images are selected from the satellite data.

[0029] The plurality of synthetic aperture radar images are combined and paired two by two using pre-set time and space baseline limits to generate a plurality of small baseline sets.

[0030] A baseline threshold is used to generate an unwrapped interferogram corresponding to each small baseline set.

[0031] After the interference phase unwrapping of each unwrapped interferogram, the obtained unwrapped interference phase is composed into an interference phase matrix, and a least square method is used to obtain a ground surface deformation rate based on the interference phase matrix.

[0032] Based on the ground surface deformation rate, a ground surface deformation rate map is generated.

[0033] Preferably, after the interference phase unwrapping of each unwrapped interferogram, the obtained unwrapped interference phases are composed into an interference phase matrix, and based on the interference phase matrix, a least square method is used to obtain the surface deformation rate, including:

[0034] Based on the unwrapped interference phase related parameters, an unwrapped interference phase calculation formula is used to calculate the unwrapped interference phase of each unwrapped interferogram.

[0035] According to all the unwrapped interference phases, an interference phase matrix is generated.

[0036] A least square method is used for inversion to extract a surface deformation time sequence from the interference phase matrix, and a surface deformation rate is obtained according to the surface deformation time sequence.

[0037] Preferably, the unwrapped interference phase calculation formula is as follows:

[0038]

[0039] Where, Δφ j is the unwrapped interference phase of the jth unwrapped interferogram, IU j is the start index of the interferometric pair measurement time period of the jth unwrapped interferogram, IE j is the end index of the interferometric pair measurement time period of the jth unwrapped interferogram, k is the measurement time of the interferometric pair measurement time period, t k is the kth time in the interferometric pair measurement time period, t k-1 is the k-1th time in the interferometric pair measurement time period, v k is the surface velocity deformation rate at t k .

[0040] Based on the same inventive concept, the present patent application also provides a surface landslide subsidence risk area automatic detection system, including: a satellite data acquisition module, a surface deformation rate map extraction module and a surface landslide subsidence risk area identification module.

[0041] The satellite data acquisition module is used to acquire satellite data of a target area.

[0042] The surface deformation rate map extraction module is used to obtain a surface deformation rate map of the target area based on the satellite data by using a small baseline set aperture radar interferometric measurement technology.

[0043] The surface landslide subsidence risk area identification module is used to identify a surface landslide subsidence risk area in the target area based on the surface deformation rate map by using a pre-constructed improved pyramid scene analysis network.

[0044] The improved pyramid scene analysis network comprises an improved inverted residual layer, a multi-pyramid pooling structure and a hollow pyramid pooling structure.

[0045] Preferably, the system further comprises an improved pyramid scene analysis network construction module, wherein the improved pyramid scene analysis network construction module comprises:

[0046] A historical satellite data processing submodule is configured to obtain a historical ground surface deformation rate map and a deformation label of the historical ground surface deformation rate map based on historical satellite data of a target region.

[0047] A sample set generation submodule is configured to take the historical ground surface deformation rate map and the deformation label of the historical ground surface deformation rate map as a sample set.

[0048] A sample set division submodule is configured to divide the sample set into a training sample set and a verification sample set.

[0049] An iterative training submodule is configured to perform iterative training on the improved inverted residual layer, the multi-pyramid pooling structure and the hollow pyramid pooling structure in the pyramid scene analysis network based on the training sample set, and obtain a trained pyramid scene analysis network.

[0050] A model correction submodule is configured to correct the trained pyramid scene analysis network based on the verification sample set, and complete construction of the pyramid scene analysis network.

[0051] Preferably, the historical satellite data processing submodule is specifically configured to:

[0052] The historical satellite data processing submodule is configured to obtain the historical ground surface deformation rate map based on historical satellite data of the target region by using a small-baseline set aperture radar interferometry technology.

[0053] The historical satellite data processing submodule is configured to perform deformation identification on the historical ground surface deformation rate map, and obtain the deformation label of the historical ground surface deformation rate map.

[0054] Preferably, the iterative training submodule comprises:

[0055] An inverted residual layer training unit is configured to take a historical ground surface deformation rate map in the training sample set as input, perform efficient feature compression on the historical ground surface deformation rate map by using the improved inverted residual layer, and obtain a deformation rate feature vector.

[0056] A deformation rate fusion feature map generation unit is configured to obtain multi-scale image feature information by using the multi-pyramid pooling structure and the hollow pyramid pooling structure based on the deformation rate feature vector, and generate a deformation rate fusion feature map according to the multi-scale image feature information.

[0057] an upsampling unit, configured to perform upsampling on the deformation rate fusion feature map by using a bilinear interpolation method to obtain a predicted deformation label of the historical ground surface deformation rate map;

[0058] a structure parameter updating unit, configured to compare the predicted deformation label of the historical ground surface deformation rate map with a deformation label corresponding to a historical ground surface deformation rate map in a training sample set, and update weights of the improved inverted residual layer, the multi-pyramid pooling structure and the dilated pyramid pooling structure according to a comparison result, and iteratively train a number of times until the number of training iterations reaches a pre-set number of training iterations, and complete training of the pyramid scene analysis network.

[0059] Preferably, the point-by-point convolution includes a first point-by-point convolution and a second point-by-point convolution; and the inverted residual layer training unit is specifically configured to:

[0060] expand a deformation rate feature based on the historical ground surface deformation rate map in the training sample set by using the first point-by-point convolution to obtain a first point-by-point convolution feature map;

[0061] perform a convolution operation on the first point-by-point convolution feature map by using the deep convolution to obtain a deep convolution feature map;

[0062] obtain a deformation rate feature vector based on the deep convolution feature map by using the second point-by-point convolution.

[0063] Preferably, the satellite data acquisition module includes:

[0064] a data screening sub-module, configured to screen a plurality of synthetic aperture radar images from the satellite data;

[0065] a small baseline set generation sub-module, configured to combine and pair the plurality of synthetic aperture radar images two by two by using pre-set time and space baseline limit values to generate a plurality of small baseline sets;

[0066] an unwrapped interferogram generation sub-module, configured to generate an unwrapped interferogram corresponding to each small baseline set by using a pre-set baseline threshold value;

[0067] a ground surface deformation rate extraction sub-module, configured to, after performing interferometric phase unwrapping on each unwrapped interferogram, compose an interferometric phase matrix from the obtained unwrapped interferometric phase, and obtain a ground surface deformation rate by using a least square method based on the interferometric phase matrix;

[0068] a ground surface deformation rate map generation sub-module, configured to generate a ground surface deformation rate map based on the ground surface deformation rate.

[0069] Preferably, the ground surface deformation rate extraction sub-module is specifically configured to:

[0070] Based on the unwrapped interference phase correlation parameters, an unwrapped interference phase calculation formula is used to calculate the unwrapped interference phase of each of the unwrapped interferograms;

[0071] According to all of the unwrapped interference phases, an interference phase matrix is generated;

[0072] A least square method is used to perform inversion to extract a surface deformation time sequence from the interference phase matrix, and a surface deformation rate is obtained according to the surface deformation time sequence.

[0073] Preferably, the unwrapped interference phase calculation formula is as follows:

[0074]

[0075] wherein, Δφ j is the unwrapped interference phase of the jth unwrapped interferogram, IU j is the start index of the interference pair measurement time period of the jth unwrapped interferogram, IE j is the end index of the interference pair measurement time period of the jth unwrapped interferogram, k is the measurement time of the interference pair measurement time period, t k is the kth time in the interference pair measurement time period, t k-1 is the k-1th time in the interference pair measurement time period, v k is the surface velocity deformation rate at t k .

[0076] Based on the same inventive concept, the present patent application also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected through a bus;

[0077] The memory is used for storing one or more programs;

[0078] When the one or more programs are executed by the at least one processor, a surface landslide subsidence risk area automatic detection method as described above is realized.

[0079] Based on the same inventive concept, the present patent application also provides a readable storage medium, which has a computer program stored thereon, and an execution program stored thereon, and the execution program is executed to realize a surface landslide subsidence risk area automatic detection method as described above.

[0080] Compared with the closest prior art, the present patent application has the beneficial effects as follows:

[0081] The patent application provides a Sichuan-Tibet line area detection method and system, which comprises the following steps: collecting satellite data of a target area; obtaining a ground surface deformation rate map of the target area by adopting a small baseline set aperture radar interferometric measurement technology based on the satellite data; and identifying a ground surface landslide subsidence risk area in the target area by using a pre-constructed improved pyramid scene analysis network based on the ground surface deformation rate map; wherein the improved pyramid scene analysis network comprises an improved inverted residual layer, a multi-pyramid pooling structure and a hollow pyramid pooling structure, and the improved inverted residual layer is a deep separable convolution structure based on point-by-point convolution combined with deep convolution; the small baseline set aperture radar interferometric measurement technology is used to perform interferometric measurement analysis on satellite data, so that interference signals that may be introduced by natural environmental factors such as vegetation coverage, atmospheric disturbance and complex terrain conditions on the ground can be overcome, and high-precision monitoring of ground subsidence and landslide deformation can be performed by the baseline set aperture radar interferometric measurement technology; the improved pyramid scene analysis network can not only automatically identify the ground surface deformation rate extracted by the small baseline set aperture radar interferometric measurement technology, but also reduce the calculation amount in the process of identifying the ground surface subsidence risk area, so that the cost of expert analysis each time is saved, and the efficiency and reliability of large-scale landslide monitoring are improved. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 A ground surface landslide subsidence risk area automatic detection method flowchart is provided for the patent application.

[0083] Figure 2 A training loss diagram of an improved PSPNet model is provided for the patent application.

[0084] Figure 3 A sample diagram of a historical ground surface deformation rate map and a deformation label is provided for the patent application.

[0085] Figure 4 A MobilenetV2 structure diagram is provided for the patent application.

[0086] Figure 5 A general technical route diagram of the patent application is provided for the patent application.

[0087] Figure 6 A research area and a Sentinel-1 data coverage range diagram is provided for the patent application.

[0088] Figure 7 A deformation rate map of the research area from January 2020 to March 2023 obtained based on an SBAS-InSAR technology is provided for the patent application.

[0089] Figure 8 A schematic diagram of an automatic detection system for a surface landslide subsidence risk area is provided for the patent application.

[0090] Figure 9 A structural schematic diagram of an electronic device is provided for the patent application. DETAILED DESCRIPTION

[0091] The specific embodiments of the patent application will be further described in detail below with reference to the accompanying drawings.

[0092] Example 1:

[0093] An automatic detection method for a surface landslide subsidence risk area provided by the patent application is as shown in Figure 1 , comprising:

[0094] Step 1: Collect satellite data of the target area;

[0095] Step 2: Based on the satellite data, use the small baseline set aperture radar interferometric measurement technology to obtain the surface deformation rate map of the target area;

[0096] Step 3: Based on the surface deformation rate map, identify the surface landslide subsidence risk area in the target area through the pre-constructed improved pyramid scene analysis network;

[0097] Among them, the improved pyramid scene analysis network includes an improved inverted residual layer, a multi-pyramid pooling structure and a hollow pyramid pooling structure, and the improved inverted residual layer is a deep separable convolution structure based on point-by-point convolution combined with deep convolution.

[0098] InSAR technology, especially when monitoring the rapid deformation of the target area, is prone to cause loss of coherence, which limits the range of effective monitoring. The internal constraints of InSAR technology lead to uneven distribution of generated interference points, making it difficult to fully cover the study area and affecting data integrity and analysis depth. Therefore, the accurate interpretation of landslide deformation usually needs to combine expert knowledge and consider deformation characteristics. In previous studies, the lack of efficient automated interpretation methods still hinders the application of InSAR technology in large-area landslide detection. Simple detection methods such as threshold method are easily affected by InSAR technology measurement noise, and are prone to false positives, therefore, it is of great significance to develop an automatic method to identify subsidence landslide risk, which helps to improve the efficiency and reliability of large-scale landslide monitoring.

[0099] With the rapid development of computer vision, deep learning techniques such as Convolutional Neural Networks (CNN) have made outstanding achievements in landslide automatic identification. In recent years, many studies have combined CNN with optical images to detect landslides and have achieved gratifying results. These methods can be divided into object detection-based methods and semantic segmentation-based methods. Object detection-based methods use bounding boxes to locate landslides. For example, Cheng et al. (2021) combined the YOLO-V4 (You Only Look Once-Version 4, a target detection algorithm) model and attention mechanism to detect landslides using optical remote sensing images from Google Earth Terrain Mapping. Tanatipuknon et al. (2021) found that combining two region-based RCNN (Regions with Convolutional Neural Networks) models with a simple decision tree could achieve better results in detecting landslides. Semantic segmentation-based methods (such as U-Net and FCN networks) can classify subsiding landslide and non-subsiding landslide pixels to delineate their boundaries. For example, Ghorbanzadeh et al. (2022) proposed a new strategy that combined rule-based object-based image analysis (OBIA) with a fully convolutional neural network (FCN) to detect landslides from multi-temporal Sentinel-2 images. Lu et al. (2023) proposed a dual-encoder U-Net that used Sentinel-2 and Digital Elevation Model (DEM) data for landslide detection. Inspired by deep learning techniques in the field of optical remote sensing, researchers began to study the application of deep learning in the field of InSAR. There have been some studies using CNN and InSAR to monitor and detect geological disasters such as earthquakes, volcanoes, land subsidence, and mining subsidence. These studies show that CNN has great potential in automatically interpreting InSAR results and effectively detecting geological disasters. A recent study proposed a new method using CNN and phase gradient stacking to automatically identify active landslides without phase unwrapping. Phase gradient stacking can quickly obtain landslide displacement locations, but this method is susceptible to phase noise (caused by shadows, overlaps, or DEM errors), leading to false positives in landslide detection. Overall, there have been fewer studies on subsiding landslide automatic detection based on InSAR and CNN.

[0100] Therefore, in the present application, firstly, an improved pyramid scene analysis network is constructed, and the construction process is specifically as follows:

[0101] In an implementation manner, the construction of the improved pyramid scene analysis network comprises:

[0102] Based on the historical satellite data of the target region, a historical ground surface deformation rate map and a deformation label of the historical ground surface deformation rate map are obtained;

[0103] The historical ground surface deformation rate map and the deformation label of the historical ground surface deformation rate map are taken as a sample set;

[0104] In an example, a sample set of landslide image data is created by visual interpretation of InSAR results of a research region to train the improved PSPNet (Pyramid Scene Parsing Network).

[0105] The sample set is divided into a training sample set and a verification sample set;

[0106] In an example, in order to train the improved PSPNet, 70% of images in the sample set are randomly selected as the training sample set, 20% of images are set as the verification sample set, and 10% of images are selected as a test sample set.

[0107] Based on the training sample set, the improved inverted residual layer, the multi-pyramid pooling structure and the hollow pyramid pooling structure in the pyramid scene analysis network are iteratively trained to obtain a trained pyramid scene analysis network;

[0108] Based on the verification sample set, the trained pyramid scene analysis network is corrected to complete the construction of the pyramid scene analysis network.

[0109] In the training of the improved PSPNet, the initialization of weight values, bias values, and scale factor values of batch normalization is crucial. To ensure that the network can learn effectively, we initialize these parameters with a normal distribution. Specifically, the weight values are sampled from a normal distribution with a mean of 0 and a variance of 1, the bias values are generally initialized to 0, and the scale factor of batch normalization is set to 1, so that the performance of the network in the initial stage will not be excessively affected by improper initialization of parameters. Next, the initial learning rate of the network is set to 0.01, which is a common choice when using the Stochastic Gradient Descent (SGD) optimizer. The learning rate controls the step size of each parameter update, and a too high learning rate may lead to unstable convergence, while a too low learning rate may result in slow training speed. In the setting of batch size, the batch size is set to 8 in the initial stage, i.e., 8 samples are used for backpropagation and weight update each time the improved PSPNet model parameters are updated. Subsequently, all initialized parameter data is input into the network to start training. In terms of the optimizer, the Stochastic Gradient Descent (SGD) algorithm is used for parameter update, and its update expression is as follows:

[0110]

[0111] where W t+1 is the parameter at time t+1, W t is the parameter at time t, η is the learning rate, is the gradient of the loss function with respect to W t , and λ is the weight decay factor.

[0112] The optimizer can accelerate convergence through momentum, and the momentum parameter is set to 0.9. In addition, to further improve the efficiency of training, a learning rate reduction strategy is adopted. Here, the Cosine Annealing learning rate reduction method is chosen, i.e., the learning rate decreases gradually during the training process, and its variation can be represented as:

[0113]

[0114] where η t is the learning rate at time t, η min is the minimum learning rate, η max is the maximum learning rate, and T max is the maximum learning period.

[0115] This method can effectively avoid the improved PSPNet from falling into a local optimum in the later training stage, maintaining the stability and performance of the improved PSPNet; for example,Figure 2 The training loss diagram of the improved PSPNet is shown; from the given loss function curve, it can be seen that the improved PSPNet shows good convergence in the training process. The loss value (Loss) decreases significantly with the increase of training rounds (Epoch). In the early stage of training, the loss value decreases rapidly from close to 1.8, especially in the first 20 Epochs, indicating that the improved PSPNet has quickly learned effective features in the initial stage. With the continuation of training, the loss value tends to be stable, and basically remains at about 0.2 near the 50th epoch, indicating that the training process of the improved PSPNet model has tended to converge, and there is no obvious loss decrease. This loss change trend shows that the improved PSPNet model can effectively update parameters under the given optimizer (such as stochastic gradient descent or other optimization algorithms) and learning rate adjustment strategy, and gradually approaches the optimal solution; finally we calculate the accuracy of the improved PSPNet model to be 90.72.

[0116] In recent years, the power grid engineering field has frequently encountered subsidence geological disasters, which often lead to the damage of power transmission towers and the interruption of power supply, posing a serious threat to regional power safety. The space-borne synthetic aperture interferometric measurement technology InSAR has the ability to monitor ground surface deformation all day long and all weather. In recent years, the monitoring ability of slow deformation can effectively identify and warn potential subsidence landslide and other geological disaster risks, thereby providing important support for the safe operation of power facilities. However, there are relatively few studies on the automatic detection of subsidence risks based on InSAR using deep learning technology, and the automatic detection is of great significance for non-InSAR professionals to interpret subsidence risks. In the process of automatic detection of subsidence risks, in order to overcome the identification of subsidence risk areas at different scales, an automatic detection method for power grid corridor subsidence risk areas based on PSPNet and InSAR technology is proposed to improve the efficiency and accuracy of subsidence geological disaster detection. The method uses the pyramid pooling module to capture the multi-scale context information of the deformation image by performing pooling operations on the deformation rate image obtained by InSAR at different scales, thereby improving the accuracy of subsidence risk identification; the invention creates a sample set to train the improved PSPNet through visual interpretation of the InSAR results of the study area, realizing automatic monitoring of subsidence and landslide risk areas; by performing deep analysis and feature extraction on the historical ground surface deformation rate image input to the improved PSPNet, the model can identify and classify various risk areas. After completing the training of the improved PSPNet, the improved PSPNet is deployed to monitor and update the risk assessment in real time, effectively warning potential geological disasters and improving the efficiency of emergency response.

[0117] In an implementation, the historical surface deformation rate map and the deformation label of the historical surface deformation rate map are obtained based on historical satellite data of the target region, including:

[0118] The historical surface deformation rate map is obtained based on historical satellite data of the target region by using a small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) technology.

[0119] The historical surface deformation rate map is labeled to obtain the deformation label of the historical surface deformation rate map.

[0120] For example, as Figure 3 The sample schematic diagram of the historical surface deformation rate map and the deformation label; the size of the historical surface deformation rate map is 224*224, and the historical surface deformation rate map is manually labeled to obtain the deformation label of the historical surface deformation rate map.

[0121] For example, in the small baseline subset interferometric synthetic aperture radar (SBAS-InSAR) technology, the synthetic aperture interferometric synthetic aperture radar (InSAR) not only greatly improves the identification efficiency, realizes wider monitoring coverage, and also exhibits significant economic advantages by reducing operation costs, and has more prominent comprehensive performance and development potential compared with traditional methods. Optical remote sensing technology can quickly identify ancient landslides that have undergone overall deformation and landslides with significant deformation characteristics due to its high resolution, wide coverage, and rich image information. However, its identification ability is limited when facing slow deformation and lack of obvious deformation signs. In addition, cloud cover significantly affects the quality of optical remote sensing images, making it impossible to detect at all times and in all weather. Synthetic aperture radar interferometry uses satellite-acquired radar images to monitor ground subsidence and landslide deformation with high precision through interferometry. InSAR technology can generate high-resolution ground deformation maps by analyzing the phase differences of radar images acquired at different time points, revealing subtle deformation and dynamic changes in subsidence and landslide areas. Currently, InSAR technology has developed to a relatively mature stage. This technology has the ability to accurately detect millimeter-level ground deformation by deeply analyzing the phase information of radar images. Its unique advantages are the significant improvement in measurement accuracy, immunity to cloud cover, the possibility of all-weather operation, and wide monitoring coverage, which are described in detail in the literature. InSAR technology not only overcomes the limitations of optical remote sensing in cloudy and rainy application scenarios, but also provides high-precision ground deformation data, which has great value and potential in scientific research and practical application fields such as understanding the dynamic evolution of subsidence and landslide deformation, analyzing its formation mechanism, and achieving early identification and warning.

[0122] In an implementation, the improved inverted residual layer, the multi-pyramid pooling structure, and the dilated pyramid pooling structure in the pyramid scene analysis network are iteratively trained based on the training sample set to obtain a trained pyramid scene analysis network, including:

[0123] The historical ground surface deformation rate graph in the training sample set is input, and efficient feature compression is performed on the improved inverted residual layer to obtain a deformation rate feature vector.

[0124] For example, the historical ground surface deformation rate graph in the training sample set is received, and the historical ground surface deformation rate graph is uniformly adjusted to 224x224 pixel size in the improved pyramid scene analysis network. Then, efficient feature compression is performed through the improved Bottleneck (inverted residual) layer. This process maps the deformation history ground surface deformation rate graph feature to a deformation rate feature vector with a spatial resolution of 28x28 and a channel number of 192.

[0125] Based on the deformation rate feature vector, multi-scale image feature information is obtained through the multi-pyramid pooling structure and the dilated pyramid pooling structure, respectively, and a deformation rate fusion feature map is generated according to the multi-scale image feature information.

[0126] For example, the deformation rate feature vector is sent into two core components in the PSPNet in parallel: the multi-pyramid pooling structure (Pyramid Pooling Module) and the ASPP (Atrous Spatial Pyramid Pooling) dilated pyramid pooling structure. These two structures work together to capture rich feature information in images from different scales, effectively improving the recognition ability of the PSPNet model for complex scenes and multi-scale targets. Finally, the deformation rate fusion feature map is obtained after the above feature extraction and fusion.

[0127] The deformation rate fusion feature map is upsampled using the bilinear interpolation method to obtain a predicted deformation label of the historical ground surface deformation rate graph.

[0128] For example, the bilinear interpolation method is used for upsample processing to restore and map to the size of the historical ground surface deformation rate graph, thereby realizing fine semantic segmentation of the content of the historical ground surface deformation rate graph. In this process, each pixel in the historical ground surface deformation rate graph is accurately divided into two categories: a risk-free area and a subsidence risk area, and is identified by masks 0 and 1, respectively, to intuitively display the segmentation result.

[0129] The prediction deformation label of the historical ground surface deformation rate graph is compared with the deformation label corresponding to the historical ground surface deformation rate graph in the training sample set, and the weights of the improved inverted residual layer, the multi-pyramid pooling structure and the hollow pyramid pooling structure are updated according to the comparison result, and the training times are iterated until the training iteration times reach the pre-set training iteration times, and the training of the pyramid scene analysis network is completed; the pyramid scene analysis network is trained by using the deformation rate sample image and the deformation label of the deformation rate sample image, the depth separable convolution structure can significantly reduce the feature map dimension in the convolution process, and then the calculation amount of the pyramid scene analysis network is reduced; the core components of the pyramid scene analysis network, the multi-pyramid pooling structure and the hollow pyramid pooling structure, work cooperatively to capture rich feature information in the image from different scales, effectively improve the recognition ability of the model for complex scenes and multi-scale targets, and the feature map obtained after the above feature extraction and fusion is up-sampled by using the bilinear interpolation method to restore and map to the original image size, so that fine semantic segmentation of the image content is realized, and the prediction deformation label of the deformation rate sample image is obtained.

[0130] In an implementation manner, the point-by-point convolution includes a first point-by-point convolution and a second point-by-point convolution; the historical ground surface deformation rate graph in the training sample set is input, efficient feature compression is performed through the improved inverted residual layer, and a deformation rate feature vector is obtained, including:

[0131] Based on the historical ground surface deformation rate graph in the training sample set, deformation rate feature expansion is performed through the first point-by-point convolution, and a first point-by-point convolution feature map is obtained.

[0132] Based on the first point-by-point convolution feature map, a depth convolution operation is performed through the depth convolution, and a depth convolution feature map is obtained; based on the depth convolution feature map, a deformation rate feature vector is obtained through the second point-by-point convolution.

[0133] In an example, the improved inverted residual layer of the application is a depth separable convolution MobilenetV2 (a kind of lightweight convolutional neural network model) structure, as shown in Figure 4The MobileNetV2 structure diagram is shown, and the Pointwise convolution represents the pointwise convolution.The Pointwise convolution refers to a 1*1 convolution operation, which only operates in the channel dimension, and each channel of the input is multiplied by the corresponding weight point by point, which is used to change the channel number of the feature map.The Depthwise convolution represents the depthwise convolution.The Depthwise convolution is an operation of convolution independent of each channel, that is, each channel is multiplied by a convolution kernel independently, which is different from the traditional convolution, and is mainly used to reduce the model calculation amount, and is used together with the Pointwise convolution to construct the depthwise separable convolution;The innovation of MobileNetV2 compared with the traditional convolutional neural network (CNN) is that a significant difference is that it introduces the Depthwise Separable (depthwise separable) convolution mechanism, which replaces the traditional standard convolution operation.The Depthwise Separable convolution first performs convolution operation on each input channel independently through a single-channel convolution layer, namely depthwise convolution, and then uses a 1*1 convolution kernel, namely pointwise convolution, to perform cross-channel feature fusion and dimension expansion.This convolution method significantly reduces the feature map dimension in the convolution process, so it is usually necessary to pre-1*1 pointwise convolution for feature expansion to prepare for subsequent convolution operation, and when necessary, 1*1 pointwise convolution is used for feature dimension reduction, and this design mode constitutes the improved bottleneck structure;The pointwise convolution of the application operates in the channel dimension, and each channel of the input is multiplied by the corresponding weight point by point, which is used to change the channel number and feature dimension reduction of the feature map, and the depthwise convolution can perform cross-channel feature fusion and dimension expansion, thereby reducing the calculation amount of the pyramid scene analysis network.

[0134] In an implementation manner, the satellite data of the target region is collected according to the step 1.

[0135] For example, the C-band (wavelength 55.466mm) Sentinel-1 satellite data published by the European Space Agency is used as the main data source for InSAR processing;The Sentinel-1 is an all-time and all-weather radar imaging system, which is the first satellite developed by the European Commission and the European Space Agency for the Copernicus global earth observation project, launched in April 2014, entered the application stage in October 2014, and the Sentinel-1B satellite was launched in 2016, which together constitute the Sentinel-1 SAR (Synthetic Aperture Radar) satellite system of the Copernicus plan, and the original intention of the system is to target large-scale InSAR applications.

[0136] The Sentinel-1A / B satellite adopts a short spatial baseline design and a short revisit period. In the single-satellite mode, the revisit period is 12 days, and in the dual-satellite mode, the revisit period is 6 days. The entire European region is basically covered by data every 6 days, and other major regions are covered by data every 12 days, making the Sentinel-1 satellite data the most abundant data source for InSAR applications. As of now, the Sentinel-1 historical archive data has exceeded 6PB. In addition to the advantages of global coverage, short revisit period, and wide coverage, the most important thing is that the Sentinel-1 data adopts a completely open and free policy, which is free of charge for the global public by the European Space Agency (http: / / www.copernicus.eu). Therefore, high-quality Sentinel-1 data provides reliable data support for the development of the project.

[0137] In one implementation, the step 2 obtains the ground surface deformation rate map of the target area based on the satellite data by using a small baseline set aperture radar interferometry technology, and includes the following steps:

[0138] Filtering a plurality of synthetic aperture radar images from the satellite data;

[0139] For example, from the selected C-band Sentinel-1 satellite data, synthetic aperture radar (SAR) images suitable for SBAS-InSAR processing are collected. Because the Sentinel-1 satellite data covers the entire globe and revisits every 6 to 12 days, it provides high temporal and spatial resolution data.

[0140] For example, first, select the Sentinel-1 satellite data covering the target area. A satellite cycle can obtain multiple SAR images, and SAR images with shorter time intervals (12 days) and smaller spatial baseline distances (usually less than 200 meters) are preferentially selected. In this way, data interference caused by time and space decorrelation can be effectively reduced, and the continuity and reliability of the monitoring data can be improved.

[0141] The plurality of synthetic aperture radar images are paired two by two by using preset time and spatial baseline limits to generate a plurality of small baseline sets.

[0142] For example, the SBAS (Satellite-Based Augmentation System) algorithm is used to set specific time and spatial baseline limits to combine the SAR images obtained above to form a plurality of two-by-two paired combinations, and P two-by-two paired small baseline sets are generated.

[0143] A baseline threshold is set in advance to generate an unwrapped interferogram corresponding to each small baseline set;

[0144] For each small baseline set, an unwrapped interferogram is generated by setting a proper baseline threshold.

[0145] After the interferometric phase unwrapping of each unwrapped interferogram, the obtained unwrapped interferometric phases are combined to form an interferometric phase matrix, and a surface deformation rate is obtained based on the interferometric phase matrix by using a least square method.

[0146] Based on the surface deformation rate, a surface deformation rate map is generated.

[0147] After the multiple synthetic aperture radar images are screened, a small baseline set aperture radar interferometric measurement (SBAS-InSAR) technology is applied for data processing; by establishing a small baseline set, the influence of time and space decorrelation is reduced, and the purpose of this step is to accurately extract surface deformation information from the SAR images to provide a scientific basis for landslide monitoring; the small baseline set generated according to the synthetic aperture radar images can reduce the influence of time and space decorrelation, the distribution of the interferometric points in the differential unwrapped interferogram generated according to the small baseline set is uniform, the data integrity and analysis depth are improved, and the surface deformation rate reflects the surface deformation information of the target region.

[0148] In an implementation manner, after the interferometric phase unwrapping of each unwrapped interferogram, the obtained unwrapped interferometric phases are combined to form an interferometric phase matrix, and a surface deformation rate is obtained based on the interferometric phase matrix by using a least square method, including:

[0149] Based on the unwrapped interferometric phase related parameters, an unwrapped interferometric phase calculation formula is used to calculate the unwrapped interferometric phase of each unwrapped interferogram.

[0150] For example, the unwrapped interferogram composed of two SAR images is calculated, and the unwrapped phase at the target pixel in the unwrapped interferogram is obtained, the components contained in the phase are defined, and it is assumed that the kth unwrapped interferogram is generated by SAR images at t A and t B , t A >t B , and the calculation formula of the unwrapped phase at the pixel (x, r) is defined as:

[0151]

[0152] wherein, φ j (x, r) is the unwrapped phase of the jth unwrapped interferogram at the pixel (x, r), φ(t C , x, r) is the unwrapped phase at t CAt time t, the unwrapped phase at pixel (x,r), φ(t) A (x, r) represents the state of t. A At time t, the unwrapped phase at pixel (x,r), λ is the radar wavelength, and d(t) C (x, r) represents the state of t. C At time (x, r), the distance from the satellite to the Earth's surface is d(t). A (x, r) represents the state of t. A At time (x, r), the distance from the satellite to the Earth's surface is B. ⊥j The vertical baseline distance for the j-th observation is represented by Δz, which is the vertical spatial baseline of the radar imaging satellite during the j-th observation. Δz represents the elevation error, r represents the slant range, θ represents the angle of incidence, and φ represents the vertical baseline distance for the j-th observation. atm (t C (x, r) represents the state of t. C At time t, the atmospheric phase at pixel (x, r), φ atm (t A (x, r) represents the state of t. A At time (x, r), the atmospheric phase is Δn. j Let tP be the j-th random noise term, where j is the number of unwrapped interferograms or random noise terms, and tP is the total number of unwrapped interferograms or random noise terms.

[0153] A deformation time series for each target pixel is generated, and then the deformation time series for each target pixel is defined as follows:

[0154] φ T =[φ(t1),…,φ(t)] N )

[0155] Where, φ T Let φ(t1) represent the deformation time series of the target pixel, where T is the time of SAR image observation, and φ(t1) is the deformation information of the target pixel at time t1. N ) represents the target pixel at t N Deformation information at any given moment.

[0156] For the first time step, i.e., the deformation information φ(t0) of the target pixel at time t0 is 0, there are a total of N+1 SAR images.

[0157] The vector composed of P unwrapped interferograms obtained at the position of each pixel (x, r) is as follows:

[0158] δφ T =[δφ1,…,δφ P ]

[0159] Where, δφ TTo unwrap the interferometric phase, T is the time of SAR image observation, δφ1 is the first unwrapped interferometric phase, δφ P is the pth unwrapped interferometric phase.

[0160] For each unwrapped phase, define two index vectors of master image and slave image, the index vector is time index, the master image is defined as follows:

[0161] IM = [IM1, IM P ]

[0162] Wherein, IM is the master image of unwrapped interferogram, IM1 is the master image of the first unwrapped interferogram, IM P is the master image of the pth unwrapped interferogram.

[0163] The slave image is defined as follows:

[0164] IS = [IS1, IS P ]

[0165] Wherein, IS is the slave image of unwrapped interferogram, IS1 is the slave image of the first unwrapped interferogram, IS P is the slave image of the pth unwrapped interferogram.

[0166] And the master image and the slave image are sorted according to the acquisition time sequence, so as to satisfy the following formula:

[0167]

[0168] Wherein, IM j is the master image of the jth unwrapped interferogram, IS j is the slave image of the jth unwrapped interferogram.

[0169] Therefore, according to the above two index vectors of master image and slave image, the unwrapped phase of the unwrapped interferogram can be calculated, and the calculation formula is as follows:

[0170]

[0171] Wherein, δφ j is the unwrapped phase of the jth unwrapped interferogram, is the master image vector of the jth unwrapped interferogram, is the slave image phase of the jth unwrapped interferogram.

[0172] For the two SAR images acquired at t A and t B , the unwrapped phase of any point of the unwrapped interferogram generated by the two SAR images can be expressed by the following formula:

[0173]

[0174] where Δφ i is the unwrapped phase of any point in the i-th unwrapped interferogram, is the unwrapped phase at t C , t is the unwrapped phase at t A , and λ is the radar wavelength, is the cumulative surface deformation in the radar line-of-sight at t C , t is the cumulative surface deformation in the radar line-of-sight at t A , t is the topographic phase in the i-th unwrapped interferogram, is the atmospheric delay phase in the i-th unwrapped interferogram, is the noise phase in the i-th unwrapped interferogram.

[0175] After removing the topographic phase the atmospheric delay phase and the noise phase from the unwrapped interferogram, the calculation of the unwrapped phase of any point only contains the surface deformation phase. Meanwhile, the deformation phase in the calculation of the unwrapped phase of the pixel (x, r) is rewritten as the product of the average deformation rate between the two acquisition times and the time interval, thus obtaining the calculation of the unwrapped interferometric phase.

[0176] According to all the unwrapped interferometric phases, an interferometric phase matrix is generated;

[0177] For example, the mathematical expression of the interferometric phase matrix is as follows:

[0178] A · v = Δφ

[0179] where Δφ is the interferometric phase matrix, A is a coefficient matrix with a dimension of M x N, and v is the surface velocity deformation rate.

[0180] The least square method is used for inversion to extract the surface deformation time series from the interferometric phase matrix, and the surface deformation rate is obtained according to the surface deformation time series.

[0181] For example, the least square method is used to solve the above interferometric phase matrix to obtain the surface deformation rate.

[0182] For example, this invention utilizes small baseline ensemble aperture radar interferometry to monitor landslide geological hazards in the Sichuan-Tibet Highway region. In the context of InSAR (Interferometric Synthetic Aperture Radar) data processing, this invention typically refers to extracting time-series information on surface deformation from a series of interferograms. Specifically, this process involves several key computational steps: Unwrapping the interferometric phase: First, each interferogram needs to be unwrapped because the phase directly obtained from the interferogram is modulo 2π / 2π, meaning the phase value cycles between -π and π. The unwrapping process converts these "wrapped" phases into continuous, real physical phase changes. Constructing the matrix form of the interferometric phase: The unwrapped interferometric phases from multiple time points are combined into a matrix, where each column represents the phase map at a given time point; this matrix forms the basis for subsequent analysis. Time-series analysis: The least squares method is used to extract the time series of surface deformation from the phase matrix. This process may involve comparing and analyzing the phases at different time points to estimate the deformation of each pixel over time.

[0183] In one implementation, the formula for calculating the unwrapped interference phase is as follows:

[0184]

[0185] Where, Δφ j Let IU be the unwrapping interference phase of the j-th unwrapped interferogram. j Let IE be the starting index of the measurement time period for the interferometric pair of the j-th unwrapped interferogram. j Let t be the end index of the measurement time period for the j-th unwrapped interferogram, k be the measurement time of the measurement time period for the interferogram, and t be the end index of the measurement time period for the interferogram. k For the interferometric pair at time k within the measurement time interval, t k-1 For the interferometric pair at time k-1 within the measurement time interval, v k For in t k The surface velocity and deformation rate at that time.

[0186] In this invention, as Figure 5 The figure shows a schematic diagram of the overall technical route of the present invention. In the figure, Sentinel is the first radar imaging satellite in the Sentinel series of satellites of the European Space Agency, Sentinel-1 data is Sentinel-1 satellite data, DEM is Digital Elevation Model, SLC is Single Look Complex, and SAR is Synthetic Aperture Radar.

[0187] DEM data: InSAR, DEM (Digital Elevation Model) data is mainly used for two purposes:

[0188] (1) Remove terrain effects: InSAR technology monitors ground deformation by comparing radar images taken at different times in the same area. Since radar signals are affected by terrain (such as mountains, hills, etc.), the measured phase changes contain a component of the terrain height. DEM data can help remove this terrain effect from the InSAR phase difference, leaving only the actual deformation information of the surface. This step is usually called "terrain phase removal".

[0189] (2) Geocoding: Converting data from radar coordinate system to geographic coordinate system requires DEM data.

[0190] Precise orbit data: Precise orbit data of Sentinel-1 is used to improve the accuracy of InSAR processing. By providing accurate position and motion information of the satellite in the Earth's orbit, these data can eliminate phase bias caused by orbital errors, ensuring the accuracy of the interferogram.

[0191] SLC single look complex image: is a raw data format of SAR image, which contains amplitude and phase information of each pixel.

[0192] According to the present application, the SLC single look complex image is obtained according to the Sentinel-1 data, DEM data and precise orbit data, and the DEM coordinates are converted to the SAR coordinate system. When using InSAR technology, the coordinate system of DEM is converted to the SAR coordinate system, which is a common step. SAR image usually uses radar coordinate system, while DEM data may use geographic coordinate system. Therefore, coordinate system conversion and registration are necessary processes to ensure correct registration of SAR data and DEM data. According to the SLC single look complex image combined with the SAR coordinate system, the SAR image is registered, and the baseline graph is used for interference selection to generate the wrapped interferogram. The wrapped interferogram is an important step in InSAR processing, which is used to represent the phase difference between two satellite observations (SAR images taken at different times). The wrapped interferogram is usually obtained by calculating the interference phase between two SAR images, which can reveal surface deformation, terrain features, etc. The unwrapped interferogram is obtained from the wrapped interferogram. The unwrapped interferogram is sequentially subjected to least squares inversion, atmospheric phase removal and geocoding to obtain time series surface deformation inversion. The time series deformation can be obtained by least squares inversion based on unwrapped phase. The annual average deformation rate inversion is obtained according to the time series surface deformation inversion. The deformation rate map can be obtained by using the least squares method based on the time series deformation. The feature map is obtained by using the improved pyramid scene analysis network, and the feature map is subjected to pooling operation, upsampling operation and convolution operation to automatically identify the surface landslide subsidence risk area.

[0193] Example 2

[0194] This invention provides a specific embodiment of an automatic detection method for landslide settlement risk zones, as follows:

[0195] The experimental areas selected for verifying the effectiveness of the proposed method in this study were Shijiazhuang and Baoding cities in Hebei Province. Figure 6 The diagram shows the study area and the coverage of Sentinel-1 data. The area has typical plains, low hills and medium mountains, with significant vertical differences in the mountains, which provides the basic conditions for the occurrence of geological disasters such as landslides. In addition, the area has abundant power transmission facilities and power transmission corridors run through the mountains, providing a good test area for verifying the effectiveness of this patent.

[0196] In this experiment, we collected 188 Sentinel-1 single-view complex image data from January 2020 to March 2023 to obtain surface deformation data in areas such as Shijiazhuang, Hebei Province. After determining the monitoring area, we selected a small baseline set with a short time interval of 12 days and a small spatial baseline of less than 200 meters based on the terrain and monitoring requirements to ensure the data coherence and accuracy in subsequent processing. Based on the small baseline set, we determined the surface deformation rate map. Based on the surface deformation rate map, we used an improved PSPNet deep learning network to automatically monitor subsidence risk areas in order to implement the landslide deformation risk area detection task.

[0197] like Figure 7 The image shows the deformation rate map of the study area from January 2020 to March 2023, obtained based on SBAS-InSAR technology. Statistical analysis of the deformation rate revealed that 95.4% of the monitored targets had deformation rates ranging from -25 to 25 mm / year. Within the entire study area, we found significant localized subsidence concentrated mainly in regions A and B of the image, areas with dense power grids. We observed that these subsidences exhibited clear localized subsidence characteristics, namely, the presence of subsidence funnels. Furthermore, we found a relatively obvious trend in deformation characteristics within region A, which is helpful for interpretation... Figure 7 The localized settlement funnels shown present certain challenges. Therefore, to automate the identification of these localized settlement features, we created samples based on these settlement funnels and then used the PSPNet deep learning network for identification.

[0198] Finally, we utilized the trained and improved PSPNet across the entire research area of ​​40350 km. 2 The range was automatically detected to be 109.94 km. 2The settlement risk area is found in the surrounding area of the power grid corridor, and a large number of obvious local settlement areas are found, which will bring risks to the power facilities. The final experimental results show that the method proposed in the present application shows good applicability and result accuracy for different test areas; in the present application, an automatic detection method for settlement risk of power grid corridor based on pyramid scene analysis network PSPNet and synthetic aperture radar interferometry InSAR technology is proposed. The method can identify settlement risk areas of different scales by extracting multi-scale context information of the InSAR deformation rate map through the pyramid pooling module. We carried out experiments in Shijiazhuang and Baoding cities in Hebei Province to verify the effectiveness of the method proposed in the present application, using a total of 188 scenes of Sentinel-1 single-view complex SAR image data from January 2020 to March 2023, covering an area of 40350 square kilometers, and successfully automatically detecting a total of 109.94 square kilometers of settlement risk areas, with a model accuracy of 90.72%. The experimental results show that the method not only can effectively detect local settlement risks in a large area, but also can overcome the problem that traditional methods are difficult to detect trend deformation areas, showing high precision and adaptability. The technology proposed in the present application has important application value for the safety monitoring of power facilities, and provides a simplified and efficient settlement risk identification tool for non-InSAR professionals.

[0199] Embodiment 3:

[0200] Based on the same inventive concept, the present application also provides an automatic detection system for surface landslide settlement risk area as shown in Figure 8 , comprising a satellite data acquisition module, a surface deformation rate map extraction module and a surface landslide settlement risk area identification module.

[0201] The satellite data acquisition module is used to acquire satellite data of a target area.

[0202] The surface deformation rate map extraction module is used to obtain a surface deformation rate map of the target area based on the satellite data by using a small baseline set synthetic aperture radar interferometry technology.

[0203] The surface landslide settlement risk area identification module is used to identify the surface landslide settlement risk area in the target area based on the surface deformation rate map by using a pre-constructed improved pyramid scene analysis network.

[0204] The improved pyramid scene analysis network comprises an improved inverted residual layer, a multi-pyramid pooling structure and a hollow pyramid pooling structure, and the improved inverted residual layer is a deep separable convolution structure based on point-by-point convolution combined with deep convolution.

[0205] Preferably, the system further comprises: a construction module of the improved pyramid scene analysis network, the construction module of the improved pyramid scene analysis network comprising:

[0206] a historical satellite data processing submodule for obtaining a historical ground surface deformation rate graph and a deformation label of the historical ground surface deformation rate graph based on historical satellite data of a target region;

[0207] a sample set generation submodule for taking the historical ground surface deformation rate graph and the deformation label of the historical ground surface deformation rate graph as a sample set;

[0208] a sample set division submodule for dividing the sample set into a training sample set and a verification sample set;

[0209] an iterative training submodule for iteratively training an improved inverted residual layer, a multi-pyramid pooling structure and a hollow pyramid pooling structure in the pyramid scene analysis network based on the training sample set to obtain a trained pyramid scene analysis network;

[0210] a model correction submodule for correcting the trained pyramid scene analysis network based on the verification sample set to complete construction of the pyramid scene analysis network.

[0211] Preferably, the historical satellite data processing submodule is specifically configured to:

[0212] obtain a historical ground surface deformation rate graph based on historical satellite data of a target region by using a small-baseline set aperture radar interferometry technology;

[0213] perform deformation identification on the historical ground surface deformation rate graph to obtain a deformation label of the historical ground surface deformation rate graph.

[0214] Preferably, the iterative training submodule comprises:

[0215] an inverted residual layer training unit for taking a historical ground surface deformation rate graph in the training sample set as input, performing efficient feature compression on the historical ground surface deformation rate graph through the improved inverted residual layer to obtain a deformation rate feature vector;

[0216] a deformation rate fusion feature map generation unit for obtaining multi-scale image feature information by using the multi-pyramid pooling structure and the hollow pyramid pooling structure based on the deformation rate feature vector, and generating a deformation rate fusion feature map according to the multi-scale image feature information;

[0217] an up-sampling unit for performing up-sampling processing on the deformation rate fusion feature map by using a bilinear interpolation method to obtain a predicted deformation label of the historical ground surface deformation rate graph;

[0218] The structure parameter updating unit is configured to compare the predicted deformation label of the historical ground surface deformation rate graph with a deformation label corresponding to a historical ground surface deformation rate graph in the training sample set, update weights of the improved inverted residual layer, the multi-pyramid pooling structure and the hollow pyramid pooling structure according to a comparison result, and iterate a training iteration number until the training iteration number reaches a pre-set training iteration number, and complete training of the pyramid scene analysis network.

[0219] Preferably, the point-by-point convolution includes a first point-by-point convolution and a second point-by-point convolution; and the inverted residual layer training unit is specifically configured to:

[0220] The first point-by-point convolution is performed on the basis of the historical ground surface deformation rate graph in the training sample set to expand a deformation rate feature, and a first point-by-point convolution feature map is obtained;

[0221] The deep convolution is performed on the basis of the first point-by-point convolution feature map to obtain a deep convolution feature map;

[0222] The second point-by-point convolution is performed on the basis of the deep convolution feature map to obtain a deformation rate feature vector.

[0223] Preferably, the satellite data acquisition module includes:

[0224] The data screening submodule is configured to screen a plurality of synthetic aperture radar images from the satellite data;

[0225] The small baseline set generation submodule is configured to combine and pair the plurality of synthetic aperture radar images two by two to generate a plurality of small baseline sets by using pre-set time and space baseline limits;

[0226] The unwrapped interferogram generation submodule is configured to generate an unwrapped interferogram corresponding to each small baseline set by using a pre-set baseline threshold;

[0227] The ground surface deformation rate extraction submodule is configured to, after unwrapping the interference phase of each unwrapped interferogram, compose an unwrapped interference phase matrix from the unwrapped interference phase, and obtain a ground surface deformation rate by using a least square method on the basis of the unwrapped interference phase matrix;

[0228] The ground surface deformation rate graph generation submodule is configured to generate a ground surface deformation rate graph on the basis of the ground surface deformation rate.

[0229] Preferably, the ground surface deformation rate extraction submodule is specifically configured to:

[0230] The unwrapped interference phase of each unwrapped interferogram is calculated by using an unwrapped interference phase calculation formula on the basis of an unwrapped interference phase related parameter;

[0231] Based on all the unwrapped interference phases, generate the interference phase matrix;

[0232] The least squares method is used for inversion to extract the surface deformation time series from the interferometric phase matrix, and the surface deformation rate is obtained based on the surface deformation time series.

[0233] Preferably, the formula for calculating the unwrapped interference phase is as follows:

[0234]

[0235] Where, Δφ j Let IU be the unwrapping interference phase of the j-th unwrapped interferogram. j Let IE be the starting index of the measurement time period for the interferometric pair of the j-th unwrapped interferogram. j Let t be the end index of the measurement time period for the j-th unwrapped interferogram, k be the measurement time of the measurement time period for the interferogram, and t be the end index of the measurement time period for the interferogram. k For the interferometric pair at time k within the measurement time interval, t k-1 For the interferometric pair at time k-1 within the measurement time interval, v k For in t k The surface velocity and deformation rate at that time.

[0236] Example 4

[0237] like Figure 9 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0238] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, and the like, which are a computing core and a control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in the storage medium to implement a corresponding method flow or a corresponding function, so as to implement the steps of the surface landslide subsidence risk area automatic detection method in the above embodiment.

[0239] Embodiment 5

[0240] Based on the same inventive concept, the application further provides a readable storage medium, specifically an electronic device readable storage medium (Memory). The electronic device readable storage medium is a memory device in the electronic device, and is used for storing programs and data. It can be understood that the storage medium herein can include a built-in storage medium in the electronic device, and of course can also include an expansion storage medium supported by the electronic device. The storage medium provides a storage space, and the storage space stores an operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and the instructions can be one or more execution programs (including program codes). It should be noted that the storage medium herein can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. The processor loads and executes one or more instructions stored in the storage medium, and the steps of the surface landslide subsidence risk area automatic detection method in the above embodiment can be implemented.

[0241] Those skilled in the art should understand that the embodiments of the present application patent application can be provided as a method, a system, or a computer program product. Therefore, the present application patent application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application patent application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0242] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0243] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0244] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in one or more flow or blocks

[0245] Finally, it should be noted that the above embodiments are merely used to illustrate the technical solutions of the present application but not to limit the scope of protection of the present application, and although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that: after reading the present application, those skilled in the art can make various changes, modifications or equivalent replacements to the specific embodiments of the application, but these changes, modifications or equivalent replacements are all within the scope of protection of the claims of the present application.

Claims

1. An automatic detection method for surface landslide settlement risk zones, characterized in that, include: Collect satellite data of the target area; Based on the satellite data, a surface deformation rate map of the target area is obtained by using small baseline ensemble aperture radar interferometry. Based on the surface deformation rate map, the landslide and subsidence risk areas within the target area are identified through a pre-constructed improved pyramid scene analysis network. The improved pyramid scene analysis network includes an improved inverted residual layer, a multi-pyramid pooling structure, and a hollow pyramid pooling structure. The improved inverted residual layer is a depthwise separable convolutional structure based on pointwise convolution combined with depthwise convolution. The construction of the improved pyramid scene analysis network includes: Based on historical satellite data of the target area, historical surface deformation rate maps and deformation labels for historical surface deformation rate maps are obtained; The historical surface deformation rate map and its deformation labels are used as the sample set. The sample set is divided into a training sample set and a validation sample set; Based on the training sample set, the improved inverted residual layer, multi-pyramid pooling structure and hollow pyramid pooling structure in the pyramid scene analysis network are iteratively trained to obtain the trained pyramid scene analysis network. Based on the validation sample set, the trained pyramid scene analysis network is corrected to complete the construction of the pyramid scene analysis network.

2. The method as described in claim 1, characterized in that, Based on historical satellite data of the target area, historical surface deformation rate maps and deformation labels for these maps are obtained, including: Based on historical satellite data of the target area, a historical surface deformation rate map is obtained by using small baseline ensemble aperture radar interferometry. Deformation labels are obtained by performing deformation identification on the historical surface deformation rate map.

3. The method as described in claim 1, characterized in that, The improved inverted residual layer, multi-pyramid pooling structure, and hollow pyramid pooling structure in the pyramid scene analysis network are iteratively trained based on the training sample set to obtain the trained pyramid scene analysis network, including: Using historical surface deformation rate maps from the training sample set as input, efficient feature compression is performed through the improved inverse residual layer to obtain deformation rate feature vectors. Based on the deformation rate feature vector, multi-scale image feature information is obtained through multi-pyramid pooling structure and hollow pyramid pooling structure respectively, and deformation rate fusion feature map is generated according to the multi-scale image feature information. The deformation rate fusion feature map is upsampled using bilinear interpolation to obtain the predicted deformation label of the historical surface deformation rate map. The predicted deformation labels of the historical surface deformation rate map are compared with the deformation labels corresponding to the historical surface deformation rate map in the training sample set. Based on the comparison results, the weights of the improved inverted residual layer, multi-pyramid pooling structure and hollow pyramid pooling structure are updated, and the training iteration is iterated until the training iteration count reaches the preset training iteration count, thus completing the training of the pyramid scene analysis network.

4. The method as described in claim 3, characterized in that, The pointwise convolution includes a first pointwise convolution and a second pointwise convolution; the step of taking historical surface deformation rate maps from the training sample set as input and performing efficient feature compression through the improved inverse residual layer to obtain a deformation rate feature vector includes: Based on the historical surface deformation rate map in the training sample set, the deformation rate feature is extended through the first pointwise convolution to obtain the first pointwise convolution feature map. Based on the first pointwise convolutional feature map, a convolution operation is performed through the depthwise convolution to obtain a depthwise convolutional feature map; Based on the depth convolution feature map, the deformation rate feature vector is obtained through the second pointwise convolution.

5. The method as described in claim 1, characterized in that, Based on the satellite data, the surface deformation rate map of the target area is obtained using small baseline ensemble aperture radar interferometry, including: Multiple synthetic aperture radar images were selected from the satellite data; Using pre-set time and space baseline limits, the multiple synthetic aperture radar images are paired up in pairs to generate multiple small baseline sets; Using a pre-set baseline threshold, an unwrapped interferogram is generated for each small baseline set; After unwrapping the interferometric phases of each unwrapped interferogram, the resulting unwrapped interferometric phases are combined into an interferometric phase matrix. Based on the interferometric phase matrix, the surface deformation rate is obtained using the least squares method. A surface deformation rate map is generated based on the aforementioned surface deformation rate.

6. The method as described in claim 5, characterized in that, After unwrapping the interferometric phases of each unwrapped interferogram, the resulting unwrapped interferometric phases are combined into an interferometric phase matrix. Based on the interferometric phase matrix, the surface deformation rate is obtained using the least squares method, including: Based on the unwrapped interference phase correlation parameters, the unwrapped interference phase of each unwrapped interferogram is calculated using the unwrapped interference phase calculation formula. Based on all the unwrapped interference phases, generate the interference phase matrix; The least squares method is used for inversion to extract the surface deformation time series from the interferometric phase matrix, and the surface deformation rate is obtained based on the surface deformation time series.

7. The method as described in claim 6, characterized in that, The formula for calculating the unwrapped interference phase is as follows: in, Let j be the unwrapping interference phase of the j-th unwrapped interferogram. Let be the starting index of the measurement time period for the interferometric pair of the j-th unwrapped interferogram. Let be the end index of the measurement time period for the interference pair of the j-th unwrapped interferogram. To interfere with the measurement time of the measurement period, For the interferometric pair at time k within the measurement time period, For the interferometric pair at time k-1 within the measurement time period, In order to be in The surface velocity and deformation rate at that time.

8. An automatic detection system for surface landslide settlement risk zones, characterized in that, include: Satellite data acquisition module, surface deformation rate map extraction module, and surface landslide settlement risk zone identification module; The satellite data acquisition module is used to acquire satellite data of the target area; The surface deformation rate map extraction module is used to obtain the surface deformation rate map of the target area based on the satellite data and using small baseline ensemble aperture radar interferometry. The landslide subsidence risk zone identification module is used to identify landslide subsidence risk zones within the target area based on the land deformation rate map and through a pre-constructed improved pyramid scene analysis network. The improved pyramid scene analysis network includes an improved inverted residual layer, a multi-pyramid pooling structure, and a hollow pyramid pooling structure. The improved inverted residual layer is a depthwise separable convolutional structure based on pointwise convolution combined with depthwise convolution. The system further includes: a construction module for an improved pyramid scene analysis network, the construction module for the improved pyramid scene analysis network comprising: The historical satellite data processing submodule is used to obtain historical surface deformation rate maps and deformation labels for historical surface deformation rate maps based on historical satellite data of the target area. The sample set generation submodule is used to take the historical surface deformation rate map and the deformation label of the historical surface deformation rate map as the sample set; The sample set partitioning submodule is used to partition the sample set into a training sample set and a validation sample set; The iterative training submodule is used to iteratively train the improved inverted residual layer, multi-pyramid pooling structure and hollow pyramid pooling structure in the pyramid scene analysis network based on the training sample set, so as to obtain the trained pyramid scene analysis network. The model correction submodule is used to correct the trained pyramid scene analysis network based on the validation sample set, thereby completing the construction of the pyramid scene analysis network.

9. The system as described in claim 8, characterized in that, The historical satellite data processing submodule is specifically used for: Based on historical satellite data of the target area, a historical surface deformation rate map is obtained by using small baseline ensemble aperture radar interferometry. Deformation labels are obtained by performing deformation identification on the historical surface deformation rate map.

10. The system as described in claim 8, characterized in that, The iterative training submodule includes: The inverted residual layer training unit is used to take the historical surface deformation rate map in the training sample set as input, and perform efficient feature compression through the improved inverted residual layer to obtain the deformation rate feature vector. The deformation rate fusion feature map generation unit is used to obtain multi-scale image feature information based on the deformation rate feature vector through a multi-pyramid pooling structure and a hollow pyramid pooling structure, and generate a deformation rate fusion feature map based on the multi-scale image feature information. The upsampling unit is used to perform upsampling processing on the deformation rate fusion feature map using bilinear interpolation to obtain the predicted deformation label of the historical surface deformation rate map. The structural parameter update unit is used to compare the predicted deformation labels of the historical surface deformation rate map with the deformation labels corresponding to the historical surface deformation rate map in the training sample set. Based on the comparison results, the weights of the improved inverted residual layer, the multi-pyramid pooling structure, and the hollow pyramid pooling structure are updated, and the training number is iterated until the training iteration number reaches the preset training iteration number to complete the training of the pyramid scene analysis network.

11. The system as claimed in claim 10, characterized in that, The pointwise convolution includes a first pointwise convolution and a second pointwise convolution; the inverse residual layer training unit is specifically used for: Based on the historical surface deformation rate map in the training sample set, the deformation rate feature is extended through the first pointwise convolution to obtain the first pointwise convolution feature map. Based on the first pointwise convolutional feature map, a convolution operation is performed through the depthwise convolution to obtain a depthwise convolutional feature map; Based on the depth convolution feature map, the deformation rate feature vector is obtained through the second pointwise convolution.

12. The system as described in claim 8, characterized in that, The satellite data acquisition module includes: A data filtering submodule is used to filter out multiple synthetic aperture radar images from the satellite data; The small baseline set generation submodule is used to combine and pair the multiple synthetic aperture radar images in pairs using pre-set time and space baseline limits to generate multiple small baseline sets. The unwrapped interferogram generation submodule is used to generate unwrapped interferograms corresponding to each small baseline set using a pre-set baseline threshold. The surface deformation rate extraction submodule is used to unwrap the interferometric phase of each unwrapped interferogram, form an interferometric phase matrix from the unwrapped interferometric phases, and obtain the surface deformation rate based on the interferometric phase matrix using the least squares method. The surface deformation rate map generation submodule is used to generate a surface deformation rate map based on the surface deformation rate.

13. The system as described in claim 12, characterized in that, The surface deformation rate extraction submodule is specifically used for: Based on the unwrapped interference phase correlation parameters, the unwrapped interference phase of each unwrapped interferogram is calculated using the unwrapped interference phase calculation formula. Based on all the unwrapped interference phases, generate the interference phase matrix; The least squares method is used for inversion to extract the surface deformation time series from the interferometric phase matrix, and the surface deformation rate is obtained based on the surface deformation time series.

14. The system as described in claim 13, characterized in that, The formula for calculating the unwrapped interference phase is as follows: in, Let j be the unwrapping interference phase of the j-th unwrapped interferogram. Let be the starting index of the measurement time period for the interferometric pair of the j-th unwrapped interferogram. Let be the end index of the measurement time period for the interference pair of the j-th unwrapped interferogram. To interfere with the measurement time of the measurement period, For the interferometric pair at time k within the measurement time period, For the interferometric pair at time k-1 within the measurement time period, In order to be in The surface velocity and deformation rate at that time.

15. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, an automatic detection method for landslide subsidence risk zones as described in any one of claims 1 to 7 is implemented.

16. A readable storage medium, characterized in that, It contains an execution program, which, when executed, implements an automatic detection method for surface landslide subsidence risk zones as described in any one of claims 1 to 7.

Images