InSAR (Interferometric Synthetic Aperture Radar)-based mining area settlement funnel identification method, medium and equipment

Abstract

The invention discloses an InSAR-based mining area settlement funnel identification method, a medium and equipment, and relates to the technical field of radar remote sensing and image processing, the InSAR-based mining area settlement funnel identification method mainly comprises the following steps: preprocessing a synthetic aperture radar image of a research area to obtain a real interferogram; marking to obtain a real interferogram data set; constructing a simulation interferogram, and combining with the real interferogram data set to obtain a mixed data set; training the deep learning model by using the mixed data set to obtain a target detection model; and predicting the to-be-identified aperture radar image of the research area by using the target detection model to obtain a settlement funnel distribution diagram. By implementing the InSAR-based mining area settlement funnel identification method, the medium and the equipment provided by the invention, the settlement funnel can be quickly, accurately and automatically detected on a large-range interferogram.

Description

Technical Field

[0001] The invention relates to the technical field of radar remote sensing and image processing, and more specifically to a method, medium and equipment for identifying a mining area settlement funnel based on InSAR. Background Art

[0002] The mining process of underground mineral resources changes the original stress balance of the soil, causing large-scale collapse of the ground to form a settlement funnel. Mine collapse disaster is a common geological disaster that threatens people's lives and property safety and restricts the sustainable development of the mining area economy. Accurately and effectively identifying the deformation position of the settlement funnel is helpful to prevent and control geological disasters and monitor illegal mining. Therefore, it is of great significance to identify a large range of settlement funnels in mining areas. Although traditional geodetic methods have high measurement accuracy, they have limitations such as large human and material requirements, low spatial resolution, and limited monitoring area. Interferometric synthetic aperture radar (InSAR) technology is a new type of earth observation technology that works all day and all weather. It has the characteristics of wide monitoring range, large temporal and spatial scale, and the ability to obtain rich deformation information. InSAR technology has been widely used to monitor the settlement of mining areas. However, due to the influence of decoherence, the quality of phase unwrapping is uneven, making it difficult to quickly interpret deformation maps. In addition, considering the strict and complex operating procedures of InSAR technology, it is obviously time-consuming and laborious to detect subsidence funnels with a radius of tens to hundreds of meters in large-scale images. Therefore, combining InSAR technology with advanced deep learning algorithms provides a new solution for the automatic identification of subsidence funnels in large-scale mining areas.

[0003] The YOLOv8 model has been applied in target detection based on remote sensing data and has achieved some results. However, there are still some shortcomings when applying the YOLOv8 model to sedimentation funnel detection on interferograms. For example, the sample morphological characteristics of the sedimentation funnel dataset constructed only based on InSAR data are relatively simple; although the YOLOv8 model can maintain detection accuracy while having high computational efficiency, it has poor adaptability when identifying multi-scale sedimentation funnel targets; in addition, when training the model, there is an imbalance problem between difficult and easy samples in the sedimentation funnel sample set. Summary of the invention

[0004] The purpose of the present invention is to provide a method, medium and equipment for identifying subsidence funnels in mining areas based on InSAR, which can automatically detect subsidence funnels quickly and accurately on a large-scale interference pattern.

[0005] The present invention provides a method for identifying settlement funnels in a mining area based on InSAR, comprising the following steps: S1: preprocessing a synthetic aperture radar image of a study area to obtain a real interference map, annotating the real interference map to obtain a real interference map data set; S2: constructing a simulated interference map, and obtaining a mixed data set according to the simulated interference map and the real interference map data set; S3: using the mixed data set to train a deep learning model to obtain a target detection model; S4: using the target detection model to predict the synthetic aperture radar image to be identified in the study area to obtain a settlement funnel distribution map.

[0006] Furthermore, step S1 of the above-mentioned InSAR-based mining subsidence funnel identification method specifically includes: S11: cropping and aligning the synthetic aperture radar images of the study area to obtain the master-slave image registration results; S12: interferometrically processing the master-slave image registration results and removing the terrain phase to obtain a differential interferometric phase image; S13: segmenting, screening and standardizing the differential interferometric phase image to obtain a real interferometric map data set.

[0007] Furthermore, step S2 of the above-mentioned InSAR-based mining settlement funnel identification method specifically includes: S21: using the probability integral method to obtain the mining settlement deformation phase; S22: using the fractal Perlin noise simulation method to obtain the atmospheric noise phase; S23: using the complex Gaussian white noise simulation method to obtain the decorrelation noise phase; S24: according to the mining settlement deformation phase, the atmospheric noise phase and the decorrelation noise phase, a simulated interference pattern is obtained; S25: according to the simulated interference pattern and the real interference pattern data set, a mixed data set is obtained; Furthermore, step S3 of the above-mentioned InSAR-based mining subsidence funnel identification method specifically includes: S31: dividing the mixed data set according to a preset ratio to obtain a training set and a validation set; S32: using the training set and the validation set to train the deep learning model to obtain a target detection model.

[0008] Furthermore, the preset ratio of the above-mentioned InSAR-based mining subsidence funnel identification method is 0.85:0.15.

[0009] Furthermore, the deep learning model of the above-mentioned InSAR-based mining subsidence funnel identification method is obtained by improving YOLOv8; the improvement based on YOLOv8 includes the following steps: replacing the two c2f modules connected to the high-dimensional features of the backbone network part of the YOLOv8 model with the DWR-c2f module; adding the Dyhead module before the three detection heads of the Head part of the YOLOv8 model; replacing the CIoU of the IoU loss part of the YOLOv8 model with Inner-WIoU.

[0010] Furthermore, the above-mentioned InSAR-based mining subsidence funnel identification method also includes: evaluating the deep learning model using a real interference pattern data set and a mixed data set.

[0011] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned InSAR-based mining area settlement funnel identification method are implemented.

[0012] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the above-mentioned InSAR-based mining area settlement funnel identification method are implemented.

[0013] The present invention also provides a computer program product, including a computer program, which implements the steps of the above-mentioned InSAR-based mining area settlement funnel identification method when executed by a processor.

[0014] The InSAR-based mining funnel identification method, medium and equipment provided by the present invention have the following beneficial effects: The present invention takes into account the surface deformation law of the mining area and uses the differential interferometry (DInSAR) method in InSAR to obtain the real interference map of the study area; considering the difficulty of obtaining the real interference map and the single feature of the settlement funnel, the simulated interference map is synthesized based on the probability integral method, Polin noise and complex Gaussian white noise, and combined with the real interference map as a data set, which enriches the data features and solves the problem of the single feature of the InSAR settlement funnel data set; the present invention takes into account the multi-scale targets and the imbalance of difficult and easy samples in the process of interference map settlement funnel detection, combines the YOLOv8 model, and improves the feature extraction module based on the idea of ​​DWRseg network in the backbone network part of the model. A new module DWR-c2f is added, and the adaptive detection head Dyhead and the regression box loss function Inner-WIoU are introduced. A new mining area subsidence funnel automatic identification model SFD-YOLO is proposed to further improve the model recognition accuracy and generalization, and solve the multi-scale target problem and the imbalance of difficult and easy samples in the sample set when detecting large-scale subsidence funnels. After training on a mixed data set, the SFD-YOLO model further improves the detection accuracy of subsidence funnels. Compared with other deep learning methods, it can detect subsidence funnels more comprehensively and accurately, thereby realizing fast and accurate automatic detection of subsidence funnels on a large range of interference patterns, which is of great significance for preventing and controlling geological disasters and monitoring illegal mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1It is a flow chart of the InSAR-based mining area settlement funnel identification method provided by the present invention; Figure 2 The overall schematic diagram of the InSAR-based mining area settlement funnel identification method provided by the present invention; Figure 3 A schematic diagram of the deep learning model structure provided by the present invention; Figure 4 A schematic diagram comparing the recognition results of various deep learning models provided by the present invention; Figure 5 A schematic diagram of the sedimentation funnel identification result provided by the present invention; Figure 6 It is a structural block diagram of the computer device provided by the present invention. DETAILED DESCRIPTION

[0016] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0017] Figure 1 The schematic diagram of the InSAR-based mining area settlement funnel identification method of this embodiment is shown. In this embodiment, the InSAR-based mining area settlement funnel identification method includes the following steps: S1: Preprocess the synthetic aperture radar image of the study area to obtain the real interferogram, annotate the real interferogram, and obtain the real interferogram dataset; Specifically, step S1 includes: S11: crop and register the synthetic aperture radar images of the study area to obtain the master-slave image registration results; Specifically, in step S11, Sentinel-1A synthetic aperture radar (SAR) data of several consecutive years in the study area are obtained as data sources; in this embodiment, Sentinel-1A SAR ascending orbit data of Liulin and Yangquan mining areas in Shanxi from January 6, 2022 to April 25, 2024 and Sentinel-1A SAR ascending orbit data of Huainan mining area in Anhui from June 11, 2022 to January 8, 2024 are selected as data sources; the SAR data are imported, cropped, and registered, and the specific method is: an area including the target mining area is selected within the image range of several scenes of original SAR data, and all single look complex (SLC) original images are cropped to the same range of the selected area; a certain scene image of vertical co-directional polarization VV is selected as the main image, and the remaining slave images are registered to the main image space to obtain the main-slave image registration result; Specifically, the registration method is: identifying and matching the feature points of the master image and the slave image, and using a geometric transformation method to make the spatial positions of the slave image consistent with the master image, ensuring that they are aligned in the same coordinate system; wherein the geometric transformation method includes translation, rotation and scaling; S12: performing interference processing on the master-slave image registration result and removing the terrain phase to obtain a differential interference phase image; Specifically, the SAR data of the master-slave image registration result is subjected to interference processing to generate an interference pair with a time baseline between 12 days and 84 days and a spatial baseline less than 150 meters; then the terrain phase is removed from the original interference image to obtain a differential interference phase image; in this embodiment, the size of the differential interference phase image is 2410×2541 pixels; S13: segmenting, screening and standardizing the differential interferometric phase image to obtain a real interferogram data set; Specifically, the above-mentioned large-scale differential interference phase image was segmented into interference pattern patches of 300×300 pixels; several interference pattern patches were screened to remove pure background patches that did not contain sedimentation funnels; the screened interference pattern patches were manually labeled with mining sedimentation funnels using labelimg software, sample labels were made, and image annotation files with bounding box coordinates were generated; finally, a real interference pattern data set consisting of 1489 300×300 interference pattern patches was obtained; S2: construct a simulated interferogram, and obtain a mixed data set based on the simulated interferogram and the real interferogram data set; Specifically, Figure 2 FIG. 1 is an overall schematic diagram of a mining area settlement funnel identification method based on InSAR, wherein step S2 includes: S21: Using the probability integral method, the mining settlement deformation phase is obtained; Specifically, the basis of the probability integral method is the random medium theory. According to the basic principle of the probability integral method, the settlement deformation value W of any point (x, y) on the surface caused by unit mining is:

[0018]

[0019]

[0020] in, is the main influence radius, is the average mining depth, is the main influencing angle, is the maximum sinking angle, For the The plane coordinates of the cell center point, is the coordinate of any point on the surface; by setting the empirical parameters of the probability integral model, the surface deformation caused by mining can be simulated; considering the uncertainty of the signal in the actual deformation result, the elastic transformation algorithm is introduced to randomize the shape, position and stripe density of the deformation; the elastic transformation generates a random standard deviation in the interval (-1, 1) for each dimension of the pixel point, and uses a Gaussian filter (0, sigma) to filter the deviation matrix of each dimension, and finally uses the amplification factor alpha to control the deviation range; thus, get ; Generally speaking, the smaller the alpha, the larger the sigma, the smaller the deviation, and the closer to the original image; S22: Using the fractal Perlin noise simulation method, the atmospheric noise phase is obtained; Specifically, water vapor turbulence in the atmosphere often causes a phase delay called atmospheric noise phase. Perlin noise has been shown to be applicable to simulate the atmospheric phase in SAR interferograms. Fractal Perlin noise is obtained by superimposing Perlin noise with different frequencies and amplitudes to generate a random local atmospheric turbulence phase. S23: using the complex Gaussian white noise simulation method, the decorrelated noise phase is obtained; Specifically, complex Gaussian white noise is added to simulate the decorrelation effect, and the noise is generated as follows:

[0021] in, represents a standard normally distributed random matrix of the same size as the training sample, represents the normalized phase gradient, which is randomly modulated by the following linear mapping function:

[0022] Among them, the parameter NL determines the standard deviation STD of the maximum noise, and the parameter The minimum NL is constrained and randomly set during the simulation; independent noise matrices with the same NLs are generated as real and imaginary parts, which are then multiplied into a complex noise matrix for the simulated interferogram; S24: obtaining a simulated interference pattern according to the mining settlement deformation phase, the atmospheric noise phase and the decorrelation noise phase; Specifically, the mining settlement deformation phase, the atmospheric noise phase and the decorrelation noise phase are superimposed together and then wrapped to [-π, π], and the simulated interference pattern samples are obtained after coloring; in this embodiment, by adjusting the relevant parameters, a large number of interference pattern samples containing settlement funnels of various shapes can be quickly generated; finally, a total of 1000 simulated interference patterns of 300×300 pixels in size are generated to form a simulated sample set; S25: obtaining a mixed data set according to the simulated interference pattern and the real interference pattern data set; Specifically, this embodiment screened 750 real interferogram samples and 750 synthetic interferogram samples, and combined them to form a hybrid InSAR dataset containing 1500 images; S3: Use the mixed data set to train the deep learning model to obtain the object detection model; Specifically, step S3 includes: S31: Divide the mixed data set according to a preset ratio to obtain a training set and a validation set; Specifically, the preset ratio is 0.85:0.15; S32: Train the deep learning model using the training set and the validation set to obtain a target detection model; Specifically, the above-mentioned deep learning model includes but is not limited to any deep learning model of deep neural network (Single Shot MultiBox Detector, SSD), RetinaNet, YOLO-V5, YOLO-V7 and YOLO-V8; Specifically, the deep learning model is obtained based on the improvement of YOLOv8; the improvement based on YOLOv8 includes the following steps: replacing the two c2f modules connected to the high-dimensional features of the backbone network of the YOLOv8 model with the DWR-c2f module; adding the Dyhead module before the three detection heads of the Head part of the YOLOv8 model; replacing the CIoU of the IoU loss part of the YOLOv8 model with Inner-WIoU; Specifically, the above-mentioned improvement based on YOLOv8 includes the following steps: first, in the backbone network part of the model, the c2f module is improved based on the idea of ​​DWRseg to obtain DWR-c2f, so as to use the multi-scale dilated convolution of the DWR module to capture contextual information of different scales; secondly, the Dyhead module based on the attention mechanism is introduced in the Head part to further realize the adaptive detection of sinking funnel targets of various scales; finally, in the IoU loss part, Inner-WIoU is used to replace the CIoU of the YOLOV8 model to solve the balance problem of difficult and easy samples; It should be noted that the dilation-wise residual segmentation (DWRSeg) network solves the redundancy problem of deep learning models in multi-scale feature extraction through regional residualization and semantic residualization, effectively improving the efficiency and accuracy of feature extraction; at the high stage of the network, a more comprehensive view is needed to understand the image in order to obtain sufficient contextual information for accurate segmentation, so the dilation-wise residual (DWR) module integrates multiple convolution processes with different dilation rates, allowing a wider range of contextual information to be processed at the advanced network stage; Figure 3 As shown, by learning the idea of ​​DWRseg network, this embodiment improves the two c2f modules connected to the high-dimensional features of the YOLOv8 model backbone network into DWR-c2f, which are specifically located after the P4 and P5 features, respectively, so as to utilize the multi-scale dilated convolution of the DWR module to capture contextual information of different scales, improve the model feature extraction efficiency and multi-scale information acquisition capability, and thus improve the sedimentation funnel detection accuracy; It should be noted that the attention mechanism is a processing technique that directs attention to relevant information while filtering out irrelevant and unrelated data; it can grasp the relative importance of different semantic levels and selectively amplify features based on the dimensions of a single object; the Dyhead module unifies the object detection head and the attention mechanism; this method unifies multiple self-attention mechanisms in terms of feature levels, spatial positions, and output channels, enabling scale awareness, spatial awareness, and task awareness to be achieved simultaneously within a unified object detection head; this greatly improves the representation capabilities of the object detection head without increasing computational overhead; the input to the detection head is a scaled feature pyramid F, consisting of a three-dimensional tensor F∈RL×S×C Representation; where L represents the number of layers in the pyramid, S = H × W, H, W and C represent the height, width and number of channels of the intermediate level features respectively; the scale-aware attention module operates on the horizontal dimension, dynamically fusing different scales to adaptively combine the importance of different feature levels; the spatial-aware attention module operates on the spatial dimension, i.e., height × width, and adaptively aggregates features on layers at the same spatial position to obtain a more discriminative representation; the task-aware attention module operates within the channel and is able to dynamically activate or deactivate functional channels to flexibly facilitate different tasks; such as Figure 3 As shown in the figure, the Dyhead module is added before the three detection heads of the model, so that the feature map passes through the Dyhead module to dynamically fuse features of different scales before entering the detection head, thereby adaptively detecting sedimentation funnel targets of different sizes; It should be noted that the loss function of YOLOv8 consists of three parts, which can be divided into two branches: classification and regression; binary cross entropy (BCE) loss is used for classification, bounding box loss (CIoU) and distribution focus loss (DFL) are used for regression; the total loss function is obtained by the weighted sum of these three loss functions, which can be expressed as:

[0023] in, , , is the weight coefficient; CIoU adopted by YOLOv8 takes into account the overlapping area, center point distance and aspect ratio of bounding box regression, but its aspect ratio describes a relative value, which has a certain ambiguity and does not consider the balance between difficult and easy samples; Wise-IoU, or WIoU, proposes a dynamic non-monotonic focusing mechanism, uses "outlier degree" instead of IoU to evaluate the quality of anchor frames, and provides a wise gradient gain allocation strategy; this strategy reduces the competitiveness of high-quality anchor frames while also reducing the harmful gradients generated by low-quality examples; this allows WIoU to focus on anchor frames of ordinary quality, thereby effectively solving the problem of imbalance between difficult and easy samples and improving the overall performance of the detector; WIoU is specifically divided into three versions: v1, v2, and v3, and the calculation formulas are as follows:

[0024]

[0025]

[0026] Among them, (x, y) and (x gt ,y gt ) represent the center coordinates of the anchor box and the target box respectively; W g , H g is the length and width of the minimum bounding box; superscript Indicates that W g , H gSeparate from the calculation graph; β is the outlier degree, which is controlled by two hyperparameters α and δ, and r is the gradient gain; Inner-IoU is a novel regression box loss function calculation method. By analyzing the border regression process and mode, based on the characteristics of the border regression problem itself, it is proposed that using a smaller auxiliary border to calculate the loss during model training has a gain effect on the regression of high IoU samples, while the opposite is true for low IoU samples; therefore, Inner-IoU uses the scale factor ratio to control the generation of auxiliary borders of different scales for calculating the loss; Inner-IoU loss inherits some characteristics of IoU loss and has its own characteristics; Like IoU loss, the value range of Inner-IoU loss is [0,1]; because the auxiliary border and the actual border only differ in scale, the loss function is calculated in the same way; compared with IoU loss, when ratio is less than 1, the auxiliary border size is smaller than the actual border, and its regression is The effective range is smaller than the IoU loss, but the absolute value of its gradient is greater than the gradient obtained by the IoU loss, which can accelerate the convergence of high IoU samples; on the contrary, when the ratio is greater than 1, the larger-scale auxiliary border expands the effective range of regression, which is beneficial for the regression of low Iou; this embodiment combines the ideas of Inner-IoU and WIoU, applies them to the model regression box loss function, and obtains a new loss function Inner-WIoU, thereby solving the problem of imbalance between difficult and easy samples, and improving the efficiency and accuracy of the regression results, and further improving the model sedimentation funnel detection performance; S4: Use the target detection model to predict the aperture radar image to be identified in the study area and obtain the sedimentation funnel distribution map; Specifically, this embodiment uses the target detection model to monitor the subsidence funnel in the mining area of ​​Shanxi Province, a major mining province in China for a long time. The data source is unified as the Sentinel-1A ascending orbit SAR image; Figure 5 As shown in the blue rectangle in the figure, 16 SAR data scenes from three adjacent tracks were selected to generate three large-scale interferograms covering the lower, middle and upper parts of Shanxi Province respectively; in order to maintain coherence, the interferometric time baseline was unified to 24 days; the SAR images were taken between December 1, 2015 and January 1, 2016. The 30m resolution SRTM DEM was used to remove the terrain phase; the center position of the sedimentation funnel was determined according to the vertex coordinates of the model detection frame, and then the distribution of the sedimentation funnel in the whole Shanxi Province was obtained, as shown in the figure below. Figure 5 As shown; In this embodiment, a total of 952 sedimentation funnels were detected in various parts of Shanxi Province using the trained SFD-YOLO; In addition, Figure 5Middle (B) shows the number of subsidence funnels in 11 prefecture-level cities in Shanxi Province; the results show that Taiyuan, Linfen and Luliang are the top three cities with the most subsidence funnels in Shanxi Province, and Yuncheng City has the least subsidence funnels; the distribution of subsidence funnels in Shanxi Province is combined with the scope of key mining areas in Shanxi Province to analyze the distribution law of subsidence funnels and verify the close connection between mining activities in mining areas and the formation of subsidence funnels; by observing the distribution of subsidence funnels, this embodiment can easily determine the approximate scope and advancement direction of mining in mining areas; therefore, the use of InSAR interferograms for rapid deformation area detection plays a vital role in monitoring ongoing mining activities.

[0027] Optionally, the InSAR-based mining area subsidence funnel identification method further includes: evaluating the deep learning model using a real interferogram dataset and a hybrid dataset; Specifically, Figure 4 The figure shows a comparison diagram of the recognition results of various deep learning models; 85% of the data in the real interference pattern dataset and the mixed dataset are used as training sets to train five comparison models, namely SSD, RetinaNet, YOLO-V5, YOLO-V7 and YOLO-V8, as well as the improved deep learning model SFD-YOLO based on YOLOv8, and the other 15% of the data are used as test sets, and the accuracy index of each model detection is calculated according to the test results; It should be noted that the most advanced deep learning target detection methods can be divided into two categories: region proposal-based methods and regression-based methods; region proposal-based methods are also called two-stage detection methods, which divide the detection framework into two stages. First, the algorithm is used to extract the candidate region, and then the candidate frame target is corrected twice, so the detection accuracy is relatively high. The main methods include R-CNN, FastRCNN, Faster-RCNN, etc.; regression-based methods are also called single-stage detection methods. This method does not have a candidate region recommendation process. The target category can be determined and the target can be located in one stage, so high inference speed can be achieved. The main methods include YOLO series and RetinaNet, etc.; although the two-stage detection method has high accuracy, the model training is complex and the calculation is large. It is difficult to achieve real-time detection; therefore, this embodiment selects five most representative single-stage target detection methods, namely SSD, RetinaNet, YOLOv5, YOLOv7 and YOLOv8, as comparison models to improve experimental efficiency; due to the diversity of object types and shapes in natural images, heavy networks are often used as the backbone; however, the shape of the sedimentation funnel is relatively simple and the detection type is single, so a heavy network is not required to detect the sedimentation funnel; therefore, this embodiment uniformly selects the light version in the version of the YOLO series model, among which YOLOv5 and YOLOv8 select the "s" version model, and YOLOv7 selects the "l" version model; It should be noted that, for both the real data set and the mixed data set, 85% of the images were randomly selected as training samples, and the remaining 15% were used as test samples, where the training samples were also divided into training set and validation set in a ratio of 0.85:0.15; the models of this embodiment were trained and tested on a computer configured with an AMD Ryzen7-5800H CPU (3.20GHz) and an NVIDIA GeForce RTX 3060 graphics card; Pytorch (GPU version) was used as the deep learning framework of the experiment, and CUDA11.6 and CuDNN 8.3 were selected for GPU parallel computing and GPU acceleration; the Adam optimizer was used to update and optimize the network parameters; the initial learning rate of training was set to 1e-4; the cosine annealing algorithm was selected for the dynamic adjustment of the learning rate; after experiments, the mini-batch size and number of training iterations of all models were set to 9 and 100, respectively; all experiments displayed the position of the sedimentation funnel in a rectangular prediction box area; It should be noted that the main indicator used to evaluate the performance of deep learning models is meanAverage Precision (mAP), which is used in many other object detection studies to show the accuracy and recall trends of various training models. mAP is actually the average value of AP in each category. AP is calculated using the difference average accuracy evaluation method, that is, the area under the Precision-Recall curve, and the formula is:

[0028]

[0029] Where n represents the number of detection points. Represents the value of accuracy when the recall rate is r; N is the number of categories, and APk is the average precision of category k; because the category of the target to be detected in this embodiment is only "sedimentation funnel", mAP is actually equal to AP in this embodiment, so AP is used as the main accuracy verification indicator; AP50 and AP75 represent IoU thresholds of 0.5 and 0.75 respectively; AP50-95 represents the average AP value under different IoU thresholds with an IoU threshold range of 0.5 ~ 0.95 and a step size of 0.05; In addition, auxiliary evaluation indicators such as Precision, Recall and F1-Score are expressed in mathematical expressions as follows:

[0030]

[0031]

[0032] Among them, TP stands for true positive, which refers to the situation where an object is correctly predicted as a positive sample; FP stands for false positive, which means that an invalid object is detected as a positive sample; FN stands for false negative, which means that a valid object is detected as a negative sample.

[0033] Optionally, the above-mentioned InSAR-based mining area settlement funnel identification method further includes: comparing the detection accuracy of each model after training with two different data sets, verifying the improvement effect of the mixed data set on the detection accuracy of the model settlement funnel; comparing the detection accuracy of the SFD-YOLO model trained with the mixed data set with the other five comparison models, verifying the performance of the improved model; It should be noted that, in order to verify the improvement effect of the hybrid InSAR dataset proposed in this embodiment on the detection accuracy of the model sedimentation funnel, five target detection models are trained using a dataset constructed by combining simulated interferogram samples and real interferogram samples, and the accuracy indicators F1-Score, AP50, AP75 and AP50-95 of each model on the test set are calculated; then, the test set detection accuracy indicators of each model trained only with the real sample set are compared; taking the YOLOv8 model as an example, after the model is retrained on the hybrid dataset, the detection effect of the sedimentation funnel in the same area is significantly improved; It should be noted that, in order to verify the improvement effect of the improved model SFD-YOLO based on YOLOv8 proposed in this embodiment on the detection accuracy of sedimentation funnels, it is compared with the other five target detection methods including the YOLOv8 basic model; combined with the experimental results, it can be seen that when the SFD-YOLO8 model proposed in this embodiment is trained on real data sets and mixed data sets, the AP index of the model is better than that of other comparison models, and the Params index and FLOPs index of the model are smaller, and the FPS index is the largest, that is, it has a smaller model complexity and the fastest model transmission speed; in general, under the same experimental method, the SFD-YOLO model is the best among the six deep learning models. The best method among the learning methods; among them, when SFD-YOLO is trained on the mixed data set, its AP50, AP75 and AP50-95 can reach 92.0%, 60.0% and 54.1% respectively; in addition, the SFD-YOLO model trained on the mixed data set can achieve the highest detection accuracy among all the methods in this embodiment. Compared with the YOLOv8 basic model trained on the real data set, the AP50, AP75 and AP50-95 of the model detecting the sedimentation funnel are improved by 2.2%, 2.8% and 2.0% respectively; when trained on the mixed data set, SFD-YOLO has a greater advantage than other models in the detection effect of the sedimentation funnel in the same local area.

[0034] Optionally, the above-mentioned InSAR-based mining area settlement funnel identification method further includes: testing the robustness of the proposed improved model SFD-YOLO to noise and the generalization to different terrains and multi-source data; It should be noted that due to the speckle effect and various other forms of noise, coherent speckle noise is usually present in the InSAR interferogram, which has an adverse effect on subsequent processing steps; many advanced speckle filters are usually used to reduce these noise sources; however, considering that in actual scenarios, the quality of the interferogram is often uneven even after filtering, resulting in residual noise, the real and simulated samples used for training in this embodiment are not speckle filtered; the trained SFD-YOLO model is used to detect the sedimentation funnel in the same local area in the interferogram with and without speckle noise, respectively, to verify the robustness of the model proposed in this embodiment to speckle noise; It should be noted that the features of the regions with different surface types in the interferograms are often very different; for example, in the interferograms of mountainous areas, regular dense stripes caused by the overlapping phenomenon of SAR images often appear, while a large amount of decoherent noise appears near water areas; these unique features will affect the model's extraction of sedimentation funnel features, resulting in a rapid decrease in the detection accuracy of the model in these areas; in order to verify the generalization of the model proposed in this embodiment to different surface types, the SFD-YOLO model and four comparison models, YOLOv5, YOLOv7, SSD and RetinaNet, are used to perform local sedimentation funnel detection in mountainous areas, near water areas and flat areas; the sedimentation funnel detection accuracy of each model in the corresponding different areas is compared to verify the generalization of the model proposed in this embodiment to data of different surface types; Specifically, in order to verify the generalization ability of the model proposed in this embodiment for multi-source data, the interferogram generated by using Radarsat2 SAR data as the data source was tested; two representative large-scale interferograms of mining areas were selected; the sizes of the two interferograms were both 3155 × 3813 pixels; the time baseline of the interferogram was 24 days, from December 31, 2015 to January 24, 2016, and from January 24, 2016 to February 17, 2016; first, the entire image was cropped to generate a large number of sub-images with a size of 150 × 150 pixels; then, the SFD-YOLO model trained on the mixed data set was used to detect the sedimentation funnel on each sub-image; finally, the detection results of each sub-image were spliced; by verifying the detection effect of the model proposed in this embodiment on the sedimentation funnel on the interferogram generated by Radarsat2 SAR data, the generalization ability of the model for multi-source data was verified.

[0035] This embodiment provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned method for identifying a mining area settlement funnel based on InSAR are implemented. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory (Flash Memory), a hard disk (HDD) or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above-mentioned types of memory.

[0036] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the above-mentioned InSAR-based mining area settlement funnel identification method are implemented.

[0037] like Figure 6 As shown, the computer device may include: at least one processor 121, such as a CPU (Central Processing Unit), at least one communication interface 123, a memory 124, and at least one communication bus 122. Among them, the communication bus 122 is used to realize the connection and communication between these components. Among them, the communication interface 123 may include a display screen (Display), a keyboard (Keyboard), and the optional communication interface 123 may also include a standard wired interface and a wireless interface. The memory 124 may be a high-speed RAM memory (Random Access Memory, volatile random access memory), or a non-volatile memory (non-volatile memory), such as at least one disk memory. The memory 124 may also be at least one storage device located away from the aforementioned processor 121. Among them, the memory 124 stores an application program, and the processor 121 calls the program code stored in the memory 124 to perform any of the above method steps. Among them, the communication bus 122 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 122 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6Only one line is used to represent it, but it does not mean that there is only one bus or one type of bus. Among them, the memory 124 may include a volatile memory (English: volatile memory), such as a random access memory (English: random-access memory, abbreviated: RAM); the memory may also include a non-volatile memory (English: non-volatile memory), such as a flash memory (English: flash memory), a hard disk drive (English: hard disk drive, abbreviated: HDD) or a solid-state drive (English: solid-state drive, abbreviated: SSD); the memory 124 may also include a combination of the above-mentioned types of memory. Among them, the processor 121 may be a central processing unit (English: central processing unit, abbreviated: CPU), a network processor (English: network processor, abbreviated: NP) or a combination of CPU and NP. Among them, the processor 121 may further include a hardware chip. The above-mentioned hardware chip may be an application-specific integrated circuit (English: application-specific integrated circuit, abbreviated: ASIC), a programmable logic device (English: programmable logic device, abbreviated: PLD) or a combination thereof. The PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof. Optionally, the memory 124 is also used to store program instructions. The processor 121 can call the program instructions to implement the InSAR-based mining area settlement funnel identification method as in this embodiment.

[0038] This embodiment provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the above-mentioned InSAR-based mining subsidence funnel identification method.

[0039] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for identifying mining area settlement funnel based on InSAR, characterized in that: The following steps are involved: S1: preprocessing the synthetic aperture radar image of the study area to obtain a real interferogram, annotating the real interferogram to obtain a real interferogram data set; S2: constructing a simulated interference pattern, and obtaining a mixed data set according to the simulated interference pattern and the real interference pattern data set; S3: Using the mixed data set to train a deep learning model to obtain a target detection model; S4: Using the target detection model, the aperture radar image to be identified in the study area is predicted to obtain a sedimentation funnel distribution map.

2. The InSAR-based mining funnel identification method according to claim 1, characterized in that: Step S1 specifically includes: S11: crop and register the synthetic aperture radar images of the study area to obtain the master-slave image registration results; S12: performing interference processing on the master-slave image registration result and removing the terrain phase to obtain a differential interference phase image; S13: segmenting, screening and standardizing the differential interference phase image to obtain a real interference pattern data set.

3. The InSAR-based mining area subsidence funnel identification method according to claim 1, characterized in that: Step S2 specifically includes: S21: Using the probability integral method, the mining settlement deformation phase is obtained; S22: Using the fractal Perlin noise simulation method, the atmospheric noise phase is obtained; S23: using the complex Gaussian white noise simulation method, the decorrelated noise phase is obtained; S24: obtaining a simulated interference pattern according to the mining settlement deformation phase, the atmospheric noise phase and the decorrelation noise phase; S25: Obtain a mixed data set according to the simulated interference pattern and the real interference pattern data set.

4. The method for identifying mining area subsidence funnel based on InSAR according to claim 1, characterized in that: Step S3 specifically includes: S31: Divide the mixed data set according to a preset ratio to obtain a training set and a validation set; S32: Use the training set and the validation set to train the deep learning model to obtain a target detection model.

5. The InSAR-based mining area settlement funnel identification method according to claim 4, characterized in that: The preset ratio is 0.85:0.

15.

6. The InSAR-based mining area settlement funnel identification method according to claim 4, characterized in that: The deep learning model is obtained based on the improvement of YOLOv8; the improvement based on YOLOv8 The method comprises the following steps: replacing two c2f modules connected to the high-dimensional features of the backbone network of the YOLOv8 model with DWR-c2f modules; Add the Dyhead module before the three detection heads in the Head part of the YOLOv8 model; replace the CIoU in the IoU loss part of the YOLOv8 model with Inner-WIoU.

7. The InSAR-based mining area settlement funnel identification method according to claim 1, characterized in that: The InSAR-based mining subsidence funnel identification method also includes: evaluating the deep learning model using the real interference pattern data set and the mixed data set.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying a mining subsidence funnel based on InSAR as described in any one of claims 1 to 7 are implemented.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the InSAR-based mining subsidence funnel identification method as described in any one of claims 1-7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the InSAR-based mining subsidence funnel identification method described in any one of claims 1 to 7 are implemented.

Images