A detection method for mining subsidence areas based on MaskR-CNN model

Abstract

The present invention discloses a method for detecting mining subsidence areas based on a MaskR‑CNN model. First, the original SAR image is obtained, and then the original SAR image is co-registered to a common main image, and an InSAR observation is performed to obtain a winding interference image; based on the winding interference image, the winding interference image is sliced ​​to generate a color image, and after marking the ground subsidence characteristics of the mining area on the sliced ​​color interference map to make a data set, a Gaussian surface function is used in combination with GACOS to simulate the winding interference map to expand the data set; based on the Mask R‑CNN model, the data set is input for training to obtain a mining area ground subsidence detection model, and the mining subsidence area is detected and its position and boundary are determined; after obtaining the historical and latest mining subsidence areas, its evolution state is determined, and the surface activity level (activity time and frequency of each mining subsidence area) is evaluated. The present invention is easy to operate, has good effect, saves time and storage space, and is applicable to a wide range and does not limit the use of platforms.

Description

Technical Field

[0001] The present invention relates to the technical field of soil monitoring, and in particular to a method for detecting a mining subsidence area based on a Mask R-CNN model. Background Art

[0002] Synthetic aperture radar interferometry (InSAR) is an advanced geodetic tool with the characteristics of strong all-day and all-weather working ability, wide spatial coverage, high spatial resolution, high measurement accuracy and no need for ground instruments. It has become a powerful technology for wide-area surface deformation measurement.

[0003] In the prior art, mining subsidence area detection methods generally include two methods: (i) setting a deformation rate threshold, and (ii) using a machine learning (ML) algorithm.

[0004] For the first method of setting deformation rate threshold, it mainly studies the average deformation rate map obtained by InSAR stacking or time series algorithm. Although it is easy to implement, it cannot detect areas with limited mining subsidence, such as areas with mining subsidence within a limited period of time. For the second method of using machine learning (ML) algorithms, more and more ML-based algorithms have been developed to detect mining subsidence areas using SAR interferograms, but this method is still in its early stages and the technical means are not mature enough. Summary of the invention

[0005] Based on this, it is necessary to provide a detection method for mining subsidence areas based on the Mask R-CNN model to address the above technical problems.

[0006] The embodiment of the present invention provides a method for detecting mining subsidence areas based on a Mask R-CNN model, comprising:

[0007] Obtaining the original SAR image, as well as the orbital data and terrain data corresponding to the original SAR image;

[0008] Combining orbital data with terrain data, two adjacent SAR images are used for InSAR observation to obtain an entangled interferometric image.

[0009] Produce data sets using entanglement interferometry images;

[0010] Based on the Mask R-CNN model, the optimal mining area ground subsidence detection model was obtained through data set training and machine learning evaluation indicators as evaluation indicators of model training results;

[0011] Input the entanglement interference image into the mining area ground subsidence detection model to obtain the latest mining subsidence area;

[0012] After obtaining historical and recent mining subsidence areas, determine their evolution status and assess the level of surface activity;

[0013] In addition, the determination of the winding interference image comprises the following steps:

[0014] Interfere adjacent SAR images to obtain preliminary interferometric images;

[0015] Use external DEM to remove flat land and terrain phase effects;

[0016] Use precision orbits to remove the effects of orbit errors;

[0017] Multi-view operation is performed on the interferometric images to suppress noise;

[0018] Obtain the entanglement interference image.

[0019] In addition, the step of preparing a data set using the entanglement interference image comprises the following steps:

[0020] The surface deformation caused by mining is modeled using Gaussian surface function, and mining subsidence areas with different settlement amounts, boundary ranges and shapes are obtained by adjusting parameters.

[0021] The atmospheric phase is simulated by differencing the atmospheric delay map obtained by GACOS;

[0022] Use continuous uniformly distributed random functions to simulate noise phase;

[0023] The acquired surface deformation, atmospheric error and random noise are added and then twisted to obtain the data set.

[0024] In addition, the determination of the mining area ground subsidence detection model includes the following steps:

[0025] The real data set is divided into a training data set and a validation data set at a ratio of 4:1 to train the Mask R-CNN basic model;

[0026] The model is trained by increasing the number of iterations, and the precision, recall, accuracy and F1 comprehensive index are used as the model accuracy evaluation indicators. When the comprehensive accuracy reaches the highest, the mining area ground subsidence detection model using the real data set is obtained;

[0027] The Mask R-CNN basic model is executed using a mixed dataset of real and simulated data to obtain a mining area ground subsidence detection model using a mixed dataset;

[0028] The accuracy evaluation indicators of the mining subsidence detection model using the real data set and the mixed data set were compared to determine the final mining area ground subsidence detection model.

[0029] In addition, the method of obtaining the mining subsidence area includes:

[0030] Perform the same operations as those for the historical interferometric images to generate a winding interferogram together with the latest SAR image in the historical SAR dataset, and input the obtained winding interferogram into the mining area ground subsidence detection model to identify the latest mining subsidence area.

[0031] Additional indicators for assessing the level of surface activity include the duration and frequency of collapse in mining subsidence areas.

[0032] In addition, the duration of collapse in the mining subsidence area was obtained by multiplying the number of collapses by the time interval for acquiring SAR images.

[0033] The mining subsidence area detection method based on the Mask R-CNN model provided in the embodiment of the present invention has the following beneficial effects compared with the prior art:

[0034] The present invention uses Mask R-CNN as the basic detection model, and the original entangled interferogram generated from two adjacent SAR images is used as the data set of the detection model. A method for generating a mining area simulation data set is proposed, which helps to improve the accuracy of the model; a mining subsidence detection model based on the Mask R-CNN model is constructed, which can effectively detect the mining subsidence area from the entangled interferogram, saves calculation time, and reduces the additional uncertainty caused by phase unwrapping and time series inversion. It is a new method for wide-area mining subsidence detection that is easy to operate, has good effects, saves resources, has a wide range of applications, and is not restricted by the use of platforms. Rapid detection and long-term monitoring of ground subsidence in mining areas are important contents of monitoring mining activities, which are crucial to the economic development of mining cities and the safety of people and property. At the same time, it can be used to guide underground mining work, promote supervision of mining activities, and promptly discover illegal mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A mining subsidence detection technology framework diagram of a mining subsidence area detection method based on a Mask R-CNN model provided in one embodiment;

[0036] Figure 2 A flowchart for generating a simulated data set for a method for detecting mining subsidence areas based on a Mask R-CNN model provided in one embodiment;

[0037] Figure 3 A surface activity map based on wide-area mining subsidence detection results of a mining subsidence area detection method based on a Mask R-CNN model is provided in one embodiment. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0039] In one embodiment, a method for detecting mining subsidence areas based on a Mask R-CNN model is provided, the method comprising:

[0040] like Figure 1-3 As shown, the present invention utilizes the Mask R-CNN basic model and uses two adjacent SAR entangled interferograms to perform mining subsidence detection, aiming to save time and storage space while simplifying the operation method, obtain a better mining subsidence detection model, and evaluate the surface activity of the mining area on this basis.

[0041] First, the original SAR image, as well as the orbital data and terrain data corresponding to the original SAR image, are obtained. Only two adjacent SAR images are used for InSAR observation to obtain the entangled interferometric image:

[0042] Step 1: Interfere the aligned adjacent SAR images to obtain a preliminary interferometric image;

[0043] Step 2: Use external DEM to remove flat land and terrain phase effects;

[0044] Step 3: Use precision orbit to remove the influence of orbit error;

[0045] Step 4: Perform multi-view operation on the interference image to suppress noise;

[0046] Step 5: Generate a color image from the interference image.

[0047] like Figure 1 , a data set is made using winding interference images, and the ground subsidence features of the mining area are manually marked on the slice color interference map. The generated file contains the original image, mask information and marked mask image, which serves as the standard data set format for training network models.

[0048] In order to provide a method to generate InSAR interferograms of simulated mining subsidence according to the activity time and frequency of each mining subsidence area and the generalization ability of the model, a method is provided. Figure 2 As shown, the steps include:

[0049] Step 1: Use Gaussian surface function to model the surface deformation caused by mining, and obtain mining subsidence areas with different settlement amounts, boundary ranges and shapes by adjusting parameters;

[0050] Step 2: Differentiate the atmospheric delay map obtained by GACOS to simulate the atmospheric phase;

[0051] Step 3: Use a continuous uniformly distributed random function to simulate the noise phase;

[0052] Step 4: Add the surface deformation, atmospheric error, and random noise obtained in steps 3-1 to 3-3 and then tangle them to obtain a simulated interferogram data set.

[0053] Based on the Mask R-CNN model, the training data set and the test data set are used for training, and the machine learning evaluation index is used as the evaluation index of the model training result. Finally, the mining area ground subsidence detection model is obtained:

[0054] Step 1: Divide the real data set into a training data set and a validation data set at a ratio of 4:1 to train the Mask R-CNN basic model;

[0055] Step 2: Train the model by increasing the number of iterations, and use precision, recall, accuracy and F1 comprehensive index as model accuracy evaluation indicators. When the comprehensive accuracy reaches the highest, a mining subsidence detection model using a real data set is obtained;

[0056] Step 3: Execute steps 1 and 2 using the real and simulated mixed data set to obtain a mining subsidence detection model using the mixed data set;

[0057] Step 4: Compare the accuracy evaluation indicators of the mining subsidence detection model using the real data set and the mixed data set, and select the final mining subsidence detection model.

[0058] like Figure 1 The winding interferogram is input into the mining area ground subsidence detection model to detect the mining subsidence area and determine its location and boundary. After obtaining the historical and latest mining subsidence areas, its evolution status is determined and the surface activity level (activity time and frequency of each mining subsidence area) is evaluated.

[0059] Figure 3 The mining subsidence ground detection model can detect the mining subsidence area in the interference map. The position and boundary of the detected mining subsidence area are basically consistent with the detection results through visual identification.

[0060] Figure 3 represents the collapse duration. Since mining subsidence area detection is applied to separate interferograms with no temporal overlap and the time span of each interferogram is known, the cumulative duration of each collapse is determined by adding the time spans of all interferograms where a particular pixel is identified as “active”.

[0061] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for detecting mining subsidence areas based on the Mask R-CNN model, characterized in that: include: Obtaining the original SAR image, as well as the orbital data and terrain data corresponding to the original SAR image; Combining orbital data with terrain data, using two adjacent SAR images for InSAR observation, we can obtain an interferometric image: Produce data sets using entanglement interferometry images; Based on the Mask R-CNN model, the optimal mining area ground subsidence detection model is obtained through data set training and machine learning evaluation indicators are used as model training result evaluation indicators; the steps include: The real data set is divided into a training data set and a validation data set at a ratio of 4:1 to train the Mask R-CNN basic model; the model is trained by increasing the number of iterations, and the precision rate, recall rate, accuracy rate and F1 comprehensive index are used as the model accuracy evaluation indicators. When the comprehensive accuracy reaches the highest, the mining area ground subsidence detection model using the real data set is obtained; the Mask R-CNN basic model is executed using a real and simulated mixed data set to obtain a mining area ground subsidence detection model using the mixed data set; the accuracy evaluation indicators of the mining subsidence detection model using the real data set and the mixed data set are compared to determine the final mining area ground subsidence detection model; Input the entanglement interference image into the mining area ground subsidence detection model to obtain the latest mining subsidence area; The method of obtaining the latest mining subsidence area comprises: performing the same operation as that of operating the historical interference image, generating a winding interference map together with the latest SAR image in the historical SAR data set, and inputting the obtained winding interference map into the mining area ground subsidence detection model to obtain the latest mining subsidence area; After obtaining historical and recent mining subsidence areas, determine their evolution status and assess the level of surface activity; The step of obtaining a simulated interference pattern data set includes the following steps: The surface deformation caused by mining is modeled using Gaussian surface function, and mining subsidence areas with different settlement amounts, boundary ranges and shapes are obtained by adjusting parameters. The atmospheric phase is simulated by differencing the atmospheric delay map obtained by GACOS; Use continuous uniformly distributed random functions to simulate noise phase; The acquired surface deformation, atmospheric error and random noise are added and then twisted to obtain the simulated interferogram data set.

2. The method for detecting mining subsidence areas based on the Mask R-CNN model according to claim 1, characterized in that: The method of obtaining the entanglement interference image comprises the following steps: Interfere adjacent SAR images to obtain preliminary interferometric images; Use external DEM to remove flat land and terrain phase effects; Use precision orbits to remove the effects of orbit errors; Multi-view operation is performed on the interferometric images to suppress noise; Obtain the entanglement interference image.

3. The method for detecting mining subsidence areas based on the Mask R-CNN model according to claim 1, characterized in that: Indicators for assessing the level of surface activity include the duration and frequency of collapse in mining subsidence areas.

4. The method for detecting mining subsidence areas based on the Mask R-CNN model according to claim 3, characterized in that: The duration of collapse in the mining subsidence area is obtained by multiplying the number of collapses by the time interval for acquiring SAR images.

Images