Mine surface deformation monitoring method based on cross-scale transformer fusion of multi-source data

Abstract

The present disclosure provides a mine surface deformation monitoring method based on cross-scale Transformer fusion of multi-source data, relating to the technical field of mine surface deformation monitoring. The method comprises: dividing the multi-source data into geographic units based on a preset reference geographic unit, extracting features, identifying scales, encoding positions and tokenizing the multi-source data in each geographic unit to generate a feature token sequence carrying scale information and geographic spatial position information; inputting the feature token sequence into a cross-scale Transformer encoder, dynamically constraining the attention interaction range and weight between tokens based on the geographic spatial position distance, and fusing the feature token sequence; decoding and reconstructing the fused feature token sequence to generate a surface deformation field. The present disclosure solves the scale fragmentation problem of multi-source data through unified division and scale identification of geographic units, and reduces computational redundancy while ensuring accuracy through attention constraint driven by geographic spatial position, thereby improving monitoring accuracy, robustness and efficiency.

Description

Technical Field

[0001] This disclosure relates to the field of surface deformation monitoring technology in mining areas, and in particular to a method for monitoring surface deformation in mining areas based on cross-scale Transformer fusion of multi-source data. Background Technology

[0002] Surface deformation is an inevitable consequence of coal mining activities, seriously threatening the ecological environment and the safety of people and property in mining areas. With the rapid development of remote sensing technologies such as Synthetic Aperture Radar Interferometry (InSAR), Global Navigation Satellite System (GNSS), UAV photogrammetry, and LiDAR, collaborative monitoring using multi-source data has become the mainstream trend in surface monitoring. However, traditional data fusion methods mostly remain at a shallow level of data stitching, such as processing and fusing satellite image patches representing hundreds of square meters with UAV image patches representing only a few square meters in the same way. This leads to chaotic attention calculations and loss of correlation between features. Traditional fusion methods cannot address the core challenges of multi-source data across scales, heterogeneity, and strong spatiotemporal correlations, thus limiting the accuracy of deformation field reconstruction. Summary of the Invention

[0003] This disclosure aims to at least partially address one of the technical problems in the related art.

[0004] Therefore, the first aspect of this disclosure proposes a method for monitoring surface deformation in mining areas based on cross-scale Transformer fusion of multi-source data, comprising the following steps:

[0005] Acquire multi-source data covering the target mining area; The multi-source data is divided into geographical units based on a preset benchmark geographical unit. Features are extracted from the multi-source data within each geographical unit. After adding a scale identifier to the extracted features, a feature token carrying the corresponding scale information is generated, forming a feature token sequence. For each feature token, a location encoding is performed based on the geographic coordinates of its corresponding geographic unit, and the location encoding result is combined with the corresponding feature token so that each feature token carries the corresponding geospatial location information; The feature token sequence carrying the scale information and the geospatial location information is input into the cross-scale Transformer encoder. The cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between the feature tokens, and fuses the feature token sequence. The fused feature token sequence is decoded and reconstructed into a raster image to generate the surface deformation field of the target mining area.

[0006] In some embodiments of this disclosure, the step of extracting features from the multi-source data within each geographic unit and generating a feature token carrying corresponding scale information by attaching a scale identifier to the extracted features includes: extracting feature vectors from the multi-source data within each geographic unit using a lightweight convolutional neural network; embedding a learnable scale-type embedding vector into each feature vector to attach a scale identifier, thereby generating a feature token carrying corresponding scale information.

[0007] In some embodiments of this disclosure, the cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between feature tokens, including: determining the geospatial neighborhood of each query token based on a preset sampling distance threshold; using the geospatial neighborhood of each query token as the attention interaction range, and having the query token interact with key tokens and value tokens located within its geospatial neighborhood; generating an attention bias matrix based on the geospatial distance between feature tokens; and when the cross-scale Transformer encoder performs attention calculation, adding the corresponding bias value in the attention bias matrix to the attention score to adjust the attention weight.

[0008] In some embodiments of this disclosure, the step of decoding and reconstructing the fused feature token sequence into a raster image to generate the surface deformation field of the target mining area includes: using a pixel recombination upsampling module or a lightweight convolutional neural network decoder to decode and reconstruct the fused feature token sequence into the surface deformation field of the target mining area in the form of a continuous raster image based on its corresponding geographic coordinates.

[0009] In some embodiments of this disclosure, the method further includes: outputting the prediction variance of each pixel deformation value in parallel at the end of the pixel reconstruction upsampling module or the lightweight convolutional neural network decoder to generate an uncertainty evaluation map.

[0010] In some embodiments of this disclosure, the step of acquiring multi-source data covering the target mining area includes: collecting initial multi-source data covering the target mining area; preprocessing the initial multi-source data to obtain the first multi-source data; and spatiotemporally aligning the first multi-source data to the same spatial coordinate system and time reference to obtain the registered multi-source data.

[0011] In some embodiments of this disclosure, the multi-source data includes at least two of the following: synthetic aperture radar interferometry data, global navigation satellite system data, unmanned aerial vehicle imagery data, and lidar data.

[0012] In some embodiments of this disclosure, the surface deformation field includes at least one of the following: a surface deformation rate map, a cumulative deformation map, and a deformation time series map.

[0013] The second aspect of this disclosure proposes a surface deformation monitoring device for mining areas based on cross-scale Transformer fusion of multi-source data, comprising: The acquisition module is used to acquire multi-source data covering the target mining area; The tokenization processing module is used to divide the multi-source data into geographical units based on a preset benchmark geographical unit, extract features from the multi-source data in each geographical unit, and generate feature tokens carrying corresponding scale information by attaching scale identifiers to the extracted features, thus forming a feature token sequence. The location encoding module is used to encode the location of each feature token according to the geographic coordinates of its corresponding geographic unit, and combine the location encoding result with the corresponding feature token so that each feature token carries the corresponding geospatial location information. The feature fusion module is used to input the feature token sequence carrying the scale information and the geospatial location information into the cross-scale Transformer encoder. The cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between feature tokens, and fuses the feature token sequence. The decoding module is used to decode and reconstruct the fused feature token sequence into a raster map to generate the surface deformation field of the target mining area.

[0014] A third aspect of this disclosure provides an electronic device, including: a processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method described in the first aspect above.

[0015] This disclosure presents a method for monitoring surface deformation in mining areas based on cross-scale Transformer fusion of multi-source data. By dividing the multi-source data into geographic units using a baseline geographic unit, and extracting features, identifying scales, and tokenizing the multi-source data within each geographic unit, it generates feature token sequences carrying scale information within a unified geospatial framework. This fundamentally solves the problems of scale fragmentation and semantic misalignment in multi-source data. By explicitly injecting physical priors such as geographic coordinates and relative distances into the model, reasonable deformation estimates can still be generated based on complementary multi-source information and physical constraints, even in low-coherence areas or when a single data source is temporarily missing. This enhances the system's robustness and the interpretability of the results. Through constraints on the attention interaction range and attention weights, it queries a small number of related cross-scale key values ​​within the dynamically focused geographic neighborhood, achieving efficient and accurate feature fusion. This overcomes the computational redundancy and ineffective associations of standard attention methods, significantly reducing the computational complexity of data processing, improving processing efficiency, and reducing hardware resource dependence and energy consumption while ensuring monitoring accuracy. This disclosure can be accurately adapted to complex scenarios involving multi-source data across scales and strong spatiotemporal correlations in surface deformation monitoring in mining areas, enabling high-precision, high-efficiency, and high-reliability deformation field reconstruction.

[0016] Additional aspects and advantages of this disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this disclosure. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 A flowchart illustrating a method for monitoring surface deformation in mining areas based on cross-scale Transformer fusion of multi-source data, provided in this embodiment of the disclosure; Figure 2 This is a schematic diagram of a surface deformation monitoring device for mining areas based on cross-scale Transformer fusion of multi-source data, provided as an embodiment of this disclosure. Detailed Implementation

[0018] Embodiments of this disclosure are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting this disclosure.

[0019] Specifically, the following describes an embodiment of the mining area surface deformation monitoring method based on cross-scale Transformer fusion of multi-source data, with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart illustrating a method for monitoring surface deformation in mining areas based on cross-scale Transformer fusion of multi-source data, as provided in this disclosure. Figure 1 As shown, the method for monitoring surface deformation in mining areas based on cross-scale Transformer fusion of multi-source data may include the following steps: Step 101: Obtain multi-source data covering the target mining area.

[0021] Optionally, the multi-source data includes at least two of the following: synthetic aperture radar interferometry data, global navigation satellite system data, UAV imagery data, and lidar data.

[0022] For Synthetic Aperture Radar Interferometry (InSAR) data, the Sentinel-1C band Synthetic Aperture Radar satellite (ALOS-2L band satellite optional) can be selected, with the orbit mode being Interferometric Wide Swath (IW), a time baseline of 12 days, a spatial baseline of ≤100m, and the imagery needing to cover the entire target mining area and a surrounding 5km buffer zone. The goal is to acquire single-view complex (SLC) imagery data of the target mining area for more than 6 consecutive months for the initial capture of large-scale deformation trends.

[0023] For Global Navigation Satellite System (GNSS) data, 15-20 monitoring points can be set up in the mining area (at a density of 1 point per 5 km², and in subsidence hazard areas, the density can be increased to 1 point per 2 km²). Each monitoring point is equipped with a GNSS signal receiver with a sampling frequency of 1 Hz, and continuous observation is carried out for 24 hours. The latitude, longitude and elevation data of the monitoring point are recorded synchronously for high-precision absolute deformation calibration.

[0024] For drone imagery data, drones equipped with cameras for surveying operations can be used to acquire aerial photographs covering the entire target mining area by flying at an altitude of 120m, with a forward overlap of 85% and a lateral overlap of 80%.

[0025] For LiDAR data, airborne LiDAR equipment can be integrated into the aforementioned UAV to simultaneously acquire three-dimensional point cloud data of the mining area surface for terrain elevation calibration and small-scale crack extraction.

[0026] In one embodiment, the following conditions must be met during the data acquisition process: different types of data correspond to the same time period, the coordinates of monitoring points are uniformly based on the CGCS-2000 coordinate system, and the flight path of the UAV avoids high-voltage lines and tall buildings in the mining area to ensure data integrity.

[0027] In some embodiments of this disclosure, to ensure monitoring quality, the initial multi-source data collected can be preprocessed. Initial multi-source data covering the target mining area is collected; the initial multi-source data is preprocessed to obtain first multi-source data; the first multi-source data is spatiotemporally aligned to the same spatial coordinate system and time reference to obtain registered multi-source data.

[0028] The following steps can be used as a reference for different data preprocessing methods: 1. Preprocessing of Synthetic Aperture Radar Interferometry (InSAR) Data: Using image processing software, the earliest acquired SAR image is used as the master image, and the rest are slave images. The SAR images are registered using the phase gradient method. Then, the flat phase is removed based on the orbital parameters. Next, SRTM30mDEM data is used to help remove the terrain phase. Finally, phase unwrapping and geocoding are performed to output interferograms and deformation maps.

[0029] 2. GNSS data preprocessing: Precise point positioning (PPP) mode can be used, combined with IGS precise ephemeris and clock error data, to calculate the three-dimensional coordinates of the monitoring point; then, the 3σ criterion is used to remove outlier data (σ is the standard deviation of coordinate observations), and the Kalman filter algorithm is used to output the hourly high-precision absolute coordinates of the monitoring point.

[0030] 3. UAV image data preprocessing: Photogrammetry software can be used to perform aerial triangulation and orthorectification on the photos, and then stitch them together to generate DOM and DSM.

[0031] 4. LiDAR data preprocessing: Point cloud processing software can be used for noise reduction and filtering to remove elevation anomalies and separate ground points from non-ground points, generating DSM and DEM.

[0032] Spatially, the local coordinate system of the target mining area can be used as a reference. Based on coordinate transformation algorithms (such as the seven-parameter transformation method), synthetic aperture radar interferometry data, global navigation satellite system data, UAV imagery data, and lidar data can be uniformly transformed to this coordinate system. Using bilinear interpolation, all data are resampled to a unified grid coordinate system to ensure consistent spatial location references. Temporally, the time of GNSS data can be used as a reference. Linear interpolation is employed to interpolate the time of InSAR data (time baseline of 12 days), UAV data (time baseline of 7 days), and LiDAR data (time baseline of 30 days), generating daily time-series datasets.

[0033] Step 102: Divide the multi-source data into geographical units based on the preset benchmark geographical units, extract features from the multi-source data in each geographical unit, and generate feature tokens carrying the corresponding scale information by attaching scale identifiers to the extracted features, thus forming a feature token sequence.

[0034] In this embodiment, for each registered data layer, instead of using a fixed-size pixel grid, geographic units representing the same real-world geographic area are divided based on a reference geographic unit. For example, the reference geographic unit is set to 50m × 50m. For a SAR deformation map with a resolution of 10m, the image patch size corresponding to each token (each geographic unit) after division is 5×5 pixels; for a UAV DEM with a resolution of 0.1m, the image patch size corresponding to each token is 500×500 pixels; and GNSS point data is assigned to its respective geographic unit grid. Through geographic unit division and tokenization, regardless of which data is from the multi-source data, each token represents the same ground area, thereby achieving semantic alignment.

[0035] In one implementation, a lightweight convolutional neural network can be used to extract feature vectors from multi-source data within each geographic unit. A learnable scale-type embedding vector is embedded into each feature vector with an additional scale identifier to identify the data source scale. Feature tokens carrying corresponding scale information are generated through tokenization (one feature token for each data source in each geographic unit), mapping data at different scales to have the same feature dimension, further forming a feature token sequence. Here, the feature token sequence refers to the sequence obtained by arranging the tokens from all geographic units and all data sources in spatial or scale order.

[0036] Step 103: For each feature token, perform location encoding based on the geographic coordinates of its corresponding geographic unit, and combine the location encoding result with the corresponding feature token so that each feature token carries the corresponding geospatial location information.

[0037] In one implementation, the two-dimensional planar coordinates of the center point of the geographic unit can be used as geographic coordinates. The geographic coordinates are then mapped into a high-dimensional location vector P through a multilayer perceptron to locate the absolute position of the token on the map.

[0038] In some embodiments of this disclosure, if the model input is time-series data, a time-series location code may be further added to represent the observation time.

[0039] Step 104: Input the feature token sequence carrying scale information and geospatial location information into the cross-scale Transformer encoder. The cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between feature tokens, and fuses the feature token sequence.

[0040] In some embodiments of this disclosure, a sequence of feature tokens carrying scale information and geospatial location information can be input into a cross-scale Transformer encoder. Based on the scale information, the input feature tokens are grouped according to scale type, establishing an indexing mechanism for cross-scale interaction. Taking multi-source data including InSAR data, UAV imagery data, and GNSS data as an example, the aforementioned multi-source data correspond to large, medium, and small-scale tokens, respectively. When faced with an InSAR data query, the cross-scale Transformer encoder only needs to select two sets of tokens—UAV imagery data and GNSS data—from its geospatial neighborhood, and then select and fuse them according to attention weights. Each query does not need to traverse all tokens; it utilizes a cross-scale deformable attention mechanism to accurately interact from a reduced subset, improving the efficiency and accuracy of cross-scale attention calculation.

[0041] In some embodiments of this disclosure, the geospatial neighborhood of each query token can be determined based on a preset sampling distance threshold. The geospatial neighborhood of each query token is used as the attention interaction range, allowing the query token to interact with key and value tokens located within its geospatial neighborhood. In the attention layer, the query comes from all tokens. For each query, the model can dynamically sample a subset of relevant key and value tokens from its geospatial neighborhood for interaction, rather than interacting with all tokens. This sampling location is determined by the offset Δp predicted by the model based on the query content and location information, ensuring that it only samples and performs attention calculations from geospatial adjacent, cross-scale key-value tokens, rather than using full connections. This mechanism forces the model to learn to find the most relevant cross-scale features within the geospatial neighborhood for interaction, greatly improving efficiency and accuracy. Secondly, multiple cascaded cross-scale attention layers can be designed. The first few layers fuse coarse-scale features at a lower cost to establish a global deformation trend; the later layers introduce higher-resolution tokens for local refinement, thereby collaboratively modeling global trends and local details.

[0042] Before attention calculation, an attention bias matrix B can be generated based on the geospatial distance between feature tokens. During attention calculation in the cross-scale Transformer encoder, the corresponding bias values ​​in the attention bias matrix are added to the attention scores to adjust the attention weights. For any two tokens i and j, their geographical distance is d. ij Bias b ij = -γ×d ijγ is a learnable parameter. The greater the geographical distance between token pairs, the larger the absolute value of the negative bias applied, thus softly suppressing the attention weights between irrelevant tokens that are spatially far apart, solving the computational redundancy and inaccurate correlation problems of global attention. If the input is time-series data, a time-series location code is added to represent the observation time.

[0043] As an example, suppose the input includes token A and token B: Token A: From GNSS data, corresponding to the center coordinates of the geographic unit (1000m, 2000m). Token B: From UAV imagery data, corresponding to the center coordinates of the geographic unit (1010m, 2010m). S1 maps geographic unit coordinates to location vectors P using a multilayer perceptron (MLP). A and P B ; S2, calculate geographical distance d AB = ≈14.14m; S3, calculate bias b AB =-γ×14.14. If γ=0.1, then b AB =-1.414; S4, In attention calculation, when token A queries the original attention score of token B (calculated based on content features), a bias b is added. AB Since the bias is negative, the greater the distance, the less attention token A will pay to token B.

[0044] Step 105: Decode and reconstruct the fused feature token sequence into a raster map to generate the surface deformation field of the target mining area.

[0045] In some embodiments of this disclosure, a pixel shuffle upsampling module or a lightweight convolutional neural network decoder can be used to decode and reconstruct the fused feature token sequence, and the fused feature token sequence can be reconstructed into a continuous, high-resolution raster image of the surface deformation field of the target mining area based on its corresponding geographic coordinates.

[0046] The surface deformation field may include at least one of the following: surface deformation rate map, cumulative deformation map, and deformation time series map.

[0047] Optionally, the predicted variance of each pixel deformation value can be output in parallel at the end of the pixel recombination upsampling module or the lightweight convolutional neural network decoder to generate an uncertainty evaluation map, thereby improving the reliability of the results.

[0048] By implementing the embodiments of this disclosure, multi-source data is divided into geographic units based on a baseline geographic unit. Feature extraction, scale labeling, and tokenization are then performed on the multi-source data within each geographic unit. This allows data of different spatial resolutions to generate feature token sequences carrying scale information within a unified geospatial framework, fundamentally solving the problems of scale fragmentation and semantic misalignment in multi-source data. By explicitly injecting physical priors such as geographic coordinates and relative distances into the model, reasonable deformation estimates can still be generated based on complementary multi-source information and physical constraints, even in low-coherence areas or when a single data source is temporarily missing. This enhances the robustness of the system and the interpretability of the results. Through constraints on the attention interaction range and attention weights, a small number of related cross-scale key values ​​within the dynamically focused geographic neighborhood are queried, achieving efficient and accurate feature fusion. This overcomes the computational redundancy and ineffective associations of standard attention, significantly reducing the computational complexity of data processing while ensuring monitoring accuracy, improving processing efficiency, and reducing hardware resource dependence and energy consumption. This disclosure can accurately adapt to complex scenarios of multi-source data with cross-scale and strong spatiotemporal correlations in mine surface deformation monitoring, achieving high-precision, high-efficiency, and high-reliability deformation field reconstruction.

[0049] Figure 2 This is a schematic diagram of a surface deformation monitoring device for mining areas based on cross-scale Transformer fusion of multi-source data, provided as an embodiment of this disclosure. Figure 2 As shown, the surface deformation monitoring device for mining areas based on cross-scale Transformer fusion of multi-source data may include: an acquisition module 201, a tokenization processing module 202, a location encoding module 203, a feature fusion module 204, and a decoding module 205.

[0050] Module 201 is used to acquire multi-source data covering the target mining area; The tokenization processing module 202 is used to divide multi-source data into geographical units based on preset benchmark geographical units, extract features from multi-source data within each geographical unit, and generate feature tokens carrying corresponding scale information by attaching scale identifiers to the extracted features, thus forming a feature token sequence. The location encoding module 203 is used to encode the location of each feature token according to the geographic coordinates of its corresponding geographic unit, and combine the location encoding result with the corresponding feature token so that each feature token carries the corresponding geospatial location information. The feature fusion module 204 is used to input the feature token sequence carrying scale information and geospatial location information into the cross-scale Transformer encoder. The cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between feature tokens, and fuses the feature token sequence. Decoding module 205 is used to decode and reconstruct the fused feature token sequence into a raster map to generate the surface deformation field of the target mining area.

[0051] In some embodiments of this disclosure, the tokenization processing module 202 is specifically used to: extract feature vectors from multi-source data within each geographic unit using a lightweight convolutional neural network; embed a learnable scale-type embedding vector into each feature vector to add a scale identifier, thereby generating a feature token carrying the corresponding scale information.

[0052] In some embodiments of this disclosure, the feature fusion module 204 is specifically used to: determine the geospatial neighborhood of each query token based on a preset sampling distance threshold; use the geospatial neighborhood of each query token as the attention interaction range, and enable the query token to perform attention interaction with the key token and value token located in its geospatial neighborhood; generate an attention bias matrix based on the geospatial distance between feature tokens; and add the corresponding bias value in the attention bias matrix to the attention score when performing attention calculation in the cross-scale Transformer encoder to adjust the attention weight.

[0053] In some embodiments of this disclosure, the decoding module 205 is specifically used to: decode and reconstruct the fused feature token sequence using a pixel recombination upsampling module or a lightweight convolutional neural network decoder, and reconstruct the fused feature token sequence into a surface deformation field of the target mining area in the form of a continuous raster map based on its corresponding geographic coordinates.

[0054] In some embodiments of this disclosure, the acquisition module 201 is specifically used to: collect initial multi-source data covering the target mining area; preprocess the initial multi-source data to obtain first multi-source data; and perform spatiotemporal alignment on the first multi-source data to unify it to the same spatial coordinate system and time reference to obtain registered multi-source data.

[0055] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0056] To implement the above embodiments, this disclosure also proposes an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0057] To implement the above embodiments, this disclosure also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0058] To implement the above embodiments, this disclosure also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0059] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0060] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0061] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this disclosure includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of this disclosure pertain.

[0062] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0063] It should be understood that various parts of this disclosure can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0064] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0065] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0066] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present disclosure have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present disclosure. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present disclosure.

Claims

1. A method for monitoring surface deformation in mining areas based on cross-scale Transformer fusion of multi-source data, characterized in that, Includes the following steps: Acquire multi-source data covering the target mining area; The multi-source data is divided into geographical units based on a preset benchmark geographical unit. Features are extracted from the multi-source data within each geographical unit. After adding a scale identifier to the extracted features, a feature token carrying the corresponding scale information is generated, forming a feature token sequence. For each feature token, a location encoding is performed based on the geographic coordinates of its corresponding geographic unit, and the location encoding result is combined with the corresponding feature token so that each feature token carries the corresponding geospatial location information; The feature token sequence carrying the scale information and the geospatial location information is input into the cross-scale Transformer encoder. The cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between the feature tokens, and fuses the feature token sequence. The fused feature token sequence is decoded and reconstructed into a raster image to generate the surface deformation field of the target mining area.

2. The method according to claim 1, characterized in that, The step of extracting features from the multi-source data within each geographic unit and generating a feature token carrying corresponding scale information by attaching a scale identifier to the extracted features includes: A lightweight convolutional neural network is used to extract feature vectors from the multi-source data within each geographic unit; For each feature vector, embed a learnable scale-type embedding vector with an additional scale identifier to generate a feature token carrying the corresponding scale information.

3. The method according to claim 1, characterized in that, The cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between feature tokens, including: Based on a preset sampling distance threshold, the geospatial neighborhood of each query token is determined; The geospatial neighborhood of each query token is used as the attention interaction range, and the query token interacts with the key token and value token located in its geospatial neighborhood. An attention bias matrix is ​​generated based on the geospatial distance between feature tokens; When the cross-scale Transformer encoder performs attention calculations, the corresponding bias value in the attention bias matrix is ​​added to the attention score to adjust the attention weights.

4. The method according to claim 1, characterized in that, The step of decoding and reconstructing the fused feature token sequence into a raster image to generate the surface deformation field of the target mining area includes: The fused feature token sequence is decoded and reconstructed using a pixel recombination upsampling module or a lightweight convolutional neural network decoder. The fused feature token sequence is then reconstructed into a continuous raster image of the surface deformation field of the target mining area based on its corresponding geographic coordinates.

5. The method according to claim 4, characterized in that, Also includes: The prediction variance of each pixel deformation value is output in parallel at the end of the pixel reconstruction upsampling module or the lightweight convolutional neural network decoder to generate an uncertainty evaluation map.

6. The method according to claim 1, characterized in that, The acquisition of multi-source data covering the target mining area includes: Collect initial multi-source data covering the target mining area; The initial multi-source data are preprocessed to obtain the first multi-source data; The first multi-source data is spatiotemporally aligned and unified to the same spatial coordinate system and time reference to obtain the registered multi-source data.

7. The method according to any one of claims 1-6, characterized in that, The multi-source data includes at least two of the following: synthetic aperture radar interferometry data, global navigation satellite system data, UAV imagery data, and lidar data.

8. The method according to any one of claims 1-6, characterized in that, The surface deformation field includes at least one of the following: surface deformation rate map, cumulative deformation map, and deformation time series map.

9. A surface deformation monitoring device for mining areas based on cross-scale Transformer fusion of multi-source data, characterized in that, include: The acquisition module is used to acquire multi-source data covering the target mining area; The tokenization processing module is used to divide the multi-source data into geographical units based on a preset benchmark geographical unit, extract features from the multi-source data in each geographical unit, and generate feature tokens carrying corresponding scale information by attaching scale identifiers to the extracted features, thus forming a feature token sequence. The location encoding module is used to encode the location of each feature token according to the geographic coordinates of its corresponding geographic unit, and combine the location encoding result with the corresponding feature token so that each feature token carries the corresponding geospatial location information. The feature fusion module is used to input the feature token sequence carrying the scale information and the geospatial location information into the cross-scale Transformer encoder. The cross-scale Transformer encoder dynamically constrains the attention interaction range and attention weight of each query token based on the geospatial distance between feature tokens, and fuses the feature token sequence. The decoding module is used to decode and reconstruct the fused feature token sequence into a raster map to generate the surface deformation field of the target mining area.

10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-8.

Images