Fusion insar coal seam mining condition under the aquifer seepage ground surface response identification and early warning system, method, device and medium

Abstract

This application proposes a system, method, device, and medium for identifying and warning of surface response to aquifer seepage under coal seam mining conditions, integrating InSAR, belonging to the interdisciplinary field of remote sensing monitoring and computer vision. The system includes: a multi-source feature fusion module for extracting surface deformation features, optical remote sensing features, and topographic features, and constructing an index feature set to reflect the seepage mechanism of underground aquifers under coal seam mining conditions; a multi-source feature fusion module for processing and fusing the features extracted by the multi-source feature extraction module to form physical prior features with clear physical meaning; a seepage response identification module; a spatiotemporal evolution modeling module; and a risk warning output module. This application can accurately extract seepage areas with small ranges and weak responses, with no fragmentation problem in the identification results, clear boundary delineation, and the ability to achieve large-scale continuous monitoring. Moreover, it has high accuracy in spatiotemporal evolution prediction and can accurately capture the dynamic laws of underground aquifer seepage under coal seam mining conditions.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of remote sensing monitoring and computer vision, and in particular to a system, method, device and medium for identifying and warning of surface response to aquifer seepage under coal seam mining conditions by integrating InSAR. Background Technology

[0002] Under coal seam mining conditions, seepage from underground aquifers is a significant factor inducing secondary disasters such as surface subsidence, landslides, and ground fissures. Currently used monitoring methods include:

[0003] Traditional monitoring methods, such as well point measurement and ground-penetrating radar, have problems such as limited spatial coverage, high cost, and poor real-time performance.

[0004] Small baseline set interferometric synthetic aperture radar (SBAS-InSAR) technology: Although it can monitor surface deformation over a wide area, it is not sensitive enough to indirect features such as weak humidity changes, vegetation response, and surface temperature fluctuations caused by seepage from underground aquifers.

[0005] Existing research on Synthetic Aperture Radar Interferometry (InSAR) technology mainly focuses on deformation inversion, failing to establish effective models to collaboratively analyze InSAR deformation sequences, optical image features (such as Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI), and topographic factors (such as slope and Topographic Moisture Index (TWI)). This results in poor anti-interference capabilities and makes it difficult to accurately separate weak and ambiguous signals caused by seepage from complex background changes in mining areas (such as mining subsidence and natural vegetation growth).

[0006] Limitations of identification methods: seepage in underground aquifers is a process with spatiotemporal continuity and evolutionary patterns. Existing segmentation models are unable to effectively extract seepage areas with ambiguous boundaries and spatial discontinuities, and lack spatiotemporal evolution modeling and dynamic early warning mechanisms.

[0007] The main problems existing in the prior art can be summarized as follows:

[0008] It has a weak ability to respond to indirect characteristics of seepage in underground aquifers: it relies on a single deformation phase and ignores multi-source information such as optical and thermal infrared.

[0009] The multi-source data fusion mechanism is lacking: a collaborative analysis model for multi-modal data such as deformation, optical, and terrain data has not been established;

[0010] Insufficient spatiotemporal continuity modeling: the identification results are fragmented, making it difficult to capture the dynamic evolution process of the seepage zone;

[0011] The physical mechanism coupling is missing: the identification results lack prior hydrogeological support and have poor interpretability.

[0012] Therefore, there is an urgent need in this field for a high-precision identification system and method that can overcome the above-mentioned shortcomings, comprehensively utilize multi-source remote sensing data, and model the spatiotemporal evolution of seepage. Summary of the Invention

[0013] To address the aforementioned technical issues, this application proposes a system, method, apparatus, and medium for identifying and warning of surface response to aquifer seepage under coal seam mining conditions, which integrates InSAR.

[0014] The technical solution adopted in this application is: a surface response identification and early warning system for aquifer seepage under coal seam mining conditions that integrates InSAR, including:

[0015] Multi-source feature extraction module: used to extract surface deformation features, optical remote sensing features, and topographic features to obtain a multi-source heterogeneous feature set. Then, based on the physical mechanism between surface deformation features, optical remote sensing features, and topographic features, an index feature set is constructed to reflect the seepage mechanism of underground aquifers under coal seam mining conditions.

[0016] Multi-source feature fusion module: This module is used to standardize the features extracted by the multi-source heterogeneous feature extraction module, and to group and weight them according to the physical processes they reflect, forming a multi-channel fusion feature map with clear physical meaning as a physical prior feature.

[0017] The seepage response recognition module uses a deep learning network consisting of a dynamic snake convolutional encoder, a physically guided bidirectional cross-attention bridging module, and a multi-scale receptive field adaptive decoder to process multi-channel fused feature maps and generate a high-precision seepage probability distribution map with clear boundaries.

[0018] Spatiotemporal evolution modeling module: Based on the physical information neural operator, a spatiotemporal state field is constructed, the groundwater diffusion equation is introduced as a physical constraint, and combined with multi-scale time branches, the future evolution trend of the seepage area is predicted, and a seepage evolution trend prediction map is generated.

[0019] Risk warning output module: Generates risk level maps and warning information based on seepage probability distribution maps and seepage evolution trend prediction maps.

[0020] Furthermore, surface deformation characteristics include deformation rate, deformation gradient, and deformation fluctuation;

[0021] Optical remote sensing features include Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Land Surface Temperature (LST).

[0022] Topographic features include slope, aspect, and topographic humidity index (TWI).

[0023] The index feature set includes a deformation-humidity response index for quantifying the temporal correlation between surface deformation and surface humidity changes, a temperature-vegetation anomaly index for highlighting “cold and wet” anomalies that are inconsistent with the regional context and may be caused by seepage, and an improved confluence index that takes into account slope aspect for more accurate prediction of shallow aquifer accumulation tendency.

[0024] Furthermore, the principle of grouped weighted fusion is as follows: the surface deformation features, optical remote sensing features, topographic features, and index feature sets are divided into three feature groups according to the physical processes they reflect: deformation response group, surface state group, and topographic control group. The weight vector of each feature in each feature group is calculated, and feature fusion is performed according to priority, where the priority is set based on hydrogeological knowledge: deformation response group > surface state group > topographic control group.

[0025] Furthermore, the deep learning network is a symmetric encoder-decoder network structure. The dynamic snake convolution encoder contains four downsampling stages. Each downsampling stage performs dynamic snake convolution, batch normalization, ReLU and 2×2 max pooling twice. The dynamic snake convolution encoder uses snake convolution with learnable offsets and adds continuity regularization constraints to smoothly extend the sampling points.

[0026] The bidirectional cross-attention bridging module performs bidirectional cross-attention interaction between the physical prior features calculated by the multi-source feature fusion module and the depth feature map output by the downsampling stage at the end of the dynamic snake convolutional encoder to obtain deep fused features.

[0027] The multi-scale receptive field adaptive decoder comprises four upsampling stages. Each upsampling stage first upsamples through a 2×2 transposed convolution and concatenates it with the corresponding features from the dynamic snake-like convolutional encoder, then runs two branches in parallel:

[0028] The large receptive field branch uses serial dilated convolutions with dilation rates of 3, 5, and 7 to capture the global context;

[0029] The detail-preserving branch uses a standard 3×3 deformable convolution to precisely preserve edge details;

[0030] The outputs of the two branches are dynamically fused using learnable weights.

[0031] Furthermore, the physical constraint loss function is a composite loss function that incorporates physical constraints, and its expression is:

[0032] ;

[0033] in, The standard binary cross-entropy loss; The Dice loss is used to optimize the overlap between the predicted and real regions. These are the weighting coefficients for the Dice loss; To achieve boundary-weighted loss, higher weights are assigned to the boundary pixels of the true labels before loss calculation, thus focusing the learning on boundary regions. The weighting coefficients for the boundary-weighted loss; For physical consistency loss, Here are the weighting coefficients for the physical consistency loss, and the expression for the physical consistency loss is:

[0034] ;

[0035] The first term in the above formula is a mechanism consistency constraint, which constrains the seepage probability predicted by the network. Should be related to the deformation-humidity response index There is a positive correlation. This serves as a unique index for each individual cell within the study area that participates in the calculation of physical consistency loss. For the Sigmoid function, The first term is a learnable scaling factor; the second term is a terrain rationality penalty, applied to the predicted seepage probability located in a high-slope area. Apply additional penalties, where pixels in high-slope regions are set express, For pixels The slope value at that location, This is the penalty coefficient.

[0036] A method for identifying and warning of surface response to aquifer seepage under coal seam mining conditions using InSAR, comprising the following steps:

[0037] S1: Multi-source feature extraction;

[0038] S2: Multi-source feature fusion;

[0039] S3: Construction of seepage probability distribution map;

[0040] S4: Construction of seepage evolution trend prediction map;

[0041] S5: Output of seepage risk level map and early warning information.

[0042] Furthermore, step S3 specifically includes:

[0043] S31: Multi-scale feature encoding and progressive feature decoding: A symmetrical encoder-decoder network structure is adopted to achieve feature abstraction and high-resolution prediction, and to obtain feature maps after decoding and detail restoration.

[0044] S32: Physics-guided network training and loss calculation: A composite loss function that incorporates physical constraints is used to train the network model;

[0045] S33: Use the trained network model to output the past The permeability probability distribution map of each time step is generated, and the generated permeability probability distribution map is post-processed and optimized.

[0046] Furthermore, step S4 specifically includes:

[0047] S41: Spatiotemporal State Field Construction: Input the past output of step S3 The sequence of seepage probability distribution maps at each time step, along with the static physical field data provided in steps S1 and S2, are used to discretize the study area into regular grid cells, with each grid cell serving as a state point. The data at each state point at time step S1 are then used to construct the flow probability distribution at each time step S2. The state vector is used to obtain the historical spatiotemporal state field of the entire region, which serves as the basis for subsequent evolutionary modeling.

[0048] S42: Evolutionary mapping learning based on physical information neural operators: Input the historical spatiotemporal state field sequence output in step S41, and learn a mapping function from the historical spatiotemporal state field to the future spatiotemporal state field using the physical information neural operator framework;

[0049] The physical information neural operator framework has three parallel time branches to handle short-term, medium-term and long-term evolution patterns respectively. The outputs of each time branch are fused through learnable gating parameters to finally output the predicted spatiotemporal state field of the next moment.

[0050] Furthermore, the groundwater seepage and diffusion equation is introduced as a soft constraint during the training process of the physical information neural operator framework.

[0051] S43: Prediction result extraction and output: Input the spatiotemporal state field of the next moment output in step S42, extract the seepage probability component from the state vector of each state point to form a seepage probability distribution prediction map of the whole region; and output the seepage probability distribution prediction map of the next time unit.

[0052] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0053] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0054] The advantages of this application over the prior art are as follows:

[0055] 1. Significantly improves the accuracy of aquifer seepage zone identification under coal seam mining conditions, outputting results with clear boundaries and spatial continuity:

[0056] This application constructs an index feature set with clear hydrogeological and physical significance (deformation-humidity response index, temperature-vegetation anomaly index, and improved confluence index considering slope aspect), and adopts a grouping weighted fusion strategy based on physical mechanisms to enhance the synergistic effect of multi-source features. Furthermore, by combining a dynamic serpentine convolutional encoder (fitting to curved seepage channels), a physically guided bidirectional cross-attention bridging module (aligning the spatiotemporal misalignment of physical quantities such as deformation, humidity, and temperature), and a physically constrained loss function (suppressing spurious responses caused by non-seepage), the final output seepage probability distribution map significantly outperforms existing methods relying solely on single deformation features in terms of boundary clarity, regional integrity, and spatial continuity, avoiding fragmentation and boundary ambiguity issues.

[0057] 2. Achieve high-precision, physically consistent prediction of the spatiotemporal evolution of aquifer seepage under coal seam mining conditions:

[0058] This application models seepage evolution as a dynamic evolution process of a spatiotemporal state field, employing a physical information neural operator framework. This framework is resolution-invariant and can accept inputs and outputs with arbitrary spatial resolutions. By introducing the groundwater diffusion equation as a soft constraint, the prediction results automatically follow the basic principles of seepage diffusion. Simultaneously, three parallel time branches—short-term, medium-term, and long-term—are set up, and through learnable gating parameters, the evolutionary patterns at different time scales are effectively captured. Compared to traditional time series models (Autoregressive Moving Average (ARIMA) and Long Short-Term Memory (LSTM) networks) or fixed-grid methods, the prediction results of this application more closely match the actual development process of aquifer seepage under coal seam mining conditions, providing a reliable basis for dynamic early warning.

[0059] 3. Enhanced ability to identify weak-response seepage areas, significantly reducing the false negative rate:

[0060] This application simultaneously extracts surface deformation, optical remote sensing data (water index, vegetation index, and surface temperature), and topographic features to construct a multi-source heterogeneous feature set. By standardizing the dimensions using Z-scores and performing weighted fusion based on physical mechanisms, the model becomes highly sensitive to indirect signals such as subtle changes in humidity, vegetation anomalies, and surface temperature fluctuations caused by aquifer seepage. Compared to InSAR methods that rely solely on deformation features or spatially sparse wellpoint measurements, this application significantly improves the recall rate for identifying seepage areas with weak deformation and weak response, effectively reducing the risk of missed detections.

[0061] 4. Achieve fully automated monitoring and early warning throughout the entire process to meet the needs of mining area engineering applications:

[0062] This application constructs an end-to-end automated processing flow, from inputting raw SAR / optical / DEM data, to feature extraction, fusion, seepage identification, spatiotemporal evolution prediction, and finally to risk level determination and structured early warning information generation. It can periodically output seepage probability distribution maps, evolution trend maps, and risk level maps without manual intervention, and automatically trigger early warnings when preset conditions are met (such as the proportion of high-risk areas, spatial expansion trends, and distance from important facilities). This design significantly improves the timeliness and operability of monitoring and is suitable for long-term, large-scale mine disaster monitoring. Attached Figure Description

[0063] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0064] Figure 1 System block diagram provided for embodiments of this application;

[0065] Figure 2 A flowchart illustrating the method provided in this application embodiment;

[0066] Figure 3 The images show the results of the seepage regions identified in four different sets of images using the system of this application and different ablation experiments;

[0067] Figure 4 This is a block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0068] like Figures 1 to 4 As shown, this application provides a surface response identification and early warning system for aquifer seepage under coal seam mining conditions that integrates InSAR, including:

[0069] Multi-source feature extraction module: used to extract surface deformation features, optical remote sensing features, and topographic features to obtain a multi-source heterogeneous feature set. Then, based on the physical mechanism between surface deformation features, optical remote sensing features, and topographic features, an index feature set is constructed to reflect the seepage mechanism of underground aquifers under coal seam mining conditions. Among them, surface deformation features are extracted from time-series SAR images, optical remote sensing features are extracted from optical remote sensing images, and topographic features are extracted from digital elevation model (DEM) data.

[0070] Multi-source feature fusion module: This module is used to standardize the features extracted by the multi-source heterogeneous feature extraction module, and to group and weight them according to the physical processes they reflect, forming a multi-channel fusion feature map with clear physical meaning as a physical prior feature.

[0071] The seepage response recognition module employs a deep learning network consisting of a dynamic serpentine convolutional encoder, a physically guided bidirectional cross-attention bridging module, and a multi-scale receptive field adaptive decoder to process multi-channel fused feature maps and generate a high-precision seepage probability distribution map with clear boundaries.

[0072] Spatiotemporal evolution modeling module: Based on the physical information neural operator, a spatiotemporal state field is constructed, the groundwater diffusion equation is introduced as a physical constraint, and combined with multi-scale time branches, the future evolution trend of the seepage area is predicted, and a seepage evolution trend prediction map is generated.

[0073] Risk warning output module: Generates risk level maps and warning information based on seepage probability distribution maps and seepage evolution trend prediction maps.

[0074] By sequentially connecting and coordinating the above five core modules, a complete aquifer seepage identification and early warning system under coal seam mining conditions is constructed, realizing automated processing from raw SAR data to seepage early warning information, and providing an effective technical means for aquifer seepage monitoring under coal seam mining conditions.

[0075] The multi-source feature extraction module processes the input time-series SAR imagery to obtain a surface deformation rate map and deformation time-series data. In this embodiment, the mature SBAS-InSAR processing workflow can be used, and preferably, the following processing steps are performed using SARscape software.

[0076] Data preprocessing: Input no less than 20 SAR images, use the Delaunay triangulation construction function in SARscape software to generate an interferometric pair combination network, set the screening conditions of time baseline ≤100 days and spatial baseline ≤150 meters, use precise orbit data to perform orbit fine correction, and control the registration accuracy within 0.1 pixels;

[0077] Interferometry processing: Differential interferograms were generated using SARscape software, and noise suppression was performed using the Goldstein filtering algorithm with filter parameter α. g The value is set to 0.7, and the influence of terrain phase is removed through elevation phase simulation, followed by atmospheric phase correction.

[0078] Deformation inversion: Phase unwrapping was performed based on the minimum cost flow algorithm in SARscape software. The unwrapping coherence coefficient threshold was set to 0.3. Deformation observation equations were constructed, and deformation time series data were solved by singular value decomposition.

[0079] Deformation feature extraction: Calculate the annual average deformation rate based on the surface deformation rate map and generate deformation rate contour maps, extract deformation gradient features, and calculate deformation fluctuation features based on deformation time series data.

[0080] The extraction process of optical remote sensing features is as follows: Normalized water index (NDWI) and normalized vegetation index (NDVI) are extracted from Landsat-8 / 9 images, and land surface temperature (LST) is obtained by radiometric calibration, atmospheric correction and single-window algorithm.

[0081] The process for extracting terrain features is as follows: calculate slope, aspect, and topographic humidity index (TWI) based on 30-meter resolution DEM data.

[0082] The index feature set is an index feature set with clear hydrogeological indicative significance constructed based on the physical mechanism of surface response to underground aquifer seepage under coal seam mining conditions. This index feature set includes the following indices:

[0083] Deformation-humidity response index This is used to quantify the temporal correlation between changes in land surface deformation and land surface moisture, and its expression is as follows:

[0084] ;

[0085] in, For the two-dimensional spatial coordinates of a pixel, For pixels in Surface deformation at time, This represents the NDWI value at the corresponding time. , Surface deformation and NDWI values ​​in the study time series, respectively. The mean within. High The value indicates that surface deformation may be dominated by underground aquifer activity.

[0086] Temperature-Vegetation Anomaly Index This is used to highlight "cold and wet" anomalies that are inconsistent with the regional background and may be caused by seepage. Its expression is as follows:

[0087] ;

[0088] in, and The value is the result of time-series Z-score normalization. and This represents the spatial mean of LST and NDVI across the entire study area during the research period.

[0089] Improved confluence index considering slope aspect This is used to more accurately predict the accumulation tendency of shallow groundwater aquifers, and its expression is as follows:

[0090] ;

[0091] in, The upstream catchment area, For slope, To prevent zero small constant, In terms of water flow direction, It is a slope.

[0092] To address the challenges of slender, curved seepage areas and spatiotemporal misalignment of physical responses in mining areas, the seepage response identification module incorporates a dynamic serpentine convolutional encoder, a physically guided bidirectional cross-attention bridging module, and a multi-scale receptive field adaptive decoder.

[0093] Dynamic serpentine convolution encoder: Traditional square convolution kernels struggle to conform to curved seepage channels. This application's dynamic serpentine convolution encoder employs a serpentine convolution with learnable offsets and incorporates continuity regularization constraints to smoothly extend sampling points. Let the standard convolution kernel sampling position be... The learnable offset is The actual sampling location is The learnable offsets are dynamically generated by the offset generator in the dynamic serpentine convolution based on the current feature map, and continuity regularization is introduced. Forced smoothing of adjacent sampling point offsets prevents the convolution kernel from "jumping out" of the seepage channel. For the continuity regularization loss term, For the first A learnable two-dimensional offset vector at each sampling location. This is the index of the sampling point within the convolution kernel. The dynamic serpentine convolution encoder consists of four downsampling stages, each of which performs dynamic serpentine convolution, batch normalization, ReLU, and 2×2 max pooling twice.

[0094] A physics-guided bidirectional cross-attention bridging module addresses the temporal misalignment among different physical quantities (deformation, humidity, temperature) in a mining area, which is difficult to align using a single data-driven approach. This module facilitates bidirectional cross-attention interaction between the physical prior features calculated by the multi-source feature fusion module and the deep features.

[0095] First, map the physical priors as... Then calculate:

[0096] Feature to Physical Attention This is used to retrieve seepage-related information from prior knowledge;

[0097] Physical to Feature Attention It is used to correct spurious responses in the data (such as surface deformation caused by non-seepage).

[0098] Fusion features Then, through spatial and channel-based collaborative gating output, a deep integration of physical mechanisms and data characteristics is achieved.

[0099] In the above formula: This is an embedding mapping representation of physical prior features. The query matrix is ​​generated from the deep features of the encoder. The key matrix is ​​generated by physical prior embedding. The dimension of the embedding vector is used to scale the dot product attention and prevent gradient vanishing. The Value matrix is ​​generated by physical prior embedding. The query matrix is ​​generated by physical prior embedding. The key matrix is ​​generated from the deep features of the encoder. The Value matrix is ​​generated from the deep features of the encoder. This is the depth feature map output at the end of the downsampling stage (i.e., the fourth downsampling stage) of the dynamic snake convolutional encoder.

[0100] Multi-scale receptive field adaptive decoder: Under coal seam mining conditions, underground aquifer seepage is controlled by both the large-scale water catchment background and the details of local fractures. This multi-scale receptive field adaptive decoder consists of four upsampling stages. Each upsampling stage first upsamples using a 2×2 transposed convolution and concatenates it with the corresponding features from a dynamic serpentine convolutional encoder. Then, two branches run in parallel: a large receptive field branch uses serial dilated convolutions with dilation rates of 3, 5, and 7 to capture the global context; a detail-preserving branch uses standard 3×3 deformable convolutions to accurately preserve edge details. The outputs of both branches are processed with learnable weights. Dynamic fusion: ,in The initial value is 0.5, which is adaptively adjusted during training to balance the overall coherence of the seepage region and the clarity of its boundaries. The multi-scale receptive field adaptive decoder's end features are concatenated with the first-stage features of the dynamic serpentine convolutional encoder and reduced to 16 channels for output. For subsequent use. This is the output feature map after dynamic fusion during the current upsampling stage. The feature map output after processing the large receptive field branch through serial dilated convolution (dilation rate of 3, 5, 7) is shown. To preserve details, the branch is processed by a standard 3×3 deformable convolution to output a feature map.

[0101] This application also proposes a method for identifying and warning of surface response to aquifer seepage under coal seam mining conditions by integrating InSAR, implemented based on the above system, such as... Figure 2 As shown, the main steps are as follows:

[0102] S1: Multi-source feature extraction, including:

[0103] S11: Surface Deformation Feature Extraction: Extracting surface deformation features from time-series SAR image data;

[0104] S12: Optical remote sensing feature extraction: Normalized water index NDWI and normalized vegetation index NDVI are extracted from Landsat-8 / 9 remote sensing images acquired synchronously with SAR images, and the land surface temperature LST is obtained by inversion through radiometric calibration, atmospheric correction and single window algorithm.

[0105] S13: Topographic feature extraction: Calculate slope, aspect, and topographic humidity index (TWI) from DEM data acquired synchronously with SAR imagery based on 30-meter resolution.

[0106] S14: Based on the physical mechanism of surface response to seepage in underground aquifers under coal seam mining conditions, construct an index feature set with clear hydrogeological indication significance.

[0107] S2: Multi-source feature fusion, including:

[0108] S21: The extracted surface deformation features, optical remote sensing features, and topographic features are processed using the Z-score normalization method. For any feature value... Its standardized value The calculation formula is:

[0109] ;

[0110] in, This is the spatial mean of the feature within the study area. Its spatial standard deviation.

[0111] After processing, all features are unified to a spatial resolution of 30 meters and a WGS84 UTM geographic coordinate system.

[0112] S22: Mechanism-guided feature fusion: This application proposes a feature fusion strategy based on physical mechanism grouping and weighting, and the specific implementation steps are as follows:

[0113] S221: Feature Grouping: The 12 feature channels to be fused are divided into three feature groups based on the core physical processes they reflect. :

[0114] Deformation response group Includes deformation rate, deformation gradient, deformation fluctuation, ;

[0115] Surface state group Includes Normalized Difference Water Index (NDWI), Normalized Difference Vegetation Index (NDVI), and Land Surface Temperature (LST). ;

[0116] Terrain Control Group Includes slope, aspect, and topographic humidity index (TWI). .

[0117] S222: In-group adaptive weighting: For each feature group A lightweight channel attention submodule is applied to each feature, and the weight vectors of each feature within the group are calculated. .

[0118] S223: Inter-group priority fusion: Fusion priority is set based on hydrogeological understanding: Deformation response group > Surface state group > Topographic control group. This is achieved through a learnable scaling factor. and The three weighted features are fused to obtain the final multi-channel fused feature map. :

[0119] ;

[0120] in, This represents a channel-weighted combination based on weights. This represents the weight vector of each feature within the deformation response group. This represents the weight vector of each feature within the surface state group. This represents the weight vector of each feature within the terrain control group. Scaling factor. and These are the parameters for model training, and their initial values ​​can be set based on prior knowledge.

[0121] The multi-channel fusion feature map output in this step This refers to the physical prior features with clear physical meaning provided to the subsequent seepage response identification module.

[0122] S3: Multi-channel fusion feature map The process yields a seepage probability distribution map, including:

[0123] S31: Multi-scale Feature Encoding and Progressive Feature Decoding: This step achieves feature abstraction and high-resolution prediction through a symmetric encoder-decoder network structure, including:

[0124] S311: Encoding Process: Each downsampling stage of the dynamic serpentine convolutional encoder performs the following operations: two dynamic serpentine convolutions, batch normalization, ReLU, and 2×2 max pooling. The number of feature channels output by the four downsampling stages are 64, 128, 256, and 512, respectively.

[0125] S312: Physically Guided Bidirectional Cross-Attention Interaction: Fusing Multi-Channel Feature Maps The depth feature map output by the downsampling stage at the end of the dynamic snake convolutional encoder By performing bidirectional cross-attention interaction, deep fusion features are obtained.

[0126] S312: Decoding Process: Each upsampling stage of the multi-scale receptive field adaptive decoder performs the following operations: First, upsampling is performed using a 2×2 transposed convolution and concatenated with the corresponding features from the dynamic snake-like convolutional encoder. Then, two branches run in parallel: the large receptive field branch uses serially dilated convolutions with dilation rates of 3, 5, and 7 to capture the global context; the detail-preserving branch uses standard 3×3 deformable convolutions to precisely preserve edge details. The outputs of both branches are processed with learnable weights. Dynamic fusion: .

[0127] S313: Detail recovery: At the end of the multi-scale receptive field adaptive decoder, the final decoded feature map is concatenated with the high-resolution feature map (shallow features) of the first stage of the dynamic snake convolutional encoder and reduced to 16 channels.

[0128] S32: Physically Guided Network Training and Loss Calculation: To guide the network in learning seepage patterns that conform to hydrogeological laws, a composite loss function incorporating physical constraints is used for training. Total Loss Defined as:

[0129] ;

[0130] in, The standard binary cross-entropy loss; The Dice loss is used to optimize the overlap between the predicted and real regions. These are the weighting coefficients for the Dice loss; To achieve boundary-weighted loss, higher weights are assigned to the boundary pixels of the true labels before loss calculation, thus focusing the learning on boundary regions. The weighting coefficients for the boundary-weighted loss; For physical consistency loss, Here are the weighting coefficients for the physical consistency loss, which is the core innovation of this step. Its expression is:

[0131] ;

[0132] The first term in the above formula is a mechanism consistency constraint, which constrains the seepage probability predicted by the network. It should be consistent with the deformation-humidity response index calculated in step S2. There is a positive correlation. This serves as a unique index for each individual cell within the study area that participates in the calculation of physical consistency loss. For the Sigmoid function, The first term is a learnable scaling factor. The second term is a terrain rationality penalty, applied to the predicted seepage probability located in a high-slope region. Apply additional penalties, where pixels in high-slope regions are set express, For pixels The slope value at the location (derived from the terrain features in step S2). This is the penalty coefficient.

[0133] S33: Seepage Probability Distribution Map Generation and Post-processing Optimization: This step converts the network output into the final recognition result, including:

[0134] S331: Probabilistic Map Generation: The feature map after decoding and detail restoration is input into a 1×1 convolutional layer, the number of channels is reduced to 1, and then passed through a Sigmoid activation function to output the percolation probability value of each pixel. .

[0135] S332: Adaptive Threshold Segmentation: The global adaptive threshold of the probability map is calculated using the Otsu's method, which transforms the probability map into an initial binary segmentation map.

[0136] S333: Morphological Closure Operation: Performs a morphological closure operation on the initial binary segmentation image, which is to first perform a dilation operation and then an erosion operation, using 3×3 circular structuring elements to fill small holes and connect adjacent small fragments.

[0137] S334: Small area filtering: Calculate the area of ​​all connected regions (patches) and remove isolated patches with an area smaller than a preset threshold (e.g., equivalent to 5 pixels), classifying them as background.

[0138] S335: Spatial Neighbor Patch Merging: For two independent seepage patches whose spatial Euclidean distance is less than a preset threshold and whose average feature vector cosine similarity is higher than 0.8, merge them into a single whole.

[0139] By executing the above three steps (steps S31-S33) in sequence, the multi-source fusion features can be fully utilized to output a seepage probability distribution map with clear boundaries, spatial continuity and conformity to hydrophysical laws, providing high-quality input for the spatiotemporal evolution modeling module.

[0140] S4: The spatiotemporal evolution modeling module processes the seepage probability distribution map to obtain a seepage evolution trend prediction map. This module employs a neural operator framework, introducing the groundwater diffusion equation as a physical constraint to achieve resolution-adaptive and physically consistent spatiotemporal prediction, including:

[0141] S41: Spatiotemporal State Field Construction: Input the past output of step S3 time steps ( Take a sequence of seepage probability distribution maps (12 months). The study area is discretized into regular grid cells, with each grid cell serving as a state point. The static physical field data provided in steps S1 and S2 (including terrain slope, aspect, terrain humidity index, and improved confluence index considering aspect) is then constructed. At any moment state vector :

[0142] ;

[0143] in: For the current moment The probability of seepage; For the current moment The spatial gradient reflects the direction of seepage diffusion; The rate of change over time is approximated by the difference between adjacent time points; It is a static physical feature vector.

[0144] Output: Historical spatiotemporal state field of the entire region This serves as the foundation for subsequent evolutionary modeling. The seepage evolution problem is transformed into a dynamic evolution problem of the state field, and physical variables are introduced to lay the foundation for subsequent physical constraints.

[0145] S42: Evolutionary Mapping Learning Based on Physical Information Neural Operators: Input the historical spatiotemporal state field sequence output from step S41 A physical information neural operator framework is used to learn a mapping function from the historical spatiotemporal state field to the future spatiotemporal state field. :

[0146] ;

[0147] This physical information neural operator framework has the following characteristics:

[0148] 1. Resolution invariance: The operator acts directly on the function space, and can accept input and output with arbitrary spatial resolution, avoiding the dependence of the original image convolution method on fixed grid topology.

[0149] 2. Physical Information Embedding: During model training, the groundwater seepage diffusion equation is introduced as a soft constraint. For any state point, the time rate of change of the seepage probability should approximately satisfy:

[0150] ;

[0151] in: The diffusion coefficient (which can be obtained from the physical field or learned by the network) It is a continuous seepage probability field, consisting of the probability value of seepage occurring in each pixel at a certain moment, output by the S3 module. For divergence operators, The spatial gradient of the seepage probability represents the direction in which the seepage probability changes most rapidly in space and the magnitude of its rate of change. This constraint is added to the total loss function in the form of physical residual loss.

[0152] 3. Multi-scale temporal branching modeling: Internally, three parallel time branches are configured to handle short-term (3 months), medium-term (12 months), and long-term (36 months) evolutionary patterns respectively. The output of each branch is processed by learnable gating parameters. Fusion:

[0153] ;

[0154] in, For the model to predict the next time step ( (Time) The spatiotemporal state field of the entire region, For the first The historical spatiotemporal state field sequence corresponding to each time branch For the first Each time branch corresponds to a learnable parameter Physical information neural operator mapping function, For the index number of the multi-scale time branch, The values ​​1, 2, and 3 correspond to the short-term (3 months), medium-term (12 months), and long-term (36 months) evolutionary pattern processing branches, respectively.

[0155] Output: The predicted spacetime state field for the next time step. .

[0156] The purpose of this step is to learn the nonlinear spatiotemporal evolution of the seepage probability field. Physical constraints ensure that the prediction results conform to the basic principles of hydrogeology, and multi-scale time branches enhance the ability to capture the evolution patterns at different time scales.

[0157] S43: Prediction Result Extraction and Output: Input the spatiotemporal state field of the next time step output from step S42. .

[0158] Processing: Extract the seepage probability component from the state vector of each state point. A prediction map of the seepage probability distribution across the entire region. Optionally, the predicted graph can be post-processed (Gaussian filtering) to eliminate numerical noise.

[0159] Output: A predicted distribution map of seepage probability for the next one time unit (the next month). The spatial resolution remains consistent with the input (30 meters).

[0160] The purpose of this step is to separate the prediction results from the state field and generate the evolution trend data required by the risk warning output module.

[0161] Through the sequential execution of the above three steps (steps S41-S43), the spatiotemporal evolution modeling module, with physical information neural operators at its core, integrates physical diffusion equation constraints and multi-scale time modeling to achieve spatiotemporal evolution prediction of the seepage probability field. The spatiotemporal evolution modeling module has advantages such as resolution adaptation, physical consistency, and multi-scale fusion. Its output results can be directly used for risk level determination and early warning triggering in the risk warning output module.

[0162] S5: The risk warning output module processes the seepage probability distribution map and seepage evolution trend prediction map to obtain a seepage risk level map and warning information, including:

[0163] S51: Based on the seepage probability values ​​of each pixel in the seepage probability distribution map And the mean predicted seepage probability for the next month provided by the seepage evolution trend prediction chart. The risk level of each grid cell is determined.

[0164] S52: Calculate the trend quantification index: For each grid cell, calculate the predicted rate of change of its seepage probability. :

[0165] ;

[0166] like ,but Replace it with the predicted absolute value.

[0167] S53: Risk Level Classification: Based on the seepage probability value and evolution trend, the risk is divided into three levels.

[0168] Low risk: and ;

[0169] Medium risk: or ;

[0170] High risk: and .

[0171] S54: Warning Trigger: Three trigger conditions are set, including the high-risk area accounting for more than 5% of the total study area, the high-risk area showing a spatial expansion trend for three consecutive months, and the high-risk area being less than 500 meters away from important facilities in the mining area. A warning will be triggered if any one of the conditions is met.

[0172] S55: Early Warning Information Generation: The system automatically extracts the core attribute information of all grid cells or connected areas identified as high-risk, generating a structured electronic data table. This electronic data table generates one record for each target area, and each record contains at least the following fields: area identifier, coordinates of the center point representing its geographical location, current seepage probability value, risk level, quantitative indicator of predicted trend, area, and name and distance to the nearest key facility in the mining area.

[0173] The advantages of this application will be further explained below based on some comparative experiments and ablation experiments between this application and the prior art.

[0174] Advantage 1: The accuracy of identifying seepage zones in underground aquifers under coal seam mining conditions is significantly improved, with clear boundaries and spatial continuity. Refer to Table 1 below for explanation.

[0175] Table 1. Intersection over Union (IOU) and Boundary F1 Comparison Table.

[0176] .

[0177] In Table 1: Experiment 1 Explanation: The exponential features are directly related to the physical response mechanism of underground aquifer seepage under coal seam mining conditions. After removal, the model can only rely on basic features such as the original deformation rate and NDWI. It loses the confluence information of deformation and humidity time-series correlation, temperature and vegetation anomalies, and slope aspect correction, which greatly weakens the ability to characterize the seepage area of ​​underground aquifer under coal seam mining conditions, reduces the identification accuracy, and blurs the boundaries.

[0178] Experiment 2 Explanation: Grouped weighted fusion divides features into three groups—deformation response, surface condition, and topographic control—based on hydrogeological logic, and sets priorities and learnable weights, which can suppress the interference of redundant information and highlight key features. After switching to ordinary stitching, each feature channel is treated equally, noise and irrelevant feature interference are enhanced, and the model has difficulty focusing on the core seepage response, resulting in a significant decrease in the Intersection over Union (IOU) and boundary F1 scores.

[0179] Experiment 3 Explanation: Dynamic serpentine convolution, through learnable offsets and continuity regularization, enables the convolution kernel to extend smoothly along the curved and slender seepage channels. When using a standard square convolution kernel, the convolution sampling points cannot accurately match the actual seepage morphology, easily "jumping out" of the channel or missing edge details, thus reducing the fitting accuracy for curved seepage regions.

[0180] Experiment 4 Explanation: The bidirectional cross-attention bridging module is used to align the spatiotemporal misalignments between physical quantities such as deformation, humidity, and temperature. After removal, the interaction between data features and physical priors is severed. The model cannot use physical priors to correct spurious responses in the data, leading to a decrease in the ability to identify disturbances such as surface deformation caused by non-seepage, resulting in a slight loss of accuracy.

[0181] Experiment 5 Explanation: The physical constraint loss function includes a mechanism consistency term and a terrain rationality penalty term. The former constrains the positive correlation between the predicted probability and the deformation humidity response index, while the latter penalizes misjudgments in high-slope areas. After removing this loss, the network is no longer constrained by hydrogeological laws, making it prone to generating false seepage areas in physically unreasonable locations, leading to increased fragmentation of recognition results and poor boundary continuity.

[0182] Advantage 2: High-precision, physically consistent prediction of seepage spatiotemporal evolution. See Table 2 below for explanation.

[0183] Table 2 Comparison of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).

[0184] .

[0185] In Table 2:

[0186] Experiment 6 Explanation: The groundwater diffusion equation, acting as a soft constraint, guides the physical information neural operator to learn an evolutionary pattern that conforms to the physical laws of seepage. Removing this constraint renders the model entirely data-driven, making it susceptible to learning statistically spurious correlations. For example, it may misjudge deformation fluctuations caused by non-seepage factors as seepage diffusion trends, leading to increased prediction bias and loss of physical consistency.

[0187] Experiment 7 explains that seepage evolution patterns differ across time scales. Short-term fluctuations may be influenced by rainfall or mining activities, while long-term trends reflect the overall changes in underground aquifer seepage under coal seam mining conditions. Multi-scale branches were used to handle evolution patterns over 3 months, 12 months, and 36 months, and fused using learnable gating. Retaining only the 12-month branch resulted in insufficient ability to capture both short-term sharp fluctuations and long-term slow trends, leading to increased prediction errors.

[0188] Experiment 8 Explanation: The physical information neural operator acts directly on the function space, exhibiting resolution invariance and capable of handling inputs and outputs from arbitrary spatial grids. Fixed-grid convolutional LSTMs can only operate on predefined discrete grids, requiring retraining when the monitoring area or resolution changes. After replacement, the model's spatial generalization ability decreases, leading to poorer accuracy and significantly increased errors in modeling the continuous field of seepage and diffusion in underground aquifers under coal seam mining conditions.

[0189] Experiment 9 Explanation: A long-term input of 12 months provides the model with sufficient samples of seepage evolution, including seasonal variations and interannual trends. After shortening to 6 months, the model cannot fully observe the annual periodicity of the seepage region, resulting in decreased prediction stability and a slight increase in error.

[0190] Advantage 3: Enhanced ability to identify weak-response seepage areas. See Table 3 below for explanation.

[0191] Table 3. Comparison of recall and precision.

[0192] .

[0193] In Table 3:

[0194] Experiment 10 explains that under coal seam mining conditions, seepage from underground aquifers may not cause significant surface deformation in the early or weakly active stages, but it can lead to increased surface humidity, abnormal vegetation, or temperature changes. After removing optical remote sensing features and topographic features, the model relies solely on deformation information and cannot perceive these indirect response signals, resulting in the underestimation of a large number of weakly deformed seepage areas, causing the recall rate to drop sharply from 0.88 to 0.68.

[0195] Experiment 11 Explanation: Features from different sources have different dimensions and distribution ranges. Z-score normalization unifies them to the same scale, ensuring the model treats each feature fairly during training. Grouped weighted fusion further highlights feature channels related to seepage physics. Using ordinary Min-Max normalization without grouping and fusion results in inconsistent feature contributions, weakened response to weak indirect signals, and a significant decrease in recall.

[0196] Figure 4 A structural block diagram of a computer device according to a specific embodiment of this application is shown. Figure 4 As shown, the computer device includes a memory and a processor, the memory storing instructions executable on the processor. When the processor executes the instructions, it implements the methods described in the above embodiments. The number of memories and processors can be one or more. This computer device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The computer device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0197] The computer device may also include a communication interface for communicating with external devices and exchanging data. The devices are interconnected using different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as a display device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). The bus can be divided into address buses, data buses, control buses, etc. For ease of illustration, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0198] Optionally, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, then the memory, processor, and communication interface can communicate with each other through an internal interface.

[0199] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting advanced RISC machines (ARM) architecture.

[0200] This application provides a computer-readable storage medium (such as the memory described above) storing computer instructions that, when executed by a processor, implement the method provided in this application.

[0201] Optionally, the memory may include a stored program area and a stored data area, wherein the stored program area may store the operating system and application programs required for at least one function; the stored data area may store data created based on the use of the computer device for mapping. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the computer device for mapping via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0202] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A surface response identification and early warning system for aquifer seepage under coal seam mining conditions integrating InSAR, characterized in that: include: Multi-source feature extraction module: used to extract surface deformation features, optical remote sensing features, and topographic features to obtain a multi-source heterogeneous feature set. Then, based on the physical mechanism between surface deformation features, optical remote sensing features, and topographic features, an index feature set is constructed to reflect the seepage mechanism of underground aquifers under coal seam mining conditions. Multi-source feature fusion module: This module is used to standardize the features extracted by the multi-source heterogeneous feature extraction module, and to group and weight them according to the physical processes they reflect, forming a multi-channel fusion feature map with clear physical meaning as a physical prior feature. The seepage response recognition module uses a deep learning network consisting of a dynamic snake convolutional encoder, a physically guided bidirectional cross-attention bridging module, and a multi-scale receptive field adaptive decoder to process multi-channel fused feature maps and generate a high-precision seepage probability distribution map with clear boundaries. Spatiotemporal evolution modeling module: Based on the physical information neural operator, a spatiotemporal state field is constructed, the groundwater diffusion equation is introduced as a physical constraint, and combined with multi-scale time branches, the future evolution trend of the seepage area is predicted, and a seepage evolution trend prediction map is generated. Risk warning output module: Generates risk level maps and warning information based on seepage probability distribution maps and seepage evolution trend prediction maps.

2. The surface response identification and early warning system for aquifer seepage under coal seam mining conditions, which integrates InSAR, as described in claim 1, is characterized in that: Surface deformation characteristics include deformation rate, deformation gradient, and deformation fluctuation; Optical remote sensing features include Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Land Surface Temperature (LST). Topographic features include slope, aspect, and topographic humidity index (TWI). The index feature set includes a deformation-humidity response index for quantifying the temporal correlation between surface deformation and surface humidity changes, a temperature-vegetation anomaly index for highlighting "cold and wet" anomalies that are inconsistent with the regional context and may be caused by seepage, and an improved confluence index that takes into account slope aspect for more accurate prediction of shallow aquifer accumulation tendency.

3. The surface response identification and early warning system for aquifer seepage under coal seam mining conditions, which integrates InSAR, as described in claim 2, is characterized in that: The principle of grouped weighted fusion is as follows: the surface deformation features, optical remote sensing features, topographic features, and index feature sets are divided into three feature groups according to the physical processes they reflect: deformation response group, surface state group, and topographic control group. The weight vector of each feature in each feature group is calculated, and the features are fused according to the priority, which is set based on hydrogeological knowledge: deformation response group > surface state group > topographic control group.

4. The surface response identification and early warning system for aquifer seepage under coal seam mining conditions, which integrates InSAR, as described in claim 1, is characterized in that: The deep learning network is a symmetric encoder-decoder network structure. The dynamic snake convolution encoder contains four downsampling stages. Each downsampling stage performs dynamic snake convolution, batch normalization, ReLU and 2×2 max pooling twice. The dynamic snake convolution encoder uses snake convolution with learnable offsets and adds continuity regularization constraints to smoothly extend the sampling points. The bidirectional cross-attention bridging module performs bidirectional cross-attention interaction between the physical prior features calculated by the multi-source feature fusion module and the depth feature map output by the downsampling stage at the end of the dynamic snake convolutional encoder to obtain deep fused features. The multi-scale receptive field adaptive decoder comprises four upsampling stages. Each upsampling stage first upsamples through a 2×2 transposed convolution and concatenates it with the corresponding features from the dynamic snake-like convolutional encoder, then runs two branches in parallel: The large receptive field branch uses serial dilated convolutions with dilation rates of 3, 5, and 7 to capture the global context; The detail-preserving branch uses a standard 3×3 deformable convolution to precisely preserve edge details; The outputs of the two branches are dynamically fused using learnable weights.

5. The surface response identification and early warning system for aquifer seepage under coal seam mining conditions, which integrates InSAR, as described in claim 4, is characterized in that: The physical constraint loss function is a composite loss function that incorporates physical constraints, and its expression is as follows: ； in, The standard binary cross-entropy loss; The Dice loss is used to optimize the overlap between the predicted and real regions. These are the weighting coefficients for the Dice loss; To achieve boundary-weighted loss, higher weights are assigned to the boundary pixels of the true labels before loss calculation, thus focusing the learning on boundary regions. The weighting coefficients for the boundary-weighted loss; For physical consistency loss, Here are the weighting coefficients for the physical consistency loss, and the expression for the physical consistency loss is: ； The first term in the above formula is a mechanism consistency constraint, which constrains the seepage probability predicted by the network. Should be related to the deformation-humidity response index There is a positive correlation. This serves as a unique index for each individual cell within the study area that participates in the calculation of physical consistency loss. For the Sigmoid function, The first term is a learnable scaling factor; the second term is a terrain rationality penalty, applied to the predicted seepage probability located in a high-slope area. Apply additional penalties, where pixels in high-slope regions are set express, For pixels The slope value at that location, This is the penalty coefficient.

6. A method for identifying and warning of surface response to aquifer seepage under coal seam mining conditions using InSAR, comprising the InSAR-integrated surface response identification and warning system for aquifer seepage under coal seam mining conditions as described in any one of claims 1-5, characterized in that: Includes the following steps: S1: Multi-source feature extraction; S2: Multi-source feature fusion; S3: Construction of seepage probability distribution map; S4: Construction of a seepage evolution trend prediction map; S5: Output of seepage risk level map and early warning information.

7. The method for identifying and early warning of surface response to aquifer seepage under coal seam mining conditions by incorporating InSAR, as described in claim 6, is characterized in that: Step S3 specifically includes: S31: Multi-scale feature encoding and progressive feature decoding: A symmetrical encoder-decoder network structure is adopted to achieve feature abstraction and high-resolution prediction, and to obtain feature maps after decoding and detail restoration. S32: Physics-guided network training and loss calculation: A composite loss function that incorporates physical constraints is used to train the network model; S33: Use the trained network model to output the past The permeability probability distribution map of each time step is generated, and the generated permeability probability distribution map is post-processed and optimized.

8. The method for identifying and early warning of surface response to aquifer seepage under coal seam mining conditions by incorporating InSAR, as described in claim 6, is characterized in that: Step S4 specifically includes: S41: Spatiotemporal State Field Construction: Input the past output of step S3 The sequence of seepage probability distribution maps at each time step, along with the static physical field data provided in steps S1 and S2, are used to discretize the study area into regular grid cells, with each grid cell serving as a state point. The data at each state point at time step S1 are then used to construct the flow probability distribution at each time step S2. The state vector is used to obtain the historical spatiotemporal state field of the entire region, which serves as the basis for subsequent evolutionary modeling. S42: Evolutionary mapping learning based on physical information neural operators: Input the historical spatiotemporal state field sequence output in step S41, and learn a mapping function from the historical spatiotemporal state field to the future spatiotemporal state field using the physical information neural operator framework; The physical information neural operator framework has three parallel time branches to handle short-term, medium-term and long-term evolution patterns respectively. The outputs of each time branch are fused through learnable gating parameters to finally output the predicted spatiotemporal state field of the next moment. Furthermore, the groundwater seepage and diffusion equation is introduced as a soft constraint during the training process of the physical information neural operator framework. S43: Prediction result extraction and output: Input the spatiotemporal state field of the next moment output in step S42, extract the seepage probability component from the state vector of each state point to form a seepage probability distribution prediction map of the whole region; and output the seepage probability distribution prediction map of the next time unit.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 6-8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 6-8.

Images