Deep learning landslide identification method and system fusing geological and mining prior knowledge

Abstract

The application discloses a kind of deep learning landslide identification method and system of fusing geology and mining prior knowledge, it is related to geological disaster remote sensing identification and deep learning technical field, the method includes: constructing the geological and mining prior knowledge graph of target coal mine area, and according to prior knowledge graph generates prior feature map;Acquire the remote sensing image to be identified, prior feature map is registered with remote sensing image in space, the center point of landslide candidate area in the remote sensing image is extracted using lightweight network, and landslide candidate area is obtained by adaptive cropping;The segmentation model after target domain meta migration learning training is input into landslide candidate area, and the landslide probability graph of each candidate area is obtained;Through channel attention mechanism, the landslide probability graph under multiple prior assumptions is adaptively fused, and the final landslide identification mask is generated, and the landslide disaster identification result is output according to landslide identification mask.In this way, the recognition accuracy and generalization ability are improved under the condition of sample scarcity.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing identification of geological hazards and deep learning technology, and in particular to a deep learning method and system for landslide identification that integrates prior knowledge of geology and mining. Background Technology

[0002] Currently, intelligent identification of landslide hazards using high-resolution remote sensing imagery is an important means of disaster prevention and mitigation. However, in specific scenarios such as coal mining areas, traditional deep learning methods still face the following serious challenges:

[0003] (1) Extremely scarce target domain samples: Landslide events in coal mining areas are relatively isolated, making it difficult to obtain a large number of pixel-level labeled samples. Directly using a small number of samples to train a deep model is very likely to lead to overfitting and poor generalization ability.

[0004] (2) High risk of negative transfer: The data distribution (source domain) of the pre-trained model for general natural scenarios or general mountain landslides differs significantly from that of the coal mining area (target domain). If parameter fine-tuning is forcibly performed, the source domain knowledge will not only be useless, but may also damage the performance of the target domain, causing the "negative transfer" phenomenon. The theoretical research of Wu et al. (Online Transfer Learning: Negative Transfer and Effect of Prior Knowledge, 2021) points out that inappropriate prior knowledge assumptions are the root cause of negative transfer.

[0005] (3) Prior knowledge has not been effectively converted into training signals: The occurrence of landslides is strictly controlled by geomechanical factors such as stratum lithology, slope structure, and disturbance from mining operations. Existing methods usually use this structured knowledge as an additional input layer (such as Lei et al. using order tracking for data augmentation in fault diagnosis, or Wu et al. extracting semantic edge maps from dental images), but fail to deeply integrate it into the training objectives and learning process of the model. Although the prior knowledge embedding meta-transfer learning proposed by Lei et al. (PKEMTL, Mechanical Systems and Signal Processing, 2023) has achieved results in fault diagnosis, it is geared towards temporal vibration signals and cannot be directly applied to remote sensing images and spatialized geological prior knowledge. The enhancement strategy of mapping geomechanical parameters to remote sensing images needs to solve remote sensing-specific problems such as coordinate system registration, spatial resolution matching, and multi-source data fusion, and is not a simple scene replacement.

[0006] (4) Confusion between mining-specific disturbances and natural landslides: Coal mining areas have a large number of surface deformations caused by mining activities such as goaf collapse and surface subsidence, and their image characteristics are very similar to those of real landslides. Models lacking geological prior guidance are very likely to misjudge them, resulting in a large number of false positives.

[0007] Furthermore, existing methods typically use geological information as input features and directly splice them together, lacking a technical path to transform it into training constraints or learning mechanisms. This results in the model lacking the ability to distinguish between mining area disturbances and actual landslides.

[0008] Therefore, there is an urgent need for a landslide identification method that can effectively utilize prior knowledge and suppress the risk of negative migration under small sample conditions. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a deep learning landslide identification method and system that integrates prior geological and mining knowledge. By constructing a prior-driven enhancement mechanism and a multi-constraint transfer learning framework, the identification accuracy and generalization ability are improved under the condition of scarce samples.

[0010] In a first aspect, embodiments of the present invention provide a deep learning-based landslide identification method that integrates prior geological and mining knowledge, including:

[0011] Construct a prior knowledge graph of geology and mining for the target coal mining area, and generate a prior feature graph based on the prior knowledge graph;

[0012] The remote sensing image to be identified is acquired, the prior feature map is spatially registered with the remote sensing image, the center point of the landslide candidate area in the remote sensing image is extracted using an independently trained lightweight network, and the landslide candidate area is adaptively cropped.

[0013] The landslide candidate regions are input into the segmentation model trained by target domain meta-transfer learning to obtain the landslide probability map of each candidate region;

[0014] The landslide probability maps under multiple prior assumptions are adaptively fused using a channel attention mechanism to generate a final landslide identification mask, and the landslide disaster identification result is output based on the landslide identification mask.

[0015] As a preferred implementation, a prior knowledge graph of the geology and mining of the target coal mining area is constructed, and a prior feature map is generated based on the prior knowledge graph, including:

[0016] Obtain the mining engineering plan, geological map, and digital elevation model of the target coal mining area;

[0017] Extract the boundary of the goaf and its distance from the mining face from the mining engineering plan to generate a mining prior map;

[0018] The stratigraphic lithological boundaries and distances from faults are extracted from the geological map, and a lithological softening coefficient map is generated based on lithological mechanical parameters as a geological prior map.

[0019] Slope and aspect maps are calculated from the digital elevation model to generate a topographic prior map;

[0020] The mining prior map, geological prior map, and topographic prior map are spliced ​​together to generate a multi-channel prior feature map.

[0021] In a preferred embodiment, the prior feature map is spatially registered with the remote sensing image, the center points of landslide candidate areas in the remote sensing image are extracted using an independently trained lightweight network, and the landslide candidate regions are adaptively cropped, including:

[0022] The prior feature map is spatially registered with the remote sensing image, and the registered prior feature map is input into an independently trained lightweight U-Net network to predict the heat map of the geometric center point of the landslide body.

[0023] Connectivity analysis is performed on the heat map, and center points are filtered based on a preset confidence threshold to locate candidate center points for each landslide.

[0024] The cropping size is adaptively determined based on the slope information of the location of each landslide candidate center point, and candidate image blocks are cropped from the remote sensing image with the center point as the center.

[0025] An overlapping area is set between adjacent candidate image blocks, and the pixel values ​​in the overlapping area are fused by distance-weighted averaging during stitching to obtain landslide candidate areas.

[0026] In a preferred embodiment, the landslide candidate regions are input into a segmentation model trained by target domain meta-transfer learning to obtain a landslide probability map for each candidate region, including:

[0027] For each landslide candidate area, multiple sets of different spatial adaptive enhancement parameters are used to enhance its image, generating multiple enhanced images;

[0028] Each enhanced image is input into the segmentation model to obtain a corresponding landslide probability map; wherein, multiple landslide probability maps corresponding to each landslide candidate region are used for subsequent channel attention fusion.

[0029] As a preferred implementation, the landslide probability maps under multiple prior assumptions are adaptively fused using a channel attention mechanism to generate a final landslide identification mask, including:

[0030] Multiple landslide probability maps corresponding to each landslide candidate area are stitched together along the channel dimension to obtain a multi-channel probability map.

[0031] The multi-channel probability map is input to the channel attention fusion module. The fusion module includes a first fully connected layer, an activation function layer, and a second fully connected layer connected in sequence. The number of input channels of the first fully connected layer is the same as the number of the multiple landslide probability maps, and the number of output channels of the second fully connected layer is the same as the number of input channels of the first fully connected layer. The fusion module outputs a weight vector, and the dimension of the weight vector is the same as the number of the multiple landslide probability maps.

[0032] The weight vector is weighted and summed with the multiple landslide probability maps channel by channel to obtain the fused single-channel probability map.

[0033] The single-channel probability map is thresholded and binarized to generate the final landslide identification mask.

[0034] In a preferred embodiment, the segmentation model trained by target domain meta-transfer learning is obtained in the following way:

[0035] Based on the prior knowledge graph, spatial adaptive enhancement and geometric micro-deformation with slope aspect constraints are performed on a small number of real landslide samples to generate enhanced samples and corresponding enhancement masks. Enhanced remote sensing image meta-tasks are constructed based on the enhanced samples, wherein each meta-task includes a support set and a query set.

[0036] The basic segmentation network is pre-trained on the source domain general landslide sample database to obtain the pre-trained segmentation network;

[0037] Meta-transfer learning is performed on the pre-trained segmentation network on the enhanced remote sensing image meta-task. The total loss function of the meta-training includes perceptual consistency loss and domain confusion loss. The perceptual consistency loss is used to constrain the distribution consistency between the enhanced samples and the original samples in the feature space. The domain confusion loss is used to align the feature distributions of the source domain and the target domain to suppress negative transfer.

[0038] In a preferred embodiment, the perceived consistency loss is calculated according to the following steps:

[0039] In the construction of the meta-task, a query set consisting of augmented samples is obtained, wherein each augmented sample in the query set corresponds to an original sample;

[0040] Extract the feature maps of the enhanced sample and its corresponding original sample from the output of the last residual block of the encoder;

[0041] After the feature map is activated by ReLU and a smoothing term is added, the channel feature vector at each spatial location is normalized to obtain a normalized feature distribution.

[0042] Calculate the Jensen-Shannon divergence between the normalized feature distributions of the enhanced sample and the original sample at the same spatial location, and use it as the distribution difference value at that spatial location;

[0043] The perceptual consistency loss is obtained by averaging the distribution differences across all spatial locations and all sample pairs.

[0044] In a preferred embodiment, the geometric micro-deformation includes the following constraints:

[0045] Obtain the original image and its corresponding labeled mask;

[0046] Regularly distributed control points are set on the original image. The main offset direction of each control point is determined to be the slope direction of its location, and the displacement component along the slope direction dominates the total displacement. The upper limit of the offset amplitude of the control point is determined according to the slope of its location.

[0047] The offset of each control point is generated based on the offset direction and upper bound of the offset amplitude, and a deformation field is generated based on the offset using thin plate spline transformation.

[0048] Calculate the Jacobian determinant of the deformation field. If there is a negative value region, resample the control point offset and regenerate the deformation field based on the resampled offset until the Jacobian determinant has no negative value region.

[0049] The deformation field is applied to the original image and the labeled mask respectively to obtain the deformed enhanced sample and its corresponding enhancement mask.

[0050] As a preferred embodiment, the spatial adaptive enhancement further includes:

[0051] Calculate the relative offset between the mean gray value of each band of the enhanced sample and the mean gray value of the corresponding band of the original sample;

[0052] Calculate the relative standard deviation of the grayscale standard deviation of each band of the enhanced sample and the corresponding grayscale standard deviation of the original sample;

[0053] If the relative offset of the mean exceeds a preset first threshold or the relative offset of the standard deviation exceeds a preset second threshold, the enhancement intensity of the spatial adaptive enhancement is reduced, and the enhanced sample is regenerated based on the reduced enhancement intensity.

[0054] Secondly, embodiments of the present invention also provide a deep learning landslide identification model construction and transfer learning system that integrates prior knowledge, comprising:

[0055] The prior knowledge graph construction module is used to construct a prior knowledge graph of the geology and mining of the target coal mining area, and generate a prior feature graph based on the prior knowledge graph.

[0056] The candidate region extraction module is used to acquire the remote sensing image to be identified, spatially register the prior feature map with the remote sensing image, extract the center point of the landslide candidate area in the remote sensing image using an independently trained lightweight network, and adaptively crop to obtain the landslide candidate area.

[0057] The segmentation prediction module is used to input the landslide candidate regions into the segmentation model trained by target domain meta-transfer learning to obtain the landslide probability map of each candidate region.

[0058] The fusion output module is used to adaptively fuse the landslide probability maps under multiple prior assumptions through a channel attention mechanism, generate a final landslide identification mask, and output the landslide disaster identification result based on the landslide identification mask.

[0059] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0060] One or more processors;

[0061] Storage device for storing one or more programs;

[0062] When the one or more programs are executed by the one or more processors, the one or more processors implement a deep learning landslide identification method that integrates geological and mining prior knowledge as described in any embodiment of the present invention.

[0063] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a deep learning landslide identification method that integrates prior geological and mining knowledge as described in any embodiment of the present invention.

[0064] Compared to existing technologies, the following beneficial effects have been achieved:

[0065] This invention deeply integrates prior knowledge of geology and mining to construct a complete landslide identification framework from data augmentation to adaptive fusion. Through prior-driven spatial adaptive augmentation and aspect-constrained geometric micro-deformation, a high-precision segmentation model can be obtained with only a small number of real samples, effectively alleviating the overfitting problem caused by the scarcity of samples in coal mining areas. At the same time, by jointly optimizing perceptual consistency loss and domain confusion loss, the consistency between augmented samples and original samples is constrained at the feature level, and the source-target domain distribution is aligned, avoiding negative transfer of general knowledge and achieving safe and robust cross-domain landslide identification. Attached Figure Description

[0066] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0067] Figure 1 This is a flowchart of a deep learning landslide identification method that integrates prior geological and mining knowledge, provided by an embodiment of the present invention.

[0068] Figure 2 This is a diagram of the meta-transfer learning training structure provided in an embodiment of the present invention;

[0069] Figure 3 This is a sample efficiency curve provided in an embodiment of the present invention;

[0070] Figure 4 This is a schematic diagram of the structure of the deep learning landslide identification method that integrates prior geological and mining knowledge provided in an embodiment of the present invention;

[0071] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0073] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations (or steps) may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0074] Example 1

[0075] like Figure 1As shown, Embodiment 1 of the present invention provides a deep learning landslide identification method 100 that integrates prior geological and mining knowledge. Addressing the problems of scarce landslide samples, high negative migration risk, and easy misjudgment of disturbances in coal mining areas, this method achieves high-precision and robust intelligent landslide identification with small sample sizes through a four-step process: prior knowledge graph construction, adaptive candidate region extraction, meta-transfer learning segmentation, and channel attention fusion. The method 100 specifically includes the following steps:

[0076] Step S110: Construct a prior knowledge graph of geology and mining in the target coal mining area, and generate a prior feature graph based on the prior knowledge graph.

[0077] In a preferred embodiment, this step transforms the static geological, mining, and topographic data of the coal mining area into spatial prior features that can be used by the deep learning model, providing geomechanical constraints for subsequent sample enhancement, candidate area extraction, and model training, thus solving the problems of scarce target domain samples and insufficient utilization of prior knowledge from the source.

[0078] Specifically, in this embodiment, a 1:10000 mining engineering plan, geological map, and high-precision digital elevation model (DEM) of the target coal mining area are collected. All collected vector data are rasterized to make them completely consistent with the spatial resolution of the remote sensing image, and the following factors are extracted:

[0079] Mining Priority Map This includes a binary map of the goaf boundary and a continuous map of the distance from the mining face, used to characterize the spatial distribution of mining disturbance intensity;

[0080] Geological a priori map This includes a binary map of stratigraphic lithological boundaries and a continuous map of distances from faults, used to characterize differences in the anti-sliding capacity of strata;

[0081] Topographical Prior Map Slope and aspect maps calculated based on DEM are used to characterize the control effect of topography on landslide occurrence.

[0082] Next, the aforementioned mining prior art map Geological preliminary map and topographic prior maps Channel splicing is performed to generate multi-channel prior feature maps, realizing the spatial fusion of geological, mining, and topographical priors, which serve as a unified prior input for subsequent modules.

[0083] Step S120: Obtain the remote sensing image to be identified, spatially register the prior feature map with the remote sensing image, extract the center point of the landslide candidate area in the remote sensing image using an independently trained lightweight network, and adaptively crop to obtain the landslide candidate area.

[0084] As a preferred embodiment, to reduce the intensive computational cost on large-scale remote sensing images, this invention designs a pre-candidate region extraction module independent of meta-transfer learning. The multi-channel prior feature map generated in step S110 above quickly locates the center point of potential landslides and adaptively crops candidate image blocks, specifically including the following steps:

[0085] First, the multi-channel prior feature map generated in S110 is strictly spatially registered with the remote sensing image to be identified to ensure that the prior information corresponds one-to-one with the image pixels.

[0086] Next, the key points of the landslide are located. Specifically, the registered prior feature map is input into an independently trained lightweight U-Net network to predict the heat map of the geometric center point of each landslide body, which represents the probability center of the landslide occurrence.

[0087] The lightweight U-Net network is independent of the main meta-transfer learning process. It is trained using a subset of the target domain, evaluated using leave-one-out cross-validation, and its weights are fixed after 30 epochs of training.

[0088] Next, a connected component analysis was performed on the heatmap of the geometric center point of each landslide body. Based on the confidence threshold determined by the validation set, high confidence center points were selected. At the same time, combined with the spatial constraints of the overlapping areas of high slope and mining subsidence, candidate points were added to low confidence but high risk areas to ensure that the recall rate of candidate areas reached 92%~96%.

[0089] Finally, using the selected center point as a reference, the cropping size is adaptively determined according to the slope of the location to avoid landslide omissions or redundant calculations caused by fixed sizes. At the same time, a 32-pixel overlap area is set between adjacent candidate image blocks. During stitching, the pixels in the overlap area are fused by distance-weighted average to eliminate stitching gaps and obtain a complete landslide candidate area.

[0090] Step S130: Input the landslide candidate regions into the segmentation model trained by target domain meta-transfer learning to obtain the landslide probability map of each candidate region.

[0091] In a preferred embodiment, this step embeds prior knowledge into sample augmentation and model training, expands the small-sample training space through prior-driven augmentation meta-tasks, and suppresses cross-domain negative transfer by combining meta-transfer learning to output a high-precision landslide probability map. Specifically, it includes the following steps:

[0092] Step 1301: Prior-driven generation of enhanced meta-tasks:

[0093] This invention proposes a conditional enhancement strategy control method. Its core is to map prior parameters such as distance from the mining face and lithological softening coefficient to image enhancement control parameters, using pixels as the unit. This enhancement process does not simulate the actual remote sensing imaging mechanism, but rather statistically expands the sample distribution, enabling the model to learn robust characterization capabilities against potential geological disturbances. Specifically, it includes:

[0094] (1) Spatial adaptive enhancement:

[0095] Prior parameters such as distance from the mining face and lithological softening coefficient are mapped to image enhancement control parameters. Specifically, for the image of any real landslide sample in the target domain... Its enhanced image Generated by the following formula:

[0096] ;

[0097] in, These are pixel coordinates; Original image at coordinates Pixel value at; Enhanced image in coordinates Pixel value at; : Spatial adaptive enhancement function; The set of enhancement parameters obtained by mapping from the prior image is used to control the amplitude of local brightness perturbation, texture noise intensity, and contrast variation range.

[0098] For example:

[0099] ;

[0100] in, : Band number of the remote sensing image Gaussian noise sampled independently for each band. The noise standard deviation is defined as having an upper bound of 0.1 times the mean of the grayscale standard deviations of each band in the original image; the sampling intensity coefficient. and lithological influence coefficient The range of values ​​for is determined by statistical constraints, preferably . , ; Normalized distance map from the mining face The value at; The encoded lithological softening coefficient map is in The value at; Original image band in Pixel value at; Enhanced image band in The pixel value at that location.

[0101] (2) Next, after spatial adaptive enhancement, clipping is performed on each band separately, as shown in the following formula:

[0102] ;

[0103] in and The first The pixel grayscale mean and standard deviation of the original image of the band; after enhancement, verify the overall statistics of the band: if the mean shift exceeds 20% of the original sample mean or the standard deviation changes by more than 20%, then adjust the coefficients. Scale (multiply by 0.8) and enhance again, retrying up to 3 times; if still not satisfied, fall back to the previous state. , The minimum perturbation setting is used to ensure the consistency of statistical characteristics between the augmented data and the real remote sensing observation data; Cutting factor, set to 2.5; : Pixel value cropping operation, which restricts the value to a specified range.

[0104] (3) Geometric micro-deformation guided by landslide mechanism:

[0105] Four × four regularly distributed control points are set on a 256×256 pixel image. The offset direction of the control points is constrained by the slope aspect, with the main offset component distributed along the slope aspect and superimposed with small random disturbance components; the displacement along the slope aspect accounts for no less than 70% of the total displacement. Upper bound of the offset amplitude. Based on the slope adaptively determined at the control point, preferably... No more than 8 pixels.

[0106] The deformation field is generated using Thin Plate Spline (TPS) transformation, and the absence of negative value regions is confirmed by Jacobian determinant test; if a negative value region is detected, the control point offset is resampled. This deformation field is applied synchronously to the image and its labeled mask. The supervision mask used during the training and inference of the fusion module is the enhanced mask after synchronous deformation, rather than the original undeformed mask.

[0107] (4) Meta-task construction:

[0108] Five samples are randomly selected from the training subset of the target mining area, and one augmentation sample is provided for each sample to form the support set. Select another different enhanced variant to form the query set. This forms a 5-way 1-shot meta-task. Additionally, the original unenhanced samples corresponding to these 5 instances are passed in. Strictly ensure that the same enhanced variant does not appear simultaneously in the same landslide body. and In the middle. The same enhanced variants are allowed to be reused between different meta-tasks, but the parameters are modified at build time. Perform stratified sampling to ensure distributed coverage. Repeat this process to construct no fewer than 200 meta-tasks, preferably 200–500.

[0109] Step 1302 Meta-transfer learning model training

[0110] like Figure 2 The diagram shown is a meta-transfer learning training structure diagram provided in an embodiment of the present invention. It employs a unified model architecture with a shared encoder and dual-task heads (source domain head and target domain head) to achieve secure transfer of general knowledge from the source domain to the target domain. Specifically:

[0111] (1) Model framework description: Shared encoder: ResNet+ASP module, extracts general landslide features; Source domain head: complete DeepLabV3 + decoder, used for general landslide sample pre-training; Target domain head: Prototype matching module, used for small sample training.

[0112] (2) Source domain pre-training: On a large general landslide sample library, the encoder and source domain head are pre-trained using conventional cross-entropy and Dice loss to learn general landslide morphological features and obtain general landslide morphology recognition capabilities.

[0113] Target domain meta-training: A prototype-based segmentation meta-learning framework is adopted. During the target domain meta-training stage, the encoder is retained for feature extraction, and the feature map output by the ASPP module is used as the input for prototype calculation and cosine similarity. The original DeepLabV3+ decoder head is replaced with the prototype matching module. The shallow layers of the encoder are frozen, and gradient updates are performed only on the deep layers and ASPP parameters.

[0114] Preferably, the meta-training phase employs an alternating sampling strategy: in each step, the source domain batch recovery is computed using the full decoder output. (Source domain part); Simultaneously sample one meta-task and the corresponding 5 original unenhanced samples, and jointly optimize the composite loss.

[0115] (3) The composite loss function is as follows:

[0116] ;

[0117] in, For conventional segmentation loss, the source domain uses the full decoder output to calculate cross-entropy and Dice loss; the target domain query set prediction is a probabilistic graph, so binary cross-entropy loss and Dice loss are used directly. Perceptual consistency loss is applied to the feature map output by the last residual block of the encoder. It constrains the distribution differences between augmented and original samples in the feature space, ensuring the model maintains consistency with surface perturbations. Calculation method: For each augmented sample in the query set... and its corresponding original samples After ReLU activation of their respective encoder feature maps, smoothing terms are added along the channel dimension. ,Right now Then, L1 normalization is performed along the channel dimension. This normalization is performed pixel-by-pixel along the channel dimension, meaning that the feature vector at each spatial location is normalized to obtain a normalized distribution. The Jensen-Shannon divergence (JS divergence) was used to calculate the difference between the two distributions. The mean was taken for all pairings and all spatial locations.

[0118] ;

[0119] in, Domain confusion loss, applied to the feature vector after global average pooling of the encoder, is used to align the feature distributions of the source and target domains. It suppresses negative transfer in an adversarial manner through a domain discriminator and GRL. The discriminator structure is as follows: LeakyReLU activation, binary cross-entropy loss, GRL inversion coefficient increases from 0 to 1 as training progresses; hyperparameters The optimal value is determined by a grid search on the validation subset. The specific details can be adjusted based on the mining area data; The number of valid pairs in the current batch; The original feature map output by the encoder; Feature map after ReLU activation and the addition of a smoothing term; : Smoothing term, with values ​​of This is used to avoid division by zero errors; : Feature distribution after L1 normalization pixel by pixel along the channel dimension; : Height of the encoder output feature map; : Width of the encoder output feature map; Spatial location index (row, column) of the feature map; Enhance the spatial location of samples Normalized characteristic distribution at; The original sample in spatial location Normalized characteristic distribution at; Jensen-Shannon divergence; The number of feature channels in the final output of the encoder.

[0120] Step S140: Adaptively fuse the landslide probability maps under multiple prior assumptions through a channel attention mechanism to generate the final landslide identification mask, and output the landslide disaster identification result based on the landslide identification mask.

[0121] As a preferred embodiment, to address the prediction bias of a single enhancement hypothesis, this step further improves recognition accuracy and robustness through adaptive fusion of multiple prior hypotheses and channel attention, outputting the final landslide recognition mask, specifically including the following steps:

[0122] (1) Multiple prior hypothesis enhancement: During inference, M different prior weight coefficients are applied to the test image block. Preferably, M=3~7, and in this example, M=5 is used to generate M enhanced image variants, corresponding to M kinds of geomechanical prior assumptions.

[0123] (2) Input the M enhanced image variants into the meta-transfer learning model to obtain A single-channel landslide probability map.

[0124] (3) The M single-channel landslide probability maps are spliced ​​together and input into the lightweight channel attention fusion module. The module outputs a weight vector that matches the dimensions of the M probability maps. The weighted summation of each probability map is performed channel by channel to obtain the fused single-channel probability map. The fused probability map is then thresholded and binarized to generate the final landslide recognition mask.

[0125] Based on the landslide recognition mask, the system outputs complete recognition results, including landslide location, extent, and boundaries, thus completing the entire process of intelligent landslide recognition in coal mining areas.

[0126] Based on the above embodiments, the core beneficial effects of the present invention are as follows:

[0127] (1) This invention constructs a complete landslide identification framework from data augmentation and candidate area extraction to meta-transfer learning and adaptive fusion by deeply integrating prior knowledge of geology and mining. It utilizes prior-driven spatial adaptive augmentation and slope aspect constraint geometric micro-deformation to generate augmented meta-tasks. Combined with the meta-learning framework, a high-precision segmentation model can be obtained with only a small number of real samples, which effectively alleviates the overfitting problem caused by the scarcity of labeled samples in scenarios such as coal mining areas.

[0128] (2) This invention optimizes the consistency between the sample and the original sample by combining the perceptual consistency loss and the domain confusion loss at the feature level, and aligns the data distribution between the source domain and the target domain, thereby avoiding the damage of the general knowledge of the source domain to the performance of the target domain and realizing safe knowledge transfer.

[0129] (3) Based on the geometric deformation field controlled by slope aspect, the brightness disturbance guided by the lithological softening coefficient and the enhancement strategy of statistical consistency constraint, the model can learn the characteristics related to the essential mechanical mechanism of landslides and reduce the misjudgment of surface deformation caused by mining activities such as goaf collapse.

[0130] Example 2

[0131] This embodiment uses a coal mining area in Shanxi Province as a practical application scenario to fully illustrate the entire process of data preparation, prior art creation, model training, and performance evaluation of the method of the present invention, and to verify the technical advantages of the present invention in small sample, cross-domain, and complex mining area scenarios.

[0132] 1. Experimental data preparation and sample set division

[0133] We collected Gaofen-2 fused images (2m resolution) and a 5m resolution digital elevation model (DEM) of a coal mining area in Shanxi Province, along with mining engineering plans and geological maps, covering the entire area of ​​historical landslides and goaf areas in the study area.

[0134] Historical landslide samples from the study area were divided into training and validation subsets in a 12:3 ratio. Ten newly occurring landslide scenarios, independent of time, were selected as the final test set to ensure no data leakage. Publicly available landslide datasets (such as Bijie, HR-GLDD, and CAS datasets) can be used for source domain pre-training.

[0135] 2. Prior layer and enhancement meta-task creation

[0136] Mining, geological, and topographic factors were uniformly normalized and resampled to a 2m resolution to maintain consistency with remote sensing imagery; mining prior maps, geological prior maps, and topographic prior maps were generated and channel-stitched to obtain a three-channel prior feature map; among them, the lithological softening coefficient map was generated by normalizing the mechanical parameters of the geological exploration report.

[0137] Landslide samples were trained for 12 target domains, and parameters were enhanced through random sampling. The algorithm utilizes spatial adaptive augmentation (including statistical constraints and pruning) and aspect-constrained TPS deformation to generate approximately 240 augmented samples and corresponding augmentation masks. It constructs no fewer than 200 5-way 1-shot meta-tasks, with hierarchical sampling parameter spaces.

[0138] 3. Model Training

[0139] This embodiment is executed in four stages: candidate region extraction, source domain pre-training, target domain meta-training, and fusion module training. The parameters and process are as follows:

[0140] Candidate Region U-Net: A lightweight U-Net is trained independently and used only for landslide center point localization. Training is performed using a subset of training data, and cross-validation is performed with leave-one-out method. The confidence threshold is determined by the recall-false positive rate curve of the validation set to ensure high recall of candidate regions.

[0141] Source domain pre-training: Based on DeepLabV3+, the segmentation network is pre-trained for 100 epochs on a publicly available general landslide dataset to learn the general morphological and textural features of landslides in natural scenes.

[0142] Target domain meta-training: Switch to prototype matching head, freeze shallow encoder parameters, and only update deep layer and ASPP modules; adopt a source domain batch and target domain meta-task alternating sampling strategy, set the learning rate to 1e-4, and train for 50 epochs; jointly optimize segmentation loss, perceptual consistency loss and domain confusion loss to achieve safe cross-domain transfer.

[0143] Fusion module: With fixed meta-learning model parameters, it independently trains for 20 epochs using augmented samples and corresponding augmented masks to ensure the consistency and stability of multi-hypothesis prediction fusion during the inference stage.

[0144] 4. Performance Evaluation

[0145] Ten novel landslide scenarios were used as test subjects, with mIoU, precision, recall, and F1 score as the core metrics. Each experiment was repeated five times and the average was taken to reduce random error. The specific numerical trends are reflected in the interval description of the beneficial effects. The parameter ranges can be adjusted according to the actual data of different mining areas.

[0146] like Figure 3 As shown, the horizontal axis represents the number of source domain training samples used by the Fine-tuning method, and the vertical axis represents mIoU (%). The red dashed line represents the performance level of the method of this invention with a fixed number of 12 target domain landslide samples (61.3%), and the blue solid line represents the performance trend of the Fine-tuning baseline as the number of source domain samples increases.

[0147] This invention is significantly superior to traditional methods in three key indicators: small sample size, cross-domain migration, and anti-interference in mining areas. It can effectively distinguish between landslides and pseudo-lands in mining areas such as mining subsidence and surface subsidence.

[0148] Example 3

[0149] Figure 4 This is a schematic diagram of the structure of a deep learning landslide identification system 400 that integrates prior geological and mining knowledge, as provided in Embodiment 3 of the present invention. Figure 4 As shown, the system includes:

[0150] The prior knowledge graph construction module 410 is used to construct a prior knowledge graph of the geology and mining of the target coal mining area, and generate a prior feature graph based on the prior knowledge graph.

[0151] The candidate region extraction module 420 is used to acquire the remote sensing image to be identified, spatially register the prior feature map with the remote sensing image, extract the center point of the landslide candidate area in the remote sensing image using an independently trained lightweight network, and adaptively crop to obtain the landslide candidate area.

[0152] The segmentation prediction module 430 is used to input the landslide candidate regions into the segmentation model trained by target domain meta-transfer learning to obtain the landslide probability map of each candidate region.

[0153] The fusion output module 440 is used to adaptively fuse the landslide probability maps under multiple prior assumptions through a channel attention mechanism, generate a final landslide identification mask, and output the landslide disaster identification result based on the landslide identification mask.

[0154] The deep learning landslide identification system that integrates geological and mining prior knowledge provided in this embodiment of the invention can execute the deep learning landslide identification method that integrates geological and mining prior knowledge provided in any of the above embodiments of the invention. It has the corresponding functions and beneficial effects of executing the deep learning landslide identification method that integrates geological and mining prior knowledge. For detailed process, please refer to the relevant operations of the deep learning landslide identification method that integrates geological and mining prior knowledge in the foregoing embodiments.

[0155] Example 4

[0156] Figure 5 This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile 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 invention described and / or claimed herein.

[0157] like Figure 5As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0158] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0159] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes a deep learning landslide identification method that integrates prior geological and mining knowledge as described above.

[0160] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0161] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A deep learning-based landslide identification method integrating prior geological and mining knowledge, characterized in that, include: Construct a prior knowledge graph of geology and mining for the target coal mining area, and generate a prior feature graph based on the prior knowledge graph; The process involves acquiring a remote sensing image to be identified, spatially registering the prior feature map with the remote sensing image, extracting the center points of landslide candidate areas from the remote sensing image using an independently trained lightweight network, and adaptively cropping to obtain landslide candidate regions. This includes: spatially registering the prior feature map with the remote sensing image, inputting the registered prior feature map into an independently trained lightweight U-Net network to predict a heatmap of the geometric center points of the landslide body; performing connected component analysis on the heatmap and filtering center points based on a preset confidence threshold to locate each landslide candidate center point; adaptively determining the cropping size based on the slope information of each landslide candidate center point's location, and cropping candidate image blocks from the remote sensing image centered on the center point; setting an overlap area between adjacent candidate image blocks, and fusing the pixel values ​​within the overlap area using a distance-weighted average during stitching to obtain the landslide candidate regions. The landslide candidate regions are input into a segmentation model trained by target domain meta-transfer learning to obtain landslide probability maps for each candidate region, including: For each landslide candidate region, multiple sets of different spatial adaptive enhancement parameters are used to enhance its image, generating multiple enhanced images; each enhanced image is input into the segmentation model to obtain the corresponding landslide probability map; wherein, the multiple landslide probability maps corresponding to each landslide candidate region are used for subsequent channel attention fusion. The landslide probability maps under multiple prior assumptions are adaptively fused using a channel attention mechanism to generate a final landslide identification mask, and the landslide disaster identification result is output based on the landslide identification mask. The segmentation model trained by target domain meta-transfer learning is obtained in the following way: Based on the prior knowledge graph, spatial adaptive enhancement and geometric micro-deformation with slope aspect constraints are performed on a small number of real landslide samples to generate enhanced samples and corresponding enhancement masks. Enhanced remote sensing image meta-tasks are constructed based on the enhanced samples, wherein each meta-task includes a support set and a query set. A basic segmentation network is pre-trained on a general landslide sample library in the source domain to obtain a pre-trained segmentation network. Meta-transfer learning is then performed on the pre-trained segmentation network on the augmented remote sensing image meta-task. The total loss function of the meta-training includes perceptual consistency loss and domain confusion loss. The perceptual consistency loss is used to constrain the distribution consistency between the augmented samples and the original samples in the feature space. The domain confusion loss is used to align the feature distributions of the source domain and the target domain to suppress negative transfer.

2. The method according to claim 1, characterized in that, Constructing a prior knowledge graph of geology and mining for the target coal mining area, and generating a prior feature graph based on the prior knowledge graph, including: Obtain the mining engineering plan, geological map, and digital elevation model of the target coal mining area; Extract the boundary of the goaf and its distance from the mining face from the mining engineering plan to generate a mining prior map; The stratigraphic lithological boundaries and distances from faults are extracted from the geological map, and a lithological softening coefficient map is generated based on lithological mechanical parameters as a geological prior map. Slope and aspect maps are calculated from the digital elevation model to generate a topographic prior map; The mining prior map, geological prior map, and topographic prior map are spliced ​​together to generate a multi-channel prior feature map.

3. The method according to claim 1, characterized in that, The landslide probability maps under multiple prior assumptions are adaptively fused using a channel attention mechanism to generate a final landslide identification mask, including: Multiple landslide probability maps corresponding to each landslide candidate area are stitched together along the channel dimension to obtain a multi-channel probability map. The multi-channel probability map is input to the channel attention fusion module. The fusion module includes a first fully connected layer, an activation function layer, and a second fully connected layer connected in sequence. The number of input channels of the first fully connected layer is the same as the number of the multiple landslide probability maps, and the number of output channels of the second fully connected layer is the same as the number of input channels of the first fully connected layer. The fusion module outputs a weight vector, and the dimension of the weight vector is the same as the number of the multiple landslide probability maps. The weight vector is weighted and summed with the multiple landslide probability maps channel by channel to obtain the fused single-channel probability map. The single-channel probability map is thresholded and binarized to generate the final landslide identification mask.

4. The method according to claim 1, characterized in that, The perceived consistency loss is calculated according to the following steps: In the construction of the meta-task, a query set consisting of augmented samples is obtained, wherein each augmented sample in the query set corresponds to an original sample; Extract the feature maps of the enhanced sample and its corresponding original sample from the output of the last residual block of the encoder; After the feature map is activated by ReLU and a smoothing term is added, the channel feature vector at each spatial location is normalized to obtain a normalized feature distribution. Calculate the Jensen-Shannon divergence between the normalized feature distributions of the enhanced sample and the original sample at the same spatial location, and use it as the distribution difference value at that spatial location; The perceptual consistency loss is obtained by averaging the distribution differences across all spatial locations and all sample pairs.

5. The method according to claim 1, characterized in that, The geometric micro-deformation includes the following constraints: Obtain the original image and its corresponding labeled mask; Regularly distributed control points are set on the original image. The main offset direction of each control point is determined to be the slope direction of its location, and the displacement component along the slope direction dominates the total displacement. The upper limit of the offset amplitude of the control point is determined according to the slope of its location. The offset of each control point is generated based on the offset direction and upper bound of the offset amplitude, and a deformation field is generated based on the offset using thin plate spline transformation. Calculate the Jacobian determinant of the deformation field. If there is a negative value region, resample the control point offset and regenerate the deformation field based on the resampled offset until the Jacobian determinant has no negative value region. The deformation field is applied to the original image and the labeled mask respectively to obtain the deformed enhanced sample and its corresponding enhancement mask.

6. The method according to claim 1, characterized in that, The spatial adaptive enhancement also includes: Calculate the relative offset between the mean gray value of each band of the enhanced sample and the mean gray value of the corresponding band of the original sample; Calculate the relative standard deviation of the grayscale standard deviation of each band of the enhanced sample and the corresponding grayscale standard deviation of the original sample; If the relative offset of the mean exceeds a preset first threshold or the relative offset of the standard deviation exceeds a preset second threshold, the enhancement intensity of the spatial adaptive enhancement is reduced, and the enhanced sample is regenerated based on the reduced enhancement intensity.

7. A deep learning landslide identification system integrating prior geological and mining knowledge, characterized in that, The system for performing the method according to any one of claims 1 to 6 comprises: The prior knowledge graph construction module is used to construct a prior knowledge graph of the geology and mining of the target coal mining area, and generate a prior feature graph based on the prior knowledge graph. The candidate region extraction module is used to acquire the remote sensing image to be identified, spatially register the prior feature map with the remote sensing image, extract the center point of the landslide candidate area in the remote sensing image using an independently trained lightweight network, and adaptively crop to obtain the landslide candidate area. The segmentation prediction module is used to input the landslide candidate regions into the segmentation model trained by target domain meta-transfer learning to obtain the landslide probability map of each candidate region. The fusion output module is used to adaptively fuse the landslide probability maps under multiple prior assumptions through a channel attention mechanism, generate a final landslide identification mask, and output the landslide disaster identification result based on the landslide identification mask.

Images