Geological disaster change detection method and system based on improved twin u-net and center-surround dual-stream network

Abstract

The present application relates to a kind of geological disaster change detection method and system based on improved twin U-Net with center surround dual-flow network.The method constructs improved twin U-Net network as feature extraction main stem, integrates deformable convolution module in encoder to adaptively adjust receptive field;Center surround dual-flow network is embedded in decoding path, detail features are extracted by center flow path, and global context is obtained by surround flow path;Design feature fusion module to realize the deep fusion of double-path features, adopt gated attention unit to calibrate feature response;Introduce contrast feature learning mechanism to strengthen feature space clustering characteristics;Multi-modal degradation enhancement training is implemented to improve model robustness;Through end-to-end joint optimization, the objective function including difference perception loss and feature contrast loss is minimized.The present application can accurately identify landslide, debris flow and ground subsidence and other geological disaster change regions, with the advantages of high detection precision and strong anti-interference ability.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and remote sensing image processing technology, and in particular to a method and system for detecting geological disaster changes based on an improved twin U-Net and a central-surround dual-stream network. Background Technology

[0002] Geological hazard change detection is an important interdisciplinary field between remote sensing technology and geological hazard prevention and control, and it is of key significance for the identification, monitoring, and early warning of hazards such as landslides, debris flows, and ground subsidence. Traditional change detection methods mainly rely on pixel-level comparison or shallow feature analysis, which have problems such as poor adaptability to complex terrain, insufficient ability to preserve details, and susceptibility to environmental interference.

[0003] With the development of deep learning technology, change detection methods based on convolutional neural networks have improved detection accuracy to some extent. However, existing methods still have the following technical bottlenecks: First, traditional network structures struggle to capture both subtle local changes and global contextual information simultaneously, leading to blurred boundaries and missed detections; second, geological disaster areas often have irregular shapes and complex textures, making it difficult for the fixed receptive field of standard convolutional operations to adapt effectively; third, remote sensing images are easily affected by factors such as cloud cover, lighting changes, and terrain shadows during acquisition, resulting in insufficient robustness of existing methods; and finally, the imbalance between positive and negative samples also seriously affects the stability of model training.

[0004] Current methods based on U-Net and its variants have made some progress in feature extraction, but there is still considerable room for improvement in multi-scale feature fusion, enhancement of change regions, and adaptability to complex environments. Especially in the special application scenario of geological disaster monitoring, there is an urgent need for a change detection solution that can accurately identify multiple geological disaster types, preserve boundary details, and has strong generalization capabilities. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a change detection scheme that can accurately identify areas of geological disaster change, maintain boundary details, and has strong generalization ability.

[0006] In a first aspect, embodiments of this application provide a method for detecting geological hazard changes based on an improved twin U-Net and a central-surround dual-flow network, the method comprising:

[0007] S1. An improved twin U-Net network is constructed as the backbone for feature extraction. A dual encoder-decoder structure with shared weights is used to process geological disaster images from different periods. A deformable convolution module is integrated into the encoder to adaptively adjust the receptive field.

[0008] S2. Embed a center-surround dual-stream network structure in the decoding path of the twin U-Net network, extract detailed textures and boundary features through the center stream path, and obtain global context information through the surrounding stream path;

[0009] S3, Design Center - Surround Feature Fusion Module, fuses dual-path output features through a bidirectional feature guidance mechanism and uses a gated attention unit to adaptively calibrate the feature response;

[0010] S4. Introduce a contrastive feature learning mechanism to align the feature representations of the same changed instance in the feature embedding space and push the feature distance of the unchanged region.

[0011] S5. Implement multimodal degradation enhancement training by simulating cloud and fog obstruction, changes in lighting, and terrain shadows to construct training samples;

[0012] S6. Employ an end-to-end joint optimization strategy to minimize the objective function that includes difference perception loss and feature contrast loss;

[0013] S7. Generate geological hazard change detection maps based on the trained network, and identify and locate the areas of landslide, debris flow and ground subsidence change.

[0014] Secondly, embodiments of this application provide a geological hazard change detection system based on an improved twin U-Net and a center-surrounding dual-flow network, applied to the geological hazard change detection method based on an improved twin U-Net and a center-surrounding dual-flow network as described in the first aspect. The system includes:

[0015] The feature extraction module is used to construct an improved Siamese U-Net network as the feature extraction backbone. It adopts a dual encoder-decoder structure with shared weights to process geological disaster images from different periods. A deformable convolution module is integrated into the encoder to adaptively adjust the receptive field.

[0016] The dual-stream feature enhancement module is used to embed a center-surround dual-stream network structure in the decoding path of the twin U-Net network, extract detailed textures and boundary features through the center stream path, and obtain global context information through the surrounding stream path;

[0017] The feature fusion module is used to design the center-surround feature fusion module. It fuses dual-path output features through a bidirectional feature guidance mechanism and uses a gated attention unit to adaptively calibrate the feature response.

[0018] The contrastive learning module is used to introduce a contrastive feature learning mechanism, which aligns the feature representations of the same changed instance in the feature embedding space and pushes the feature distance of the unchanged region.

[0019] The degradation enhancement module is used to implement multimodal degradation enhancement training, and constructs training samples by simulating cloud and fog occlusion, changes in lighting and terrain shadows;

[0020] The joint optimization module is used to minimize the objective function, which includes difference-aware loss and feature contrast loss, by employing an end-to-end joint optimization strategy.

[0021] The results generation module is used to generate geological hazard change detection maps based on the trained network, and to identify and locate areas of landslide, debris flow and ground subsidence changes.

[0022] Thirdly, embodiments of this application provide an electronic device, including:

[0023] processor;

[0024] Memory used to store processor-executable instructions;

[0025] The processor is configured to implement the geological hazard change detection method based on the improved twin U-Net and the central surrounding dual-stream network as described in the first aspect when executing the instructions.

[0026] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the geological hazard change detection method based on an improved twin U-Net and a central-surround dual-stream network as described in the first aspect.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. By using an improved twin U-Net network and a center-surround dual-stream network structure, local details and global context information are effectively taken into account, significantly improving the positioning accuracy of changing boundaries.

[0029] 2. Adopting deformable convolution modules to adaptively adjust the receptive field, better adapting to the irregular shape characteristics of geological disaster areas.

[0030] 3. By enhancing training through multimodal degradation and contrastive feature learning mechanisms, the robustness and generalization ability of the model in complex environments are significantly improved.

[0031] 4. The innovative loss function design effectively alleviates the problem of imbalance between positive and negative samples and improves the stability of model training. Attached Figure Description

[0032] Figure 1 This is a schematic flowchart of a geological hazard change detection method based on an improved twin U-Net and a central surrounding dual-flow network, provided as an embodiment of this application.

[0033] Figure 2The architecture diagram of the geological hazard change detection system based on the improved twin U-Net and the central surrounding dual-flow network provided in this application.

[0034] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0036] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0037] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] Example 1

[0039] Figure 1 This is a schematic flowchart illustrating a geological hazard change detection method based on an improved twin U-Net and a central-surround dual-flow network, provided as an embodiment of this application. Figure 1 As shown, a geological hazard change detection method based on an improved twin U-Net and a central-surrounding dual-flow network includes:

[0040] S1. An improved twin U-Net network is constructed as the backbone for feature extraction. A dual encoder-decoder structure with shared weights is used to process geological disaster images from different periods. A deformable convolution module is integrated into the encoder to adaptively adjust the receptive field and enhance the ability to extract features of irregular geological disasters.

[0041] Specifically, in this embodiment, in step S1:

[0042] The improved Siamese U-Net network employs a symmetrical encoder-decoder architecture, where the encoder comprises multiple (e.g., 5) consecutive downsampling stages, each integrating the deformable convolutional module. This symmetrical encoder-decoder structure, through multi-stage downsampling, progressively extracts deep features, ensuring the systematic and hierarchical nature of feature extraction and laying the foundation for multi-scale feature fusion. For example, the encoder's 5 downsampling stages process features at different scales: Stage 1: original resolution, extracting low-level features such as edges and textures; Stage 2: 1 / 2 resolution, extracting local shape features; Stage 3: 1 / 4 resolution, extracting regional structural features; Stage 4: 1 / 8 resolution, extracting semantic features; Stage 5: 1 / 16 resolution, extracting global contextual features. The deformable convolution in each stage automatically adjusts the receptive field shape according to feature complexity; for example, for a curved landslide boundary, the convolutional kernel adaptively deforms along the boundary.

[0043] The deformable convolution module introduces a two-dimensional spatial offset into the standard convolution operation, enabling the convolution kernel sampling points to undergo adaptive spatial transformation based on the geometry of the input features. This offset is learned from the input feature map through additional convolutional layers. This step achieves adaptive transformation of the convolution kernel shape, enhancing the network's ability to extract irregular shape features of geological hazards and improving the flexibility of feature representation.

[0044] The dual encoders employ a weight-sharing mechanism, where the two encoders use the same convolution kernel parameters in each downsampling stage; convolution parameters are shared between the two encoder branches to ensure consistency in image feature extraction at different times, reduce model parameters, and improve training efficiency.

[0045] The decoder comprises multiple (e.g., 5) upsampling stages. By using skip connections, the feature maps of the corresponding stages from the encoder are concatenated with the decoder features. This progressively restores the spatial resolution of the feature maps and fuses shallow details. Combining deep semantic features with shallow detail features improves the localization accuracy of changing boundaries.

[0046] S2. A center-surround dual-stream network structure is embedded in the decoding path of the twin U-Net network. The center stream path extracts detailed textures and boundary features, while the surrounding stream path obtains global contextual information. The center-surround dual-stream network structure includes: a center stream path that extracts detailed textures and boundary features through high-resolution feature preservation and local attention mechanisms; and a surrounding stream path that obtains global contextual information through multi-scale dilated convolution and feature pyramid aggregation, forming a complementary relationship between local and global features.

[0047] Specifically, in this embodiment, S2 includes:

[0048] The centrally located dual-stream network structure is embedded in parallel at each decoding stage of the twin U-Net network. Through the dual-stream network integration structure, the dual-stream network is deeply integrated into each level of the decoder, achieving multi-level and multi-scale feature enhancement, and ensuring that the complementary characteristics of details and context are present from shallow to deep layers.

[0049] The central flow path extracts detailed texture and boundary features in the following ways: it directly receives high-resolution feature maps from the corresponding stage of the encoder; it employs a coordinate attention module to establish long-range spatial dependencies along the height and width directions respectively; it uses 1×1 convolution to compress feature dimensions and maintains feature integrity through residual connections. Through detailed extraction via the central flow path, it focuses on extracting and preserving local detailed features, strengthens spatial awareness through a coordinate attention mechanism, and uses residual connections to prevent the loss of detailed information, effectively preserving the fine features of changing boundaries.

[0050] The surrounding flow path acquires global context information in the following ways: four parallel dilated convolution branches are set with dilation rates of 2, 4, 8, and 12, respectively; a feature pyramid structure is adopted to fuse multi-scale features through a top-down path; multi-scale global context semantic information is captured through surrounding flow path context acquisition; the receptive field is expanded through multi-scale dilated convolution; and context features of different scales are integrated using the feature pyramid to enhance the semantic understanding of changing regions.

[0051] Global average pooling is applied to extract image-level semantic features. The output features of the central flow path and the surrounding flow path are concatenated along the channel dimension to form a complementary feature representation with both local details and global context. By fusing complementary features from both flows, feature complementarity between local details and global context is achieved. Channel-dimensional concatenation preserves the advantages of each feature, resulting in a more comprehensive and discriminative feature representation.

[0052] Specifically, the coordinate attention module in the central flow path is implemented in the following way:

[0053] Global pooling is performed on the input feature map along the height and width directions to generate feature descriptors in the height and width directions. The feature descriptors in the two directions are concatenated and convolved to decompose them into feature vectors in the height and width directions. Non-linear activation functions are applied to the feature vectors in the height and width directions to generate attention weight maps. The generated attention weight maps are multiplied with the original input feature map in a weighted manner to enhance the feature response at important locations.

[0054] Specifically, the coordinate attention mechanism of the central flow path, and the mathematical expression of coordinate attention: Let the input feature map... Pooling is performed along both the height and width directions:

[0055] ,

[0056] Take the first Row (index starts from 0), all columns in that row. (common Add the values ​​of each column together and then divide by the number of columns. We obtain a scalar representing the first... The average value of a row (i.e., the overall strength or feature level of that row).

[0057] ,

[0058] Take the first The column (indexed from 0, used to specify the exact location for the operation in the horizontal (width) direction) contains all rows. The sum of the values ​​of (a total of H rows, where H is the height of the feature map, i.e., the total number of rows) is then divided by the number of rows H. This yields a scalar representing the value of the first row. The average value of a column (i.e., the overall strength or characteristic level of that column).

[0059] The pooling results are concatenated and transformed by convolution:

[0060] ,

[0061] in, These are the feature vectors obtained by performing global average pooling on the feature map in the height and width directions, respectively. To combine the height-direction feature vector and the width-direction feature vector It involves splicing together elements in a spatial dimension. This represents a one-dimensional convolution operation, used to transform and reduce the dimensionality of concatenated features, fusing information from two directions. The activation function is usually ReLU or a similar nonlinear function.

[0062] Features decomposed into two directions:

[0063] ,

[0064] ,

[0065] in, , Two independent convolution or transformation functions are used to transform intermediate features. The attention weights are decomposed into the height and width directions respectively, generating two independent attention weights for each direction. This is an intermediate feature obtained after processing with a one-dimensional convolution and activation function. It encodes the dependencies of the original feature map in two spatial directions.

[0066] The final output is:

[0067] .

[0068] in, They represent the first Attention weights in the height direction, and the first Attention weights are listed in the width direction. This indicates the location of the original input feature map. The eigenvalues ​​(or eigenvectors, if multi-channel is considered) at that location. This indicates that after weighting by the coordinate attention mechanism, the output feature map is at the position The feature values ​​at a given location are the final output features, which enhance the feature responses at important spatial locations.

[0069] The feature pyramid structure in the surrounding flow path adopts a top-down fusion method: first, the deepest feature map is upsampled and concatenated with the feature map of the previous layer in the channel dimension; the concatenated feature map is fused by 3×3 convolution to reduce the number of channels and eliminate the aliasing effect caused by upsampling; the above process is repeated until feature maps of all scales are fused to form a feature representation with rich multi-scale contextual information.

[0070] Specifically, the multi-scale dilated convolution implementation around the flow path includes: output computation of four parallel dilated convolution branches:

[0071] ,

[0072] Input feature map (or image) The void ratio is The convolution operation, that is, inserting between adjacent weights of the convolution kernel. A hole (0 pixels) is created to expand the receptive field without increasing the number of parameters. The feature map obtained after dilated convolution has a receptive field that follows... It grew exponentially.

[0073] Feature pyramid fusion process:

[0074] Let the first Layer features are The fusion formula is:

[0075] ,

[0076] in, For the feature map of the corresponding layer, This is a 2x upsampling.

[0077] Dual-flow feature complementarity mechanism, feature splicing of central flow and surrounding flow:

[0078] ,

[0079] in, As a central flow output characteristic, This is the output characteristic of the surrounding flow.

[0080] Residual joins preserve detail integrity; residual joins in the central flow:

[0081] ,

[0082] Ensure that detailed information is not lost during multi-layered processing.

[0083] The feature fusion of the central flow path and the surrounding flow path adopts an adaptive weight learning mechanism: the importance of the two path features is dynamically adjusted through learnable weight parameters; 1×1 convolution is used to perform cross-channel information interaction on the concatenated features; and finally, the training process is stabilized through layer normalization.

[0084] S3, Design Center - Surround Feature Fusion Module, fuses dual-path output features through a bidirectional feature guidance mechanism and uses gated attention units to adaptively calibrate feature responses, enhancing feature representation of changing regions.

[0085] Specifically, in this embodiment, the center-surround feature fusion module in step S3 achieves feature fusion in the following way:

[0086] The bidirectional feature guidance mechanism includes: detail guidance from the central flow path to the surrounding flow path: using the detail features output by the central flow path as a gating signal, generating a spatial attention map through 1×1 convolution and Sigmoid activation function, and adjusting the spatial weights of the context features of the surrounding flow path.

[0087] Contextual Guidance from the Surrounding Flow Path to the Central Flow Path: The contextual features output from the surrounding flow path are used as gating signals. Channel attention weights are generated through global average pooling and fully connected layers to adjust the channel weights of the detailed features in the central flow path. A bidirectional feature guidance mechanism establishes a two-way interaction and modulation between detailed features and contextual features. A spatial attention map is generated using high-resolution detailed features from the central flow. This map highlights spatial locations related to fine boundaries and textures in the surrounding flow contextual features, suppressing unimportant background regions, thus guiding global contextual information to key areas of local change. A channel attention weight is generated using semantically rich contextual features from the surrounding flow. This weight evaluates and strengthens feature channels related to change in the central flow detailed features, while weakening irrelevant channels, making the representation of local details more discriminative.

[0088] The gated attention unit employs a dual-path gating mechanism: a spatial gating branch, which learns spatial selectivity weights through convolutional layers and a sigmoid function to enhance the spatial feature response of changing regions; and a channel gating branch, which learns channel importance weights through global pooling and fully connected layers to highlight feature channels relevant to change detection. This dual-path mechanism of the gated attention unit adaptively refines and calibrates the fused features after bidirectional guidance. Spatial gating further focuses on the spatial dimension, actively learning and enhancing feature responses belonging to changing regions to ensure higher activation levels in these areas. Channel gating selects and amplifies the feature channel information most effective in distinguishing between changing and non-changing regions. The combination of these two approaches achieves synergistic optimization of the fused features across the two key dimensions of space and channel, significantly improving feature quality.

[0089] The feature fusion further includes: element-wise addition of features processed by bidirectional guidance and gated attention, followed by feature reshaping through 3x3 convolution, and finally, residual connections to maintain the stability of gradient flow. Through feature reshaping and stabilization, the final integration of features is completed, ensuring the stability of the training process. Specifically, element-wise addition is an efficient feature fusion method that integrates the outputs of bidirectional guidance and gated attention. 3x3 convolution reshaping involves performing convolution operations on the fused features to learn local correlations between features, smooth the fusion result, and generate a new, more integrated feature representation. Residual connections effectively alleviate the vanishing gradient problem in deep networks, ensuring effective gradient backpropagation and making the training of this complex fusion module more stable and efficient.

[0090] S4. Introduce a contrastive feature learning mechanism to align the feature representations of the same changing instance in the feature embedding space and push the feature distance of the non-changing region, thereby improving the model's discriminability and robustness to changing regions.

[0091] Specifically, in this embodiment, the contrastive feature learning mechanism in step S4 is implemented in the following way:

[0092] Construct positive and negative sample pairs, and define the similarity and difference relationships in the feature space, including:

[0093] Anchor samples, positive samples, and negative samples are extracted from the feature map. Anchor samples and positive samples come from feature embeddings of the same change region, while negative samples come from feature embeddings of non-change regions or regions with different categories of change. Clustering criteria are established in the feature space to clarify which samples should be close and which samples should be far away. By constructing stricter discrimination boundaries through multiple types of negative samples (non-change regions and regions with different categories of change), a high-quality sample foundation is provided for contrastive learning.

[0094] Design feature space optimization, driving the structured reorganization of the feature space, including:

[0095] The comparison loss function based on cosine similarity is calculated using the following formula:

[0096] ,

[0097] in Indicates the features of the anchor sample. Indicates positive sample features. Indicates negative sample features. This represents the calculation of cosine similarity, where τ is the temperature hyperparameter. Cosine similarity is more suitable for measuring the directional consistency of high-dimensional features; by bringing positive samples closer and pushing negative samples further away, it encourages similar changing features to form tight clusters in the embedding space; the temperature hyperparameter τ controls the degree of attention given to difficult samples.

[0098] Implementing feature space constraints includes: in the feature embedding space, optimizing to maximize the cosine similarity between positive sample pairs and minimize the cosine similarity between negative sample pairs; employing a hard example mining strategy to select negative sample pairs with moderate similarity values ​​for focused optimization, thereby improving the model's ability to distinguish boundary samples. This avoids over-optimization on overly simple samples, concentrating training resources on optimizing sample pairs with similar features that are easily confused, effectively improving the clarity of the decision boundary.

[0099] The multi-scale contrastive learning approach includes: establishing contrastive learning constraints on the deep feature maps of the encoder and the shallow feature maps of the decoder; achieving spatial alignment of multi-scale feature embeddings through a feature pyramid structure; and achieving multi-level feature space consistency. Specifically, deep features ensure semantic consistency, while shallow features maintain consistency at the detail level; a unified feature representation space is constructed from local to global perspectives.

[0100] Based on the contrastive feature learning mechanism, a complete implementation framework is provided. From sample construction to loss design, optimization strategy, and multi-scale expansion, a systematic feature space optimization scheme is formed. This mechanism significantly improves the model's ability to distinguish changing regions through a structured feature space, especially its robustness in recognizing subtle changes in complex backgrounds. This contrastive learning mechanism complements the main structure of the network, utilizing the features extracted by the backbone network and improving the feature quality of the backbone network through feature space constraints, forming a virtuous cycle of collaborative optimization.

[0101] S5. Implement multimodal degradation enhancement training by simulating cloud and fog occlusion, lighting changes and terrain shadows to construct training samples and improve the model's generalization ability in practical applications.

[0102] Specifically, in this embodiment, the multimodal degradation enhancement training in step S5 is implemented in the following way:

[0103] The simulation of cloud and fog occlusion in remote sensing images includes: using randomly generated elliptical regions to simulate cloud and fog occlusion of different shapes, with the lengths of the major and minor axes of the ellipse randomly selected within the range of 10%-40% of the image size; adding Gaussian blur to the occluded areas, with the blur kernel size randomly selected within the range of 3×3 to 15×15 pixels, to simulate the loss of detail caused by clouds and fog; and applying brightness attenuation to the occluded areas, with the attenuation coefficient randomly set within the range of 0.3-0.7, to simulate the attenuation effect of light penetrating clouds and fog, thereby improving the model's robustness to cloud and fog interference.

[0104] The simulation of lighting changes under different times and weather conditions includes: randomly adjusting image brightness and saturation in the HSV color space, with brightness adjusted by ±30% of the original value and saturation adjusted by ±20% of the original value. HSV space adjustments are more consistent with human visual perception. Directional lighting effects are added, using gradient masks to simulate lighting changes at different angles, with directional lighting simulating the effects of different sun angles. Randomly selecting local areas of the image for uneven lighting processing simulates the changes in light and shadow in actual shooting. Local uneven lighting simulates the shadow effects of actual terrain, enhancing the model's adaptability to lighting changes.

[0105] Terrain shadow simulation simulates shadow changes caused by real topographic factors, including: generating realistic terrain shadows based on digital elevation model data and ensuring the authenticity of shadow shapes based on DEM data; adding random affine transformations to simulate shadow effects under different solar altitude angles, and affine transformations to simulate changes in solar position at different seasons and times; and simultaneously performing contrast reduction and color shift processing within the shadow area to avoid misjudging terrain shadows as geological disaster changes and improve detection accuracy in complex terrain areas.

[0106] The degradation enhancement strategy also includes: applying all degradation operations in random combinations with a probability of 0.5; dynamically generating different degradation samples during each training iteration; and randomly sampling degradation intensity parameters within a preset range. This constructs complex degradation scenarios close to the real world, randomly combining and simulating the complex situation of multiple degradations coexisting in the actual environment. Parameter randomization ensures the diversity of training samples, avoids overfitting the model to specific degradation patterns, and greatly improves the model's generalization ability.

[0107] Building upon the previously proposed multimodal degradation enhancement training, the above steps establish a systematic degradation simulation framework. By independently simulating and randomly combining three main degradation modes—clouds, illumination, and terrain—a training environment covering major real-world interference factors is constructed. This enhancement strategy does not rely on a large amount of real degradation data; instead, it effectively expands the distribution range of training samples through algorithmic simulation, enabling the model to better adapt to various complex real-world application scenarios during deployment. This technical solution specifically addresses the common environmental interference problems in remote sensing detection of geological disasters, significantly improving the practical value and reliability of the method in real-world applications through forward-looking data enhancement.

[0108] S6. An end-to-end joint optimization strategy is adopted to minimize the objective function that includes difference perception loss and feature contrast loss, so as to solve the problem of imbalance between positive and negative samples and optimize the overall performance of the model.

[0109] Specifically, in this embodiment, the end-to-end joint optimization strategy in step S6 is implemented in the following way:

[0110] The difference-perceived loss function consists of two parts: a weighted focus loss, used to address the class imbalance between changed and unchanged regions, expressed as:

[0111] ,

[0112] in α is the probability value predicted by the model, γ is the class weight coefficient, and γ is the focusing parameter; the focus loss, through the weight coefficient α and the modulation factor γ, focuses on solving the extreme imbalance problem of positive and negative samples (changed / unchanged pixels) in change detection, so that the model pays more attention to samples that are difficult to classify.

[0113] Boundary consistency loss: A boundary constraint term based on the Dice coefficient is used to enhance the positioning accuracy of changing boundaries.

[0114] ,

[0115] in To predict the boundary region, For the true boundary region, ε is the smoothing term; Boundary consistency loss: Based on the Dice coefficient, the boundary positioning accuracy of the changing region is specifically optimized to ensure the integrity and accuracy of the changing boundary.

[0116] The feature contrast loss function employs an improved normalized temperature-scaled cross-entropy loss: positive and negative sample pairs are constructed within a batch, and contrastive learning loss is calculated; a hard example mining mechanism is introduced, focusing on sample pairs that are difficult to distinguish in the feature space; contrastive learning constraints are implemented at the feature level, using positive and negative sample pairs to bring similar features closer together and push away dissimilar features. The hard example mining mechanism focuses on optimizing samples that are difficult to distinguish in the feature space, enhancing the discriminativeness and robustness of feature representation.

[0117] The objective function achieves collaborative optimization of pixel-level classification and feature-level learning, and its final form is:

[0118] ,

[0119] in The optimal values ​​for each loss term are determined through grid search.

[0120] The joint optimization employs an adaptive moment estimation algorithm, with an initial learning rate set to 0.001, and uses a cosine annealing strategy to dynamically adjust the learning rate, providing a stable and efficient training optimization process.

[0121] The above steps establish a complete optimization framework based on an end-to-end joint optimization strategy. It addresses the core challenge of pixel-level classification through difference-aware loss, improves feature quality through feature contrast loss, and finally achieves multi-objective collaborative optimization through weighted combination. This composite loss function design fully considers multiple key factors in change detection tasks, such as sample imbalance, boundary accuracy, and feature discriminability, and, combined with adaptive optimization algorithms, ensures the effectiveness and stability of model training.

[0122] S7. Generate geological hazard change detection maps based on the trained network, identifying and locating areas of landslides, debris flows, and ground subsidence changes. The trained network accurately identifies and locates change areas for various geological hazards, outputting detection results with detailed spatial distribution.

[0123] Specifically, in this embodiment, step S7 includes:

[0124] Post-processing of the changed region: An adaptive threshold segmentation method is used for the initial change probability map output by the network; Morphological optimization of the changed object: Area opening operation is used to remove noise points with an area of ​​less than 10 pixels; Morphological closing operation is used to fill the voids inside the changed region, and the structuring element size is 3×3; Edge smoothing is used to optimize the shape of the changed boundary to retain the morphological characteristics of the real geological disaster.

[0125] Change type classification and identification: Classification is performed based on the geometric and texture features of the changed area. Geometric features include area, perimeter, compactness, and aspect ratio; texture features include contrast, correlation, and entropy of the gray-level co-occurrence matrix; a support vector machine classifier is used to distinguish between three types of geological hazards: landslides, debris flows, and ground subsidence.

[0126] Results visualization output: Different colors are used to represent different types of geological hazard change areas; a pseudo-color map containing change intensity information is generated, with change intensity represented by color depth; a change statistics report is output, including the area of ​​the changed area, the distribution of change types, and spatial location information.

[0127] Example 2

[0128] like Figure 2 As shown, this application provides an architecture diagram of a geological hazard change detection system based on an improved twin U-Net and a center-surrounding dual-stream network, which is applied to the geological hazard change detection system based on an improved twin U-Net and a center-surrounding dual-stream network as described in Embodiment 1. It includes a feature extraction module 11, a dual-stream feature enhancement module 12, a feature fusion module 13, a contrastive learning module 14, a degradation enhancement module 15, a joint optimization module 16, and a result generation module 17.

[0129] The feature extraction module 11 is used to construct an improved twin U-Net network as the feature extraction backbone. It adopts a dual encoder-decoder structure with shared weights to process geological disaster images from different periods. A deformable convolution module is integrated in the encoder to adaptively adjust the receptive field.

[0130] The dual-stream feature enhancement module 12 is used to embed a center-surround dual-stream network structure in the decoding path of the twin U-Net network, extract detailed textures and boundary features through the center stream path, and obtain global context information through the surrounding stream path.

[0131] Feature fusion module 13 is used to design a center-surround feature fusion module, which fuses dual-path output features through a bidirectional feature guidance mechanism and uses a gated attention unit to adaptively calibrate the feature response.

[0132] The contrastive learning module 14 is used to introduce a contrastive feature learning mechanism to align the feature representations of the same changed instance in the feature embedding space and push the feature distance of the unchanged region.

[0133] The degradation enhancement module 15 is used to implement multimodal degradation enhancement training by constructing training samples by simulating cloud cover, lighting changes and terrain shadows.

[0134] Joint optimization module 16 is used to minimize an objective function that includes difference-aware loss and feature contrast loss by employing an end-to-end joint optimization strategy.

[0135] Result generation module 17 is used to generate geological hazard change detection maps based on the trained network, and to identify and locate the areas of change in landslides, debris flows and ground subsidence.

[0136] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0137] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0138] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0139] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0140] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0141] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0142] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0143] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0144] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for detecting geological hazard changes based on an improved twin U-Net and a central-surrounding dual-flow network, characterized in that, Includes the following steps: S1. An improved twin U-Net network is constructed as the backbone for feature extraction. A dual encoder-decoder structure with shared weights is used to process geological disaster images from different periods. A deformable convolution module is integrated into the encoder to adaptively adjust the receptive field. S2. Embed a center-surround dual-stream network structure in the decoding path of the twin U-Net network, extract detailed textures and boundary features through the center stream path, and obtain global context information through the surrounding stream path; S3, Design Center - Surround Feature Fusion Module, fuses dual-path output features through a bidirectional feature guidance mechanism and uses a gated attention unit to adaptively calibrate the feature response; S4. Introduce a contrastive feature learning mechanism to align the feature representations of the same changed instance in the feature embedding space and push the feature distance of the unchanged region. S5. Implement multimodal degradation enhancement training by simulating cloud and fog obstruction, changes in lighting, and terrain shadows to construct training samples; S6. Employ an end-to-end joint optimization strategy to minimize the objective function that includes difference perception loss and feature contrast loss; S7. Generate geological hazard change detection maps based on the trained network, and identify and locate the areas of change in landslides, debris flows and ground subsidence; In step S1: The improved twin U-Net network adopts a symmetric encoder-decoder architecture, in which the encoder contains multiple downsampling stages and each stage integrates a deformable convolutional module. The deformable convolution module introduces a learnable spatial offset, enabling the convolution kernel to adaptively adjust the sampling position according to the input features. The dual encoders employ a weight-sharing mechanism, using the same convolutional kernel parameters in each downsampling stage; The decoder includes multiple upsampling stages, and uses skip connections to concatenate encoder features and decoder features through channels to restore spatial resolution. S2 includes: The central flow path extracts and preserves detailed texture and boundary features by receiving high-resolution feature maps from the encoder, combined with coordinate attention mechanisms and residual connections. The surrounding flow path obtains multi-scale global context information through parallel multi-scale dilated convolution and feature pyramid structure, combined with global pooling operation; The output features of the central flow path and the surrounding flow path are concatenated in the channel dimension to form a complementary feature representation that simultaneously contains local details and global context. The center-surround feature fusion module in step S3 achieves feature fusion in the following way: The bidirectional feature guidance mechanism includes: detail guidance from the central flow path to the surrounding flow path: using the detail features output by the central flow path as a gating signal, generating a spatial attention map through convolution and activation functions, and adjusting the spatial weights of the context features of the surrounding flow path; Context guidance from the surrounding flow path to the central flow path: The context features output from the surrounding flow path are used as gating signals, and channel attention weights are generated through global average pooling and fully connected layers to adjust the channel weights of the detailed features of the central flow path. The gated attention unit employs a dual-path gating mechanism: a spatial gating branch, which learns spatial selectivity weights through convolutional layers and activation functions to enhance the spatial feature response of changing regions; and a channel gating branch, which learns channel importance weights through global pooling and fully connected layers to highlight feature channels related to change detection.

2. The geological hazard change detection method based on an improved twin U-Net and a central-surround dual-flow network according to claim 1, characterized in that, The coordinate attention module in the central flow path processes the feature map through the following steps: global pooling is performed along the height and width directions to generate feature descriptors; the obtained descriptors are concatenated and decomposed back into two directions after convolution; attention weight maps are generated by non-linear activation; and finally, they are weighted and fused with the original feature map to enhance key position features. The feature pyramid structure in the surrounding flow path fuses multi-scale features in the following manner: upsampling of deep feature maps and concatenating them with previous layer features; fusing the concatenated results through convolution to reduce the number of channels and suppress aliasing; iterating this process to integrate contextual information at all scales; The output features of the two paths are fused through the following mechanisms: dynamically adjusting the importance of the dual-path features using learnable weights; using convolution to achieve cross-channel interaction; and ensuring training stability through layer normalization.

3. The geological hazard change detection method based on an improved twin U-Net and a central-surround dual-flow network according to claim 1, characterized in that, The contrastive feature learning mechanism in step S4 is implemented in the following way: Construct positive and negative sample pairs, where the anchor sample and the positive sample come from the feature embedding of the same change region, and the negative sample comes from the feature embedding of the non-change region or a different category of change region. Using a contrastive loss function based on cosine similarity, the cosine similarity between positive sample pairs is maximized and the cosine similarity between negative sample pairs is minimized in the feature embedding space through optimization. A hard example mining strategy is adopted to select negative sample pairs with moderate similarity values ​​for focused optimization; Contrastive learning constraints are established for the deep feature maps of the encoder and the shallow feature maps of the decoder, respectively.

4. The geological hazard change detection method based on an improved twin U-Net and a central-surround dual-flow network according to claim 1, characterized in that, The multimodal degradation enhancement training in step S5 includes: The cloud and fog occlusion simulation is achieved by randomly generating elliptical occlusion regions and applying Gaussian blur and brightness attenuation. Simulate lighting changes by adjusting brightness and saturation in the HSV color space, and add directional lighting effects and localized uneven lighting processing. Terrain shadow simulation generates realistic shadows based on a digital elevation model and simulates shadow effects at different solar altitude angles through affine transformations. Each degradation operation is applied in a random combination with a set probability, and the degradation intensity parameter is randomly sampled within a preset range.

5. The geological hazard change detection method based on an improved twin U-Net and a central-surround dual-flow network according to claim 1, characterized in that, The end-to-end joint optimization strategy in step S6 is implemented in the following way: The objective function includes a difference-aware loss function and a feature contrast loss function, wherein the difference-aware loss function consists of a weighted focus loss and a boundary consistency loss; The feature contrast loss function adopts an improved normalized temperature-scaled cross-entropy loss, constructs positive and negative sample pairs within a batch, and introduces a hard example mining mechanism. The objective function is composed of a weighted sum of difference-perceived loss and feature contrast loss; An adaptive moment estimation algorithm is used for joint optimization, and a dynamic learning rate adjustment strategy is adopted.

6. The geological hazard change detection method based on an improved twin U-Net and a central-surround dual-flow network according to claim 1, characterized in that, Step S7 includes: An adaptive threshold segmentation method is used to process the initial change probability map of the network output; Morphological optimization is performed on the changed regions, including removing noise points, filling internal voids, and optimizing boundary shapes; Classification and identification are performed based on the geometric and textural features of the changing areas to distinguish different types of geological hazards; Generate visual detection results, including using different color codes to represent the type of change, displaying the intensity of the change, and outputting a report of change statistics.

7. A geological hazard change detection system based on an improved twin U-Net and a central-surrounding dual-flow network, applied to the geological hazard change detection method based on an improved twin U-Net and a central-surrounding dual-flow network as described in any one of claims 1 to 6, characterized in that, The system includes: The feature extraction module is used to construct an improved Siamese U-Net network as the feature extraction backbone. It adopts a dual encoder-decoder structure with shared weights to process geological disaster images from different periods. A deformable convolution module is integrated into the encoder to adaptively adjust the receptive field. The dual-stream feature enhancement module is used to embed a center-surround dual-stream network structure in the decoding path of the twin U-Net network, extract detailed textures and boundary features through the center stream path, and obtain global context information through the surrounding stream path; The feature fusion module is used to design the center-surround feature fusion module. It fuses dual-path output features through a bidirectional feature guidance mechanism and uses a gated attention unit to adaptively calibrate the feature response. The contrastive learning module is used to introduce a contrastive feature learning mechanism, which aligns the feature representations of the same changed instance in the feature embedding space and pushes the feature distance of the unchanged region. The degradation enhancement module is used to implement multimodal degradation enhancement training, and constructs training samples by simulating cloud and fog occlusion, changes in lighting and terrain shadows; The joint optimization module is used to minimize the objective function, which includes difference-aware loss and feature contrast loss, by employing an end-to-end joint optimization strategy. The results generation module is used to generate geological hazard change detection maps based on the trained network, and to identify and locate areas of landslide, debris flow and ground subsidence changes.

Images