Landslide image recognition method based on eigen decomposition and mixed channel spatial attention

Abstract

The present application belongs to the technical field of computer vision and geological disaster monitoring, and particularly relates to a landslide image recognition method based on feature decomposition and mixed channel spatial attention. The present application is based on a few sample conditions, and performs multi-channel feature decomposition on images to construct a multi-channel input feature map. The multi-channel input feature map is input into a landslide recognition model constructed based on a U-Net architecture for preliminary segmentation prediction. The model introduces a channel-spatial mixed attention mechanism, and adopts a data enhancement, L2 regularization and Bayesian optimization tuning strategy for training. The preliminary recognition result output by step S2 is subjected to SEEDS superpixel segmentation, normalized cut and connected component analysis to obtain a clear image containing the main target. Finally, verification is performed. The method of the present application has high precision recognition, strong feature expression capability, good generalization performance and high calculation efficiency, and is particularly suitable for high-precision and high-robustness landslide image segmentation under a sample scarcity condition.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and geological disaster monitoring technology, specifically involving a landslide image recognition method based on feature decomposition and hybrid channel spatial attention, which is suitable for high-precision and robust landslide image segmentation under conditions of scarce samples. Background Technology

[0002] Landslides, as a typical geological disaster, are characterized by their suddenness, destructive power, and wide impact, posing a serious threat to people's lives and property and infrastructure. Accurate identification of landslide deformation areas and early monitoring are crucial foundations for landslide risk warning and disaster prevention and mitigation. With the rapid development of remote sensing technology, sensor networks, and artificial intelligence, landslide monitoring technology has gradually evolved from traditional manual on-site surveys to an intelligent identification system based on multi-source data fusion. However, many challenges remain in practical applications, such as the challenges of landslide monitoring technology and landslide identification under limited sample conditions.

[0003] The development of landslide monitoring technology has roughly gone through three stages. Specifically, the first stage is the traditional application stage, which mainly relies on manual inspections and ground instrument measurements. Although it is highly accurate and intuitive, it has problems such as limited spatial coverage, high manpower and time costs, and significant constraints from terrain and climate conditions, making it difficult to achieve large-scale and continuous monitoring.

[0004] The second stage is the application stage of remote sensing and geographic information technology. Researchers began to use multi-temporal remote sensing imagery, digital elevation models, and other methods for landslide area identification and mapping. For example, they used airborne lidar data to identify rainfall-induced landslides, classified shallow landslide scenes based on point cloud analysis methods, or monitored rock slope deformation using ground-based laser scanning point cloud segmentation frameworks. Although these methods expanded the spatial scale of monitoring, they were still limited by the complexity of data processing, the influence of weather and topography, and insufficient identification accuracy. Moreover, most methods were only applicable to areas with relatively simple tones and textures.

[0005] The third stage is driven by artificial intelligence and machine learning, mainly manifested in the fact that machine learning and deep learning technologies have provided new approaches for intelligent landslide identification. Examples include semi-automatic landslide identification based on multi-layer semantic networks and semantic segmentation methods for predicting potential landslide areas. However, these methods typically rely on large-scale, high-quality labeled data for training, resulting in limitations such as high computational resource consumption and weak model generalization ability under conditions of scarce samples.

[0006] For landslide identification under limited sample conditions, the suddenness, geographical location, and environmental complexity of landslide events make it extremely difficult to obtain large amounts of accurately labeled landslide image data. On the one hand, landslides often occur in remote, high-risk areas, and data collection is limited by geographical conditions and safety factors; on the other hand, the post-disaster environment is chaotic, making image acquisition and processing difficult. Therefore, practical applications often face problems such as small sample size, uneven quality, and high annotation costs. Under limited sample conditions, existing deep learning models are prone to overfitting, and their recognition accuracy and robustness are insufficient to meet practical engineering needs. How to reduce the model's dependence on large-scale labeled data and improve landslide identification performance in limited sample scenarios has become a research hotspot and challenge in the field of landslide monitoring.

[0007] To address the aforementioned issues, several patented technologies have proposed improvement solutions. For example, the patent "Landslide Image Recognition Method, System, Device and Medium" (application number 2025109748618) enhances the discriminative power and structural integrity of feature representation through residual feature extraction, multi-scale feature fusion, and channel and spatial attention mechanism enhancement. Another patent, "A Landslide Image Recognition Method and System Based on DenseNet" (application number 2021116038437), constructs a multi-factor fusion landslide sample library and utilizes the DenseNet network to achieve efficient feature transfer, reducing the number of parameters while mitigating overfitting and improving recognition performance under conditions of limited samples.

[0008] However, existing methods still have room for improvement in areas such as multi-level feature interaction, lightweight and adaptive attention mechanisms, and preservation of edge and detail information. In particular, given the characteristics of landslide images with few samples, such as large intra-class differences, blurred boundaries, and high similarity to the background environment, further optimization of the network structure design is needed to achieve more refined feature decoupling and more efficient utilization of contextual information.

[0009] In conclusion, developing a landslide identification method that can achieve high accuracy and strong robustness under conditions of few samples is of great significance for promoting the practical and intelligent development of landslide monitoring technology. Summary of the Invention

[0010] The technical problem to be solved by this invention is to provide a landslide image recognition method based on feature decomposition and hybrid channel spatial attention. Specifically, it proposes an improved U-Net model based on multi-channel feature decomposition and efficient channel-spatial hybrid attention mechanism (EHCS Attention). This model aims to significantly improve the recognition accuracy and generalization ability of landslide images with few samples by multi-dimensional feature fusion, innovative attention mechanism and designed post-processing method, through multi-level feature fusion, lightweight attention module and post-processing optimization.

[0011] The technical solution adopted is as follows: A landslide image recognition method based on feature decomposition and hybrid channel spatial attention collects few-sample landslide image data from different regions and time periods, and labels the landslide areas in each image. A portion of the images are used as the training and validation set, and the remaining images are used as the test set. The few-sample conditional landslide recognition method includes the following steps: S1: In the preprocessing stage, the image is decomposed into multiple channels to extract feature information including color, space and frequency, so as to enrich the input features of the model and construct a multi-channel input feature map. S2: In the preliminary identification stage, the multi-channel feature map is input into the landslide identification model built on the U-Net architecture for preliminary segmentation and prediction. The model introduces a channel-space hybrid attention mechanism and is trained using data augmentation, L2 regularization and Bayesian optimization strategies. S3: In the post-processing stage of the model, the preliminary recognition results output from step S2 are refined using methods including SEEDS (Superpixels Extracted via Energy-Driven Sampling) segmentation, normalized cutting, and connected component analysis to obtain a clear image containing the main target.

[0012] S4: Verify the high accuracy and robustness of the model of this invention.

[0013] Preferably, step S1 specifically includes: Extract the RGB (Red, Green, Blue) color channels of the image as basic color features; Perform wavelet transform on the image to obtain the low-frequency channel containing the global structure and the high-frequency channel containing detailed information; Perform a Fourier transform on the image and take the amplitude spectrum, then perform a logarithmic transform to obtain the frequency domain features. The RGB color channels, wavelet transform subbands, and Fourier transform amplitude spectrum features are combined to form a multi-channel input image.

[0014] Preferably, the landslide identification model used in step S2 is based on U-Net architecture, and a high-efficiency channel-spatial hybrid attention module is embedded in its encoder and / or decoder modules; The workflow of the high-efficiency channel-spatial hybrid attention module is as follows: S21. Calculate the channel attention weights and spatial attention weights for the input feature map respectively; The channel attention weights are obtained by performing global average pooling and global max pooling on the feature map in the spatial dimension, and then summing the outputs of the two through a shared network. Spatial attention weights are obtained by performing global average pooling and global max pooling on the feature map along the channel dimension, concatenating the outputs of the two, and then performing a convolution operation. S22. Multiply the calculated channel attention weights and spatial attention weights element-wise to generate a fused attention weight map. Multiply this weight map element-wise with the original input feature map to obtain the weighted output feature map.

[0015] Preferably, the calculation steps for channel attention weights include: Global average pooling and global max pooling are performed on the input feature map in the spatial dimension to calculate the global statistics for each channel; For the input feature map, global average pooling (GAP) is performed, and its formula is shown in equation (1): (1)； in, F(i, j, k) Indicates the first k Each channel is in i , j Location feature values; H Indicates the height of the feature map. W Indicates width, C Indicates the number of channels; Global max pooling (GMP) is used for calculation, and its formula is shown in equation (2): (2)； The results of global average pooling and global max pooling are fused through a fully connected layer and an activation function to obtain the weights for each channel. A ch The calculation method is shown in equation (3): (3)； in, W 1 and W 2 It is the learned weight matrix. σ This represents the sigmoid activation function, used to normalize channel weights.

[0016] Preferably, the steps for calculating spatial attention weights include: Global average pooling and global max pooling are performed along the channel dimension to calculate the importance of each spatial location, enabling the model to adjust its attention based on the information at each location; The formula for calculating the spatial global average pooling is shown in equation (4): (4)； in, F(i, j, k) For the input feature map in i , j The characteristic value of the location.

[0017] The formula for calculating the global max pooling in space is shown in equation (5): (5)； The results from both are fused via convolution to generate spatial attention weights. A sp The calculation method is shown in equation (6): (6)； in, Conv2D It's a convolution operation. σ It is the sigmoid activation function.

[0018] Preferably, channel and spatial attention are fused by pixel-wise multiplication, as shown in Equation (7), and applied to the original feature map F to generate a weighted feature map, as shown in Equation (8). (7)； (8)； By combining optimization and convolution operations, EHCS Attention can efficiently capture important features under multi-channel input and complex spatial structures.

[0019] Preferably, in step S2, the Bayesian optimization algorithm is used to automatically tune the hyperparameter combination during the model training phase. The hyperparameter combination includes, but is not limited to: random seed, learning rate, L2 regularization coefficient, batch size, data augmentation parameters, and number of training rounds. Bayesian optimization fits the relationship between hyperparameters and model validation set performance by constructing a Gaussian process surrogate model, and selects the next hyperparameter point to be evaluated based on the acquisition function, thus efficiently finding the optimal solution.

[0020] The preferred mathematical expression for Bayesian optimization is shown in equation (9): (9)； in, θ Indicates a combination of hyperparameters. θ next This indicates the next combination of hyperparameters, where GP stands for Gaussian process model.

[0021] Preferably, step S3 specifically includes: Step S31: Use the SEEDS superpixel segmentation algorithm to over-segment the preliminary recognition result image to generate multiple superpixel regions; Step S32: Calculate the color mean of adjacent superpixel regions, construct a region adjacency graph, and use a normalization cutting method to segment the graph and merge similar superpixel regions; Step S33: Threshold binarize the merged image and remove noise areas smaller than the preset threshold through connected component analysis, while retaining the main landslide target areas.

[0022] Preferably, the landslide image recognition system based on feature decomposition and hybrid channel spatial attention includes: The feature decomposition module is used to perform the operation of step S1, converting a single landslide image into a multi-channel feature map rich in multi-dimensional information; The intelligent recognition module has a built-in landslide recognition model based on the U-Net architecture described in step S2, which is used to receive the multi-channel feature map and output a preliminary landslide area prediction map. The post-processing optimization module is used to perform the operation of step S3, refine the preliminary landslide area prediction map, and output the final identification result; The analysis and verification module is used to perform step S4 to test the remaining images, verifying high accuracy and robustness.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) High recognition accuracy: Under the condition of a small number of samples, the multi-channel feature decomposition and fusion strategy of the present invention enhances the model's representation ability under the condition of scarce samples. The Dice coefficient reaches 0.93, which is significantly better than the original U-Net (0.58) and the region growing method (0.21).

[0024] (2) Strong feature representation ability: Multi-channel decomposition provides multi-dimensional features of color, space and frequency, and the efficient hybrid channel spatial attention mechanism (EHCS) focuses on key information; (3) Good generalization performance: This method has strong robustness in high noise and complex backgrounds, and can adapt to a variety of different geological disaster scenarios. It integrates data augmentation, L2 regularization, Bayesian optimization and post-processing to comprehensively improve the robustness and practicality of the model. (4) High computational efficiency: The EHCS attention mechanism significantly reduces computational complexity and improves real-time monitoring performance through simple convolution and pooling operations.

[0025] (5) Good post-processing optimization effect: significantly improves boundary clarity and reduces false detection and missed detection. It is a lightweight, high-precision identification system suitable for geological disaster monitoring, with good engineering portability and real-time performance. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 The image shows the identification results of the method of the present invention; where b, d, f, h, j, l are the original images of the landslide area in the remaining 6 test sets, and a, c, e, g, i, k are the identification images of the original images.

[0028] Figure 3 This is a comparison chart of the accuracy of Embodiment 1 and the comparative example of the present invention. Detailed Implementation

[0029] The accompanying drawings are for illustrative purposes only; to enable those skilled in the art to better understand the technical solutions of the present invention, the preferred embodiments of the present invention are described below in conjunction with specific examples, but should not be construed as limiting the present patent.

[0030] Example 1 like Figure 1 As shown, the landslide image recognition method based on feature decomposition and hybrid channel spatial attention first collected 36 landslide images from different regions and time periods, and accurately labeled the landslide areas in each image. Of these, 30 images were allocated to the training and validation sets, and the remaining 6 were used for the test set.

[0031] The few-sample conditional slope identification method includes the following steps: S1: In the preprocessing stage, the image is decomposed into multiple channels to extract feature information including color, space and frequency, so as to enrich the input features of the model and construct a multi-channel input feature map. S2: In the preliminary identification stage, the multi-channel feature map is input into the landslide identification model built on the U-Net architecture for preliminary segmentation and prediction. The model introduces a channel-space hybrid attention mechanism and is trained using data augmentation, L2 regularization and Bayesian optimization strategies. S3: In the post-processing stage of the model, the preliminary recognition results output in step S2 are refined using methods including SEEDS superpixel segmentation, normalized cutting, and connected component analysis to obtain a clear image containing the main target.

[0032] S4: Verify the high accuracy and robustness of the model of this invention.

[0033] To meet the above analytical requirements, this invention provides a landslide image recognition system based on feature decomposition and hybrid channel spatial attention, comprising: The feature decomposition module is used to perform the operation of step S1, converting a single landslide image into a multi-channel feature map rich in multi-dimensional information; The intelligent recognition module has a built-in landslide recognition model based on the U-Net architecture described in step S2, which is used to receive the multi-channel feature map and output a preliminary landslide area prediction map. The post-processing optimization module is used to perform the operation of step S3, refine the preliminary landslide area prediction map, and output the final identification result; The analysis and verification module is used to perform step S4 to test the remaining images, verifying high accuracy and robustness.

[0034] Under limited sample conditions, improving the generalization ability of the model requires richer feature representations. This embodiment 1 extracts information from different dimensions by performing multi-channel decomposition on the image. Step S1 specifically includes: (1) In the case of a small number of samples, color difference is often the main basis for the model to distinguish different regions. Therefore, the RGB channels are used to capture information of the image in different color dimensions.

[0035] (2) Perform wavelet transform on the image to obtain the low-frequency channel containing the global structure and the high-frequency channel containing detailed information; The image is subjected to a Fourier transform, and the amplitude spectrum is extracted. After a logarithmic transform, this spectrum is used as a frequency domain feature. Specifically, wavelet transform is used to obtain the amplitude spectra of the low-frequency channel (cA) containing global structure and the high-frequency channels (cH, cV, cD) containing detailed information, reflecting the intensity of different frequency components. The high-frequency components correspond to details and edges in the image, while the low-frequency components reflect the overall structure of the image. Therefore, the amplitude spectrum is taken as one channel of the image and a logarithmic transform is performed.

[0036] The RGB color channels, wavelet transform subbands, and Fourier transform amplitude spectrum features are combined to form an 8-channel input image. This combination of channels provides the model with rich information across different dimensions, including color, spatial composition, and frequency, helping the model better understand and process complex image structures.

[0037] The landslide identification model used in step S2 is based on U-Net architecture, and an efficient channel-space hybrid attention module is embedded in its encoder and / or decoder modules.

[0038] U-Net is a convolutional neural network architecture specifically designed for image segmentation tasks. Its greatest advantage lies in preserving high-resolution features through skip connections, while simultaneously extracting contextual information step-by-step through an encoder-decoder structure. Especially under conditions of few samples, U-Net achieves good generalization performance. Therefore, this invention adopts U-Net as the basic architecture for the landslide image segmentation task.

[0039] To describe the model architecture in more detail, Table 1 shows the specific structure of the U-Net used in this invention and the operational details of each layer.

[0040] Table 1. U-Net architecture adopted in this invention To improve the model's ability to represent image features under limited sample conditions, multi-channel decomposition was introduced to artificially construct richer image inputs. While these features provide detailed information about landslide images at different scales and frequency domains, this multi-channel input necessitates a more efficient mechanism for the neural network to select and focus on the most useful features. Furthermore, landslide monitoring is a task with high real-time requirements, prompting the inventors to optimize computational efficiency to achieve rapid monitoring.

[0041] Therefore, this embodiment 1 designs an efficient channel-spatial hybrid attention mechanism (EHCSAttention) specifically for handling complex structures with multi-channel input. Specifically, the workflow of the efficient channel-spatial hybrid attention module is as follows: S21. Calculate the channel attention weights and spatial attention weights for the input feature map respectively; The channel attention weights are obtained by performing global average pooling and global max pooling on the feature map in the spatial dimension, and then summing the outputs of the two through a shared network. Spatial attention weights are obtained by performing global average pooling and global max pooling on the feature map along the channel dimension, concatenating the outputs of the two, and then performing a convolution operation. S22. Multiply the calculated channel attention weights and spatial attention weights element-wise to generate a fused attention weight map. Multiply this weight map element-wise with the original input feature map to obtain the weighted output feature map.

[0042] 1) Channel attention mechanism: By performing Global Average Pooling (GAP) and Global Max Pooling (GMP) on the input feature map in the spatial dimension, global statistics for each channel are calculated. GAP focuses on the global mean of each channel, while GMP focuses on the global maximum of each channel. Combining the two can capture the overall importance of the channels.

[0043] For the input feature map (in, H Indicates the height of the feature map. W Indicates width, C (Representing the number of channels), the calculation process for channel attention is as follows: The global average pooling (GAP) is calculated using the formula shown in equation (1): (1)； in, F(i, j, k) Indicates the first k Each channel is in i , j Location feature values; H Indicates the height of the feature map. W Indicates width, C Indicates the number of channels.

[0044] The global max pooling (GMP) is calculated using the formula shown in equation (2): (2)； The results of global average pooling and global max pooling are fused through a fully connected layer and an activation function to obtain the weights for each channel. A ch The calculation method is shown in equation (3): (3)； in, W 1 and W 2 It is the learned weight matrix. σ This represents the sigmoid activation function, used to normalize channel weights.

[0045] 2) Spatial attention mechanism: Global average pooling and global max pooling are performed along the channel dimension to calculate the importance of each spatial location. This allows the model to adjust its focus based on information at each location.

[0046] The formula for calculating the spatial global average pooling is shown in equation (4): (4)； in, F(i, j, k) For the input feature map in i , j The characteristic value of the location.

[0047] The formula for calculating the global max pooling in space is shown in equation (5): (5)； The results from both are fused via convolution to generate spatial attention weights. A sp The calculation method is shown in equation (6): (6)； in, Conv2D It is a 1×1 convolution operation. σ It is the sigmoid activation function.

[0048] 3) Integration of channel and spatial attention: Channel and spatial attention are fused by pixel-wise multiplication, as shown in Equation (7), and applied to the original feature map F to generate a weighted feature map, as shown in Equation (8). (7)； (8)； By employing joint optimization and 1×1 convolution operations, EHCS Attention is able to efficiently capture important features under multi-channel input and complex spatial structures.

[0049] Under conditions of few samples, data augmentation and regularization are important methods to improve the generalization performance of the model. Data augmentation generates more samples by performing various random transformations on the training data, thereby alleviating the overfitting problem. Therefore, in this invention, a variety of data augmentation strategies are adopted, including random rotation, horizontal and vertical translation, scaling, shearing, and horizontal flipping, to increase the model's adaptability to complex scenes in landslide images. In step S2, during the model training phase, a Bayesian optimization algorithm is used to automatically tune the hyperparameter combination, which includes, but is not limited to: random seed, learning rate, L2 regularization coefficient, batch size, data augmentation parameters, and number of training epochs.

[0050] This embodiment incorporates L2 regularization into the loss function. L2 regularization effectively suppresses the infinite growth of weight values ​​in the model by introducing a weight penalty term into the loss function, thereby enhancing the robustness of the model.

[0051] Bayesian optimization fits the relationship between hyperparameters and model validation set performance by constructing a Gaussian process surrogate model, and selects the next hyperparameter point to be evaluated based on the acquisition function, thus efficiently finding the optimal solution.

[0052] Meanwhile, choosing appropriate hyperparameters is also crucial under conditions of few samples. Traditional grid search and random search methods are inefficient due to the complexity of the model's hyperparameter space and the time-consuming evaluation process.

[0053] The mathematical expression of Bayesian optimization in this embodiment is shown in equation (9): (9)； Where θ represents the combination of hyperparameters, and GP is a Gaussian process model.

[0054] Step S3 in this embodiment specifically includes: Step S31: Use the SEEDS superpixel segmentation algorithm to over-segment the preliminary recognition result image to generate multiple superpixel regions; Step S32: Calculate the color mean of adjacent superpixel regions, construct a region adjacency graph, and use a normalization cutting method to segment the graph and merge similar superpixel regions; Step S33: Threshold binarize the merged image and remove noise areas smaller than the preset threshold through connected component analysis, while retaining the main landslide target areas.

[0055] The method described in this paper was used to segment and evaluate six images in the test set, and the results are as follows: Figure 2 As shown.

[0056] The accuracy of the recognition results in this embodiment is compared with the accuracy of recognition results from other different methods, and the results are shown in the appendix. Figure 3 As shown, the Dice coefficient of this embodiment reaches 0.93, which is significantly better than the original U-Net (0.58) and the region growing method (0.21).

[0057] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A landslide image recognition method based on feature decomposition and hybrid channel spatial attention, characterized in that, Collect few-sample landslide image data from different regions and time periods, and label the landslide areas in each image; a portion of the images are used as the training and validation set, and the remaining images are used as the test set; the few-sample conditional landslide identification method includes the following steps: S1: In the preprocessing stage, the image is decomposed into multiple channels to extract feature information including color, space and frequency, so as to enrich the input features of the model and construct a multi-channel input feature map. S2: In the preliminary identification stage, the multi-channel feature map is input into the landslide identification model built on the U-Net architecture for preliminary segmentation and prediction. The model introduces a channel-space hybrid attention mechanism and is trained using data augmentation, L2 regularization and Bayesian optimization strategies. S3: In the post-processing stage of the model, the preliminary recognition results output in step S2 are refined using methods including SEEDS superpixel segmentation, normalized cutting, and connected component analysis to obtain a clear image containing the main target. S4: Verify the high accuracy and robustness of the model of this invention.

2. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention as described in claim 1, characterized in that, Step S1 specifically includes: Extract the RGB color channels of the image as the basic color features; Perform wavelet transform on the image to obtain the low-frequency channel containing the global structure and the high-frequency channel containing detailed information; Perform a Fourier transform on the image and take the amplitude spectrum, then perform a logarithmic transform to obtain the frequency domain features. The RGB color channels, wavelet transform subbands, and Fourier transform amplitude spectrum features are combined to form a multi-channel input image.

3. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 1, characterized in that, The landslide identification model used in step S2 is based on U-Net architecture, and a high-efficiency channel-spatial hybrid attention module is embedded in its encoder and / or decoder modules. The workflow of the high-efficiency channel-spatial hybrid attention module is as follows: S21. Calculate the channel attention weights and spatial attention weights for the input feature map respectively; The channel attention weights are obtained by performing global average pooling and global max pooling on the feature map in the spatial dimension, and then summing the outputs of the two through a shared network. Spatial attention weights are obtained by performing global average pooling and global max pooling on the feature map along the channel dimension, concatenating the outputs of the two, and then performing a convolution operation. S22. Multiply the calculated channel attention weights and spatial attention weights element-wise to generate a fused attention weight map. Multiply this weight map element-wise with the original input feature map to obtain the weighted output feature map.

4. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 3, characterized in that, The steps for calculating channel attention weights include: Global average pooling and global max pooling are performed on the input feature map in the spatial dimension to calculate the global statistics for each channel; For the input feature map, global average pooling is used for calculation, and its formula is shown in equation (1): in, F(i, j, k) Indicates the first k Each channel is in i , j Location feature values; H Indicates the height of the feature map. W Indicates width, C Indicates the number of channels; The global max pooling is used for calculation, and its formula is shown in equation (2): The results of global average pooling and global max pooling are fused through a fully connected layer and an activation function to obtain the weights for each channel. A ch The calculation method is shown in equation (3): in, W 1 and W 2 It is the learned weight matrix. σ This represents the sigmoid activation function, used to normalize channel weights.

5. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 4, characterized in that, The steps for calculating spatial attention weights include: Global average pooling and global max pooling are performed along the channel dimension to calculate the importance of each spatial location, enabling the model to adjust its attention based on the information at each location; The formula for calculating the spatial global average pooling is shown in equation (4): in, F(i, j, k) For the input feature map in i , j Location feature values; The formula for calculating the global max pooling in space is shown in equation (5): The results from both are fused via convolution to generate spatial attention weights. A sp The calculation method is shown in equation (6): in, Conv2D It's a convolution operation. σ It is the sigmoid activation function.

6. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 5, characterized in that, Channel and spatial attention are fused by pixel-wise multiplication, as shown in Equation (7), and applied to the original feature map F to generate a weighted feature map, as shown in Equation (8). By combining optimization and convolution operations, EHCS Attention can efficiently capture important features under multi-channel input and complex spatial structures.

7. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 1, characterized in that, In step S2, during the model training phase, the Bayesian optimization algorithm is used to automatically tune the hyperparameter combination. The hyperparameter combination includes, but is not limited to: random seed, learning rate, L2 regularization coefficient, batch size, data augmentation parameters, and number of training rounds. Bayesian optimization fits the relationship between hyperparameters and model validation set performance by constructing a Gaussian process surrogate model, and selects the next hyperparameter point to be evaluated based on the acquisition function, thus efficiently finding the optimal solution.

8. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 1, characterized in that, The mathematical expression for Bayesian optimization is shown in equation (9): in, θ Indicates a combination of hyperparameters. θ next For the next combination of hyperparameters, GP is a Gaussian process model.

9. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Use the SEEDS superpixel segmentation algorithm to over-segment the preliminary recognition result image to generate multiple superpixel regions; Step S32: Calculate the color mean of adjacent superpixel regions, construct a region adjacency graph, and use a normalization cutting method to segment the graph and merge similar superpixel regions; Step S33: Threshold binarize the merged image and remove noise areas smaller than the preset threshold through connected component analysis, while retaining the main landslide target areas.

10. The landslide image recognition method based on feature decomposition and hybrid channel spatial attention according to claim 1, characterized in that, The landslide image recognition system based on feature decomposition and hybrid channel spatial attention includes: The feature decomposition module is used to perform the operation of step S1, converting a single landslide image into a multi-channel feature map rich in multi-dimensional information; The intelligent recognition module has a built-in landslide recognition model based on the U-Net architecture described in step S2, which is used to receive the multi-channel feature map and output a preliminary landslide area prediction map. The post-processing optimization module is used to perform the operation of step S3, refine the preliminary landslide area prediction map, and output the final identification result; The analysis and verification module is used to perform step S4 to test the remaining images, verifying high accuracy and robustness.

Images