Slope support stability prediction method and system based on remote sensing data

Abstract

The present application relates to the technical field of remote sensing geological prediction, in particular, the present application relates to the slope support stability prediction method and system based on remote sensing data, through obtaining the multispectral remote sensing image of target slope area, the normalized vegetation index is used to generate vegetation mask, combined with morphological filtering and connected region analysis, the interference pixels of vegetation cover area are removed, then the improved edge detection algorithm is used to identify the support structure characteristics, the region growing algorithm based on machine learning is used to demarcate the influence area, the area image is input into the convolution neural network model, the data enhancement technology and weight optimization layer are used to output the deformation probability graph, finally, according to the historical data, the threshold value is determined, combined with the spatial neighborhood information and voting mechanism, the slope support stability category is divided, the vegetation interference is effectively removed, the support structure is accurately identified, the prediction accuracy is improved, the slope support stability can be accurately judged in time, and reliable basis is provided for slope disaster warning and protection.

Description

A Method and System for Predicting Slope Support Stability Based on Remote Sensing Data Technical Field

[0001] This invention relates to the field of remote sensing geological prediction technology, and more specifically, to a method and system for predicting slope support stability based on remote sensing data. Background Technology

[0002] Remote sensing geological prediction technology is an important technology. In traditional slope support stability prediction work, the interference of vegetation cover on the slope surface on the prediction results is quite prominent.

[0003] Traditional methods often employ simple thresholding or single filtering techniques, making it difficult to remove interfering pixels from vegetation-covered areas when processing multispectral remote sensing images. They also fail to consider the complexity and diversity of vegetation distribution. In actual slope environments, vegetation growth varies, spectral characteristics of different vegetation types overlap, and the boundaries between vegetation and slope support structures in images are unclear. This leads to pixels in vegetation areas interfering with the identification of slope support structure features and the assessment of deformation around the support structure, resulting in biased predictions and difficulty in accurately assessing slope support stability. Consequently, this impacts early warning and protection against slope disasters. To address this technical problem, we provide a slope support stability prediction method and system based on remote sensing data. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting slope support stability based on remote sensing data, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, one of the objectives of this invention is to provide a method for predicting slope support stability based on remote sensing data, comprising the following steps:

[0006] S1. Acquire multispectral remote sensing images of the target slope area and generate a vegetation mask using the normalized vegetation index to remove interfering pixels in the vegetation-covered area.

[0007] S2. Based on the preprocessed image, the support structure features of the slope surface are identified by the edge detection algorithm, and the influence area of ​​the support structure is delineated.

[0008] S3. Input the image of the area affected by the support structure into the convolutional neural network model, and add a weight optimization layer for slope texture in the channel attention module to output the deformation probability map of the pixels around the support structure.

[0009] S4. Based on the proportion of areas where the pixel value in the deformation probability map exceeds the set threshold, the slope support stability is divided into a binary classification result of "deformed" and "undeformed".

[0010] As a further improvement to this technical solution, when generating a vegetation mask using the normalized vegetation index (NVI) in step S1, the formula for calculating the NVI is as follows:

[0011] ;

[0012] in, For near-infrared reflectivity, For red light band reflectivity, For blue light band reflectivity, For green light band reflectivity, and This is the correction factor.

[0013] As a further improvement to this technical solution, when removing interfering pixels in the vegetation-covered area in S1, a method combining morphological filtering and connected component analysis is adopted:

[0014] Morphological opening and closing operations are performed on the vegetation mask, and then connected component analysis is used to mark different vegetation connected regions. Pixels corresponding to vegetation connected regions with areas smaller than the set area threshold are removed.

[0015] As a further improvement to this technical solution, in step S2, when identifying the support structure features of the slope surface using an edge detection algorithm, the edge detection algorithm is employed:

[0016] An adaptive threshold adjustment mechanism is added to the edge detection algorithm to automatically adjust the high and low thresholds according to the local contrast of the image, and multi-scale Gaussian filtering is performed on the image before edge detection to obtain edge features;

[0017] When delineating the influence area of ​​the support structure, a machine learning-based region growth algorithm is used. The detected edge pixels of the support structure are used as seed points. Region growth is performed based on the gray value, texture features and spatial adjacency of the pixels. At the same time, spatial constraints are introduced to determine the actual shape and range of the support structure.

[0018] As a further improvement to this technical solution, in S2, when using a machine learning-based region growing algorithm, a random forest classifier is used to classify pixels. When training the random forest classifier, the selected features include the pixel's gray value, gradient magnitude, texture features, and spatial location features. Furthermore, the classifier's parameters are updated in real time during the region growing process.

[0019] As a further improvement to this technical solution, in step S3, when the image of the area affected by the support structure is input into the convolutional neural network model, data augmentation processing is performed on the image:

[0020] When adding a weight optimization layer for slope texture in the channel attention module, the weight optimization layer adopts a structure combining convolutional layers and fully connected layers. First, the local features of the slope texture are extracted through the convolutional layer, and then the local features are mapped to the global weights through the fully connected layer to adjust the weights of the channel attention module.

[0021] As a further improvement to this technical solution, S3 employs a data augmentation method based on generative adversarial networks when performing data augmentation processing on the image:

[0022] Generative adversarial networks are trained to generate synthetic images with different textures and lighting conditions than the original images to expand the training data. In the weight optimization layer, a multi-head attention structure is used to process texture features of different scales and directions separately, and then the different feature representations are fused through the multi-head attention mechanism.

[0023] As a further improvement to this technical solution, the determination of the threshold in S4 adopts a statistical analysis method based on historical data:

[0024] By analyzing historical slope support deformation cases, a relationship model between deformation probability and actual deformation is established. Based on the model, the optimal threshold is determined. Spatial neighborhood information is introduced during classification to check the classification results of adjacent pixels. Based on the check results, it is analyzed whether a secondary judgment is needed, and a voting mechanism is used to determine the final classification result.

[0025] As a further improvement to this technical solution, when establishing the relationship model between the deformation probability and the actual deformation, step S4 adopts a logistic regression model:

[0026] The proportion of regions with pixel values ​​exceeding a set threshold in the deformation probability map is used as the independent variable, and the actual slope support deformation is used as the dependent variable. The model parameters are estimated using the maximum likelihood estimation method. When making a secondary judgment, the voting mechanism adopts the majority voting principle. If the number of "deformed" categories in adjacent pixels exceeds the set proportion, the pixel is classified as "deformed"; otherwise, it is classified as "undeformed".

[0027] The second objective of this invention is to provide a system for running the slope support stability prediction method based on remote sensing data described in any one of the above-mentioned methods, comprising:

[0028] The system includes a multi-source remote sensing data acquisition unit, a vegetation removal and preprocessing unit, a support structure identification and area delineation unit, a convolutional neural network analysis unit, a stability classification and determination unit, and a feedback optimization and control unit.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] Multispectral remote sensing images were acquired and vegetation masks were generated using the normalized vegetation index (NVI). Morphological filtering and connected component analysis were combined to remove vegetation-interfering pixels, effectively improving data quality and avoiding interference from vegetation factors in prediction. Improved edge detection algorithms and machine learning-based region growing algorithms were used to identify slope support structure features and delineate the affected area, providing accurate ranges for subsequent analysis. A weight optimization layer for slope texture was added to the convolutional neural network model, and combined with a generative adversarial network-based data augmentation method, the prediction accuracy of deformation probability of pixels around the support structure was significantly improved. Thresholds were determined based on historical data statistical analysis, and spatial neighborhood information and voting mechanisms were introduced for classification, making the stability classification results more reliable, so as to more accurately assess the stability of slope support and provide strong support for slope disaster early warning and protection. Attached Figure Description

[0031] Figure 1 is a flowchart of the overall workflow of the present invention;

[0032] Figure 2 is a panoramic view of the process of this invention;

[0033] The meanings of the labels in the diagram are as follows:

[0034] 1. Multi-source remote sensing data acquisition unit; 2. Vegetation removal preprocessing unit; 3. Support structure identification and area delineation unit; 4. Convolutional neural network analysis unit; 5. Stability classification and judgment unit; 6. Feedback optimization and control unit. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please refer to Figures 1 and 2. One of the objectives of this embodiment is to provide a method for predicting slope support stability based on remote sensing data, including the following steps:

[0037] S1. Acquire multispectral remote sensing images of the target slope area and generate a vegetation mask using the normalized vegetation index to remove interfering pixels in the vegetation-covered area.

[0038] By receiving satellite remote sensing image data and leveraging the wide-area coverage of satellites, large-scale information on the geographical environment and surrounding topography of the slope is obtained. When generating a vegetation mask using the normalized vegetation index (NVI), S1 introduces an atmospheric correction factor. This factor is used to correct reflectivity deviations caused by light transmission in the atmosphere by conducting on-site measurements of atmospheric parameters in the acquisition area or estimating them based on meteorological data. Considering the influence of soil background, the correction coefficient is determined based on the soil type and color characteristics of different areas through pre-analysis of soil spectral samples. and Then the normalized vegetation index The calculation formula is:

[0039] ;

[0040] in, For near-infrared reflectivity, For red light band reflectivity, For blue light band reflectivity, For green light band reflectivity, and To correct the coefficients, the Normalized Difference Vegetation Index (NDVI) is calculated, which effectively highlights vegetation information. Since vegetation has high reflectivity in the near-infrared band and low reflectivity in the red band, the NDVI value calculated using this formula can quantitatively reflect the growth status and coverage information of vegetation. In this invention, a vegetation mask is generated based on NDVI, thereby accurately locating vegetation-covered areas. This facilitates the subsequent removal of pixels in these areas, reducing the interference of vegetation on the identification of slope support structure features and stability prediction. The vegetation mask is a binary image generated based on image processing technology. In the process of this invention, the results of the Normalized Difference Vegetation Index calculation are used to distinguish between vegetation-covered areas and non-vegetated areas in the image. In the mask image, vegetation-covered areas are usually represented by one color, and non-vegetated areas are represented by another color, as if a "film" is covered over the vegetation areas in the original image, hence the name vegetation mask. This "film" can intuitively display the vegetation distribution, providing a basis for the subsequent accurate removal of interfering pixels in vegetation-covered areas. It is a key step in improving the accuracy of slope support stability prediction.

[0041] When removing interfering pixels from vegetation-covered areas in S1, a combination of morphological filtering and connected component analysis is used:

[0042] Vegetation masks may contain isolated small noise points. These noise points can be removed using morphological opening operations. Morphological opening and closing operations are performed on the vegetation mask. A structuring element is selected, and an erosion operation is performed on the vegetation mask. This involves sliding the structuring element across the vegetation mask; if the structuring element is completely contained within the vegetation area, the point is retained; otherwise, it is set as background. A dilation operation is then performed on the eroded image. The structuring element is then slid across the eroded image; when the structuring element intersects with the image, the area covered by the structuring element is set as foreground. After the opening operation... Small holes may appear in vegetated areas, which can affect the judgment of the integrity of connected vegetation regions. Morphological closing operations can fill these holes. Again, a structuring element is selected, and a dilation operation is first performed on the vegetation mask after the opening operation. The structuring element is then slid across the image. When the structuring element intersects with the image, the area covered by the structuring element is set as the foreground. An erosion operation is then performed on the dilated image, i.e., the structuring element is slid across the dilated image. When the structuring element is completely contained within the vegetation region, that point is retained; otherwise, it is set as the background. The connected regions are more complete, facilitating accurate subsequent labeling and analysis. Through connected region analysis, different vegetation connected regions are labeled. Starting from the top left corner of the image, the image is scanned line by line. When an unlabeled foreground pixel is encountered, it is used as the starting point of a new connected region. Then, the selected connectivity algorithm is used to label all pixels within that connected region, assigning them the same label value. The image is scanned again until all foreground pixels are labeled, thus obtaining different vegetation connected regions and their corresponding labels. This facilitates independent analysis and processing of each vegetation connected region. Based on a set area threshold, pixels corresponding to vegetation connected regions with areas smaller than the threshold are removed. The area of ​​each labeled vegetation connected region is counted, which can be obtained by calculating the number of pixels within that region. An area threshold is set, and all labeled vegetation connected regions are traversed. Pixels corresponding to connected regions with areas smaller than the threshold are set as background pixels, i.e., these regions are removed from the vegetation mask. This makes the final image used for analysis more focused on areas that have a significant impact on slope support stability, improving the efficiency and quality of the analysis.

[0043] S2. Based on the preprocessed image, the support structure features of the slope surface are identified by the edge detection algorithm, and the influence area of ​​the support structure is delineated.

[0044] In S2, when identifying the support structure features of the slope surface using an edge detection algorithm, the edge detection algorithm is used as follows:

[0045] An adaptive threshold adjustment mechanism is added to the edge detection algorithm. The high and low thresholds are automatically adjusted according to the local contrast of the image. Before edge detection, the image is subjected to multi-scale Gaussian filtering. Multiple different Gaussian kernel sizes are selected, and the pre-processed image is convolved with Gaussian kernels of different scales to obtain a smoothed image. The smoothed images of different scales are weighted and fused to obtain the final image after multi-scale Gaussian filtering. This enables the subsequent edge detection algorithm to more accurately identify the true edges of the slope surface support structure. The image after multi-scale Gaussian filtering is divided into multiple local regions. The size of each local region is adjusted according to the specific situation of the image. For each local region, its local contrast index is calculated. Based on the local contrast index, the high and low thresholds for edge detection in that local region are dynamically adjusted. The adjusted high and low thresholds are used to perform edge detection in each local region. Even in images with large contrast differences, the edge features of the slope surface support structure can be completely and accurately identified.

[0046] When delineating the influence area of ​​the support structure, a machine learning-based region growth algorithm is used. The detected edge pixels of the support structure are used as seed points. Region growth is performed based on the gray value, texture features and spatial adjacency of the pixels. At the same time, spatial constraints are introduced to determine the actual shape and range of the support structure.

[0047] Edge pixels of the support structure detected by the edge detection algorithm are used as seed points. These edge pixels are the boundary points of the support structure. Starting region growth from these points can more accurately determine the range of the support structure. Features such as gray value, texture features, and spatial adjacency of each pixel are extracted. The machine learning model is trained using labeled sample data. During training, the extracted features are used as input and the sample category is used as output. The model parameters are adjusted to minimize the classification error. Starting from the seed point, its neighboring pixels are checked. For each neighboring pixel, its features are extracted and input into the trained machine learning model for classification. If the neighboring pixel is determined to belong to the support structure region, it is added to the current growth region. The process continues to check its neighboring pixels with that pixel as the center until there are no neighboring pixels that meet the conditions. During the region growth process, spatial constraints are introduced to ensure that the grown region conforms to the actual shape and range of the support structure. The resulting support structure influence area is more consistent with the actual situation, providing a more accurate basis for subsequent analysis and evaluation of the support structure.

[0048] In S2, when using a machine learning-based region growing algorithm, a random forest classifier is used to classify pixels. When training the random forest classifier, the selected features include the pixel's gray value, gradient magnitude, texture features, and spatial location features. Furthermore, the classifier's parameters are updated in real time during the region growing process.

[0049] For each pixel in the image, its grayscale value is directly obtained. The grayscale value reflects the brightness of the pixel. Different materials of the support structure and the surrounding environment usually have different grayscale values. For example, the grayscale of the concrete support is relatively uniform and high, while the grayscale of the soil area is relatively low and varies more. The gradient magnitude of the pixel is calculated using the Sobel operator, and the texture features are obtained using the gray-level co-occurrence matrix. The coordinate position of the pixel in the image is recorded and normalized to obtain the relative position information. Pixels near the top, bottom, or edge of the slope, combined with other features, help to determine whether they belong to the area affected by the support structure, because the distribution of the support structure is often related to the slope geometry. A large number of image samples containing known support structure areas and non-support structure areas are collected. Features are extracted from the pixels in each sample image using the above method, and the pixel category is labeled, i.e., support structure or non-support structure, to form a training dataset. The parameters of the number of decision trees, the maximum depth of the trees, and the minimum number of samples required for node splitting in the random forest are determined. Through multiple experiments and cross-validation, the optimal parameter combination is selected to avoid overfitting and ensure classification performance. The training dataset is input into the random forest model. A bootstrap sampling method is used, where a subset of samples is drawn with replacement each time to construct a decision tree. During node splitting, the optimal feature is selected based on metrics such as information gain and Gini impurity, allowing the decision tree to grow continuously. Ultimately, a random forest model capable of classifying pixels is trained. After each round of region growth (expanding a certain range outward from the seed point), the feature statistics of the grown region are checked, such as the mean and variance of grayscale values ​​and the range of texture feature changes. When the changes in these statistics exceed a preset threshold, the classifier parameters are updated. The pixel features and corresponding categories of the newly included grown region are used as incremental data. An online learning algorithm is used to retrain a portion of the random forest classifier's decision tree, adjusting the tree node parameters to adapt the classifier to the new data distribution. After each update, the updated classifier's performance is evaluated using validation samples not used in training, calculating accuracy and recall. If the metrics do not meet expectations, the update strategy is further optimized or model parameters are adjusted to ensure the classifier maintains high classification accuracy, providing a reliable foundation for subsequent stability analysis.

[0050] S3. Input the image of the area affected by the support structure into the convolutional neural network model, and add a weight optimization layer for slope texture in the channel attention module to output the deformation probability map of the pixels around the support structure.

[0051] In S3, when the image of the area affected by the support structure is input into the convolutional neural network model, data augmentation processing is performed on the image:

[0052] When adding a weight optimization layer for slope texture in the channel attention module, the weight optimization layer adopts a structure combining convolutional layers and fully connected layers. First, the local features of the slope texture are extracted through the convolutional layer, and then the local features are mapped to the global weights through the fully connected layer to adjust the weights of the channel attention module.

[0053] The image is randomly rotated within a certain angle range, using the image center as the rotation point. The new position of each pixel after rotation is calculated using trigonometric functions, and bilinear interpolation and other methods are used to fill any blank areas that may appear after rotation, ensuring the continuity and integrity of the image. This includes horizontal and vertical flipping. Horizontal flipping reverses the pixel order of each row, while vertical flipping reverses the pixel order of each column. This operation is simple and efficient, quickly increasing data diversity. The image is randomly scaled within a certain scaling range, and interpolation is also performed during scaling to ensure image sharpness and quality. The image brightness is randomly adjusted by multiplying by a random coefficient to change the brightness value of each pixel. Slope images under different lighting conditions are simulated. Convolutional kernels are designed to perform convolution operations on the input feature map. The convolutional kernel slides on the feature map to extract local texture features. Multiple convolutional kernels can be set to extract different types of texture features. During convolution, activation functions are used to introduce non-linearity and enhance the model's expressive power. After the convolutional layer, pooling is performed to reduce the dimensionality of the feature map, decrease computation, and retain important texture features. Pooling divides the feature map into multiple non-overlapping regions, taking the maximum or average value of each region as its output. The feature map after convolution and pooling is flattened into a one-dimensional vector and input into a fully connected layer. The neurons in the fully connected layer are connected to each element of the input vector. By learning the weight matrix, local texture features are mapped to global weights. The output of the fully connected layer is a vector with the same number of channels as the channel attention module, with each element corresponding to a channel weight adjustment value. The weight adjustment value of the fully connected layer output is multiplied by the original weights of the channel attention module to obtain the adjusted channel weights. The adjusted weights are used to weight the output of the channel attention module, highlighting channel features related to slope texture. The deformation probability map of the pixels surrounding the support structure is more accurate and can more clearly reflect the slope deformation, providing a more reliable basis for slope support stability prediction.

[0054] When performing image data augmentation, S3 employs a generative adversarial network-based data augmentation method:

[0055] Generative adversarial networks are trained to generate synthetic images with different textures and lighting conditions than the original images to expand the training data. In the weight optimization layer, a multi-head attention structure is used in the attention mechanism to process texture features of different scales and directions separately, and then the different feature representations are fused through the multi-head attention mechanism.

[0056] Design a deep convolutional neural network as a generator, whose input is a random noise vector. The network structure can use deconvolutional layers to gradually convert the low-dimensional noise vector into a high-resolution image, gradually increasing the size of the feature map, and finally outputting a synthetic image of the same size as the original image. Construct a convolutional neural network as a discriminator to determine whether the input image is the real original image or a synthetic image generated by the generator. The discriminator typically consists of multiple convolutional layers, batch normalization layers, and activation functions. Finally, a fully connected layer outputs a probability value representing the probability that the input image is a real image. This allows the convolutional neural network model to learn more slope features under different textures and lighting conditions, improving the model's generalization ability. The original image of the area affected by the support structure is used as real data input to the discriminator. The generator and discriminator are trained alternately. For the discriminator, real images and synthesized images generated by the generator are input, and the discriminator's loss function is calculated. Backpropagation updates the discriminator's parameters, enabling it to better distinguish between real and synthesized images. For the generator, a random noise vector is input to generate synthesized images, and the generator's loss function is calculated. The goal is to make the discriminator misclassify the synthesized images as real images. Backpropagation updates the generator's parameters, and training is iterative until the synthesized images generated by the generator can fool the discriminator with a high probability, reducing the risk of model overfitting. After extracting local features of the slope texture in the convolutional layers, the feature maps are input into a multi-head attention mechanism. The feature map is divided into multiple heads, each of which independently performs a linear transformation on the feature map, resulting in three matrices: query, key, and value. For each head, the similarity score between the query and key is calculated, and then the score is normalized using the softmax function to obtain the attention weights. The attention weights are multiplied by the value matrix to obtain the weighted feature representation of each head. The weighted feature representations of each head are concatenated and then linearly transformed through a fully connected layer to obtain the fused feature representation. This allows the model to more accurately focus on the texture features related to slope deformation, improving the accuracy of the output deformation probability map of pixels surrounding the support structure. The feature representation fused by the multi-head attention mechanism is input into a fully connected layer to obtain a vector with the same number of channels as the channel attention module. Each element corresponds to a channel weight adjustment value. This weight adjustment value is multiplied by the original weights of the channel attention module to obtain the adjusted channel weights, which are used to weight the output of the channel attention module. The output deformation probability map can more accurately reflect the slope deformation, providing a more reliable basis for slope support stability prediction.

[0057] S4. Based on the proportion of areas where the pixel value in the deformation probability map exceeds the set threshold, the slope support stability is divided into a binary classification result of "deformed" and "undeformed".

[0058] The threshold for S4 is determined using a statistical analysis method based on historical data:

[0059] By analyzing historical slope support deformation cases, a relationship model between deformation probability and actual deformation is established. The optimal threshold is determined based on the model. Spatial neighborhood information is introduced during classification to check the classification results of adjacent pixels. Based on the check results, it is analyzed whether a secondary judgment is needed, and a voting mechanism is used to determine the final classification result.

[0060] Historical data on slope support from multiple regions, geological conditions, and support structure types were collected, including deformation probability map data and corresponding actual deformation records. The deformation probability value of the pixels in the deformation probability map was used as the independent variable, and the actual deformation was used as the dependent variable. The historical data was used to train the logistic regression model. During training, methods such as maximum likelihood estimation are used to estimate the model parameters so that the model can best fit the historical data. The trained logistic regression model is validated using a validation dataset. Evaluation metrics are calculated under different thresholds. The threshold corresponding to the point closest to the top left corner of the curve is found by plotting the ROC curve, and this threshold is the optimal threshold. For each pixel in the deformation probability map, its neighborhood range is defined, the classification results of adjacent pixels in the neighborhood are checked, the number of pixels with different classification results in the neighborhood is counted, and a difference threshold is set. When the difference in the number of pixels with different classification results in the neighborhood exceeds the threshold, a secondary judgment is triggered to avoid overall misjudgment caused by the abnormal classification result of a single pixel, thus improving the reliability of classification. After triggering the secondary judgment, the classification results of all pixels in the neighborhood are voted on. If the number of pixels in the "deformed" category exceeds half, the pixel is finally classified as "deformed"; otherwise, it is classified as "undeformed".

[0061] S4 uses a logistic regression model to establish the relationship between deformation probability and actual deformation.

[0062] The proportion of regions with pixel values ​​exceeding a set threshold in the deformation probability map is used as the independent variable, and the actual slope support deformation is used as the dependent variable. The model parameters are estimated using the maximum likelihood estimation method. When making a secondary judgment, the voting mechanism adopts the majority voting principle. If the number of "deformed" categories in adjacent pixels exceeds the set proportion, the pixel is classified as "deformed"; otherwise, it is classified as "undeformed".

[0063] The second objective of this invention is to provide a system for implementing the slope support stability prediction method based on remote sensing data as described in any one of the above-mentioned methods, comprising:

[0064] The multi-source remote sensing data acquisition unit 1 integrates multiple devices to acquire multispectral remote sensing images, automatically acquiring data based on conditions to ensure data quality.

[0065] The vegetation removal preprocessing unit 2 uses an optimization algorithm to generate a vegetation mask, and combines filtering and analysis to remove vegetation interference pixels.

[0066] The improved algorithm of the support structure identification and area delineation unit 3 identifies the characteristics of the support structure and uses machine learning algorithms to delineate the affected area;

[0067] The convolutional neural network analysis unit 4 inputs the enhanced data into the network and optimizes it through the channel attention module, outputting a deformation probability map;

[0068] Stability classification determination unit 5 classifies according to statistical thresholds and outputs classification results by combining neighborhood information and voting mechanism;

[0069] The feedback optimization control unit 6 compares the predicted and measured results and adjusts each parameter to improve the accuracy of the prediction.

[0070] This invention acquires multispectral remote sensing images of a target slope area, generates a vegetation mask using the normalized vegetation index, and removes interfering pixels in vegetation-covered areas by combining morphological filtering and connected component analysis. Then, it identifies support structure features using an improved edge detection algorithm, delineates the affected area using a machine learning-based region growing algorithm, inputs the image of this area into a convolutional neural network model, and outputs a deformation probability map using data augmentation techniques and weight optimization layers. Finally, it determines thresholds based on historical data, and classifies slope support stability categories by combining spatial neighborhood information and a voting mechanism. This effectively removes vegetation interference, accurately identifies support structures, improves prediction accuracy, and enables timely and accurate assessment of slope support stability, providing a reliable basis for slope disaster early warning and protection.

[0071] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting slope support stability based on remote sensing data, characterized in that: Includes the following steps: S1. Acquire multispectral remote sensing images of the target slope area and generate a vegetation mask using the normalized vegetation index (NVI) to remove interfering pixels in the vegetation-covered areas; S2. Based on the preprocessed images, identify the support structure features of the slope surface using an edge detection algorithm and delineate the support structure influence area; S3. Input the image of the support structure influence area into a convolutional neural network model, and add a weight optimization layer for slope texture in the channel attention module to output a deformation probability map of pixels surrounding the support structure; S4. Based on the proportion of areas where pixel values ​​in the deformation probability map exceed a set threshold, classify the slope support stability into a binary classification result of "deformed" and "undeformed"; In S1, when generating the vegetation mask using the NVI, the calculation formula for the NVI is: ;in, For near-infrared reflectivity, For red light band reflectivity, For blue light band reflectivity, For green light band reflectivity, and The correction coefficient is used. In step S2, when identifying the support structure features of the slope surface using the edge detection algorithm, the edge detection algorithm is used: an adaptive threshold adjustment mechanism is added to the edge detection algorithm, which automatically adjusts the high and low thresholds according to the local contrast of the image, and performs multi-scale Gaussian filtering on the image before edge detection to obtain the processed image. When delineating the influence area of ​​the support structure, a machine learning-based region growing algorithm is used. The detected support structure edge pixels are used as seed points, and region growing is performed according to the gray value, texture features, and spatial adjacency of the processed image pixels. At the same time, spatial constraints are introduced to determine the actual shape and range of the support structure. In step S4, the threshold is determined using a statistical analysis method based on historical data: a relationship model between deformation probability and actual deformation is established by analyzing historical slope support deformation cases. The optimal threshold is determined according to the model. Spatial neighborhood information is introduced when classifying, and the classification results of adjacent pixels are checked. The results of the checks are analyzed to determine whether a secondary judgment is needed, and a voting mechanism is used to determine the final classification result.

2. The slope support stability prediction method based on remote sensing data according to claim 1, characterized in that, In step S1, when removing interfering pixels in vegetation-covered areas, a combination of morphological filtering and connected component analysis is used: morphological opening and closing operations are performed on the vegetation mask, and then connected component analysis is used to mark different vegetation connected regions. Pixels corresponding to vegetation connected regions with areas smaller than the set area threshold are removed according to the set area threshold.

3. The slope support stability prediction method based on remote sensing data according to claim 1, characterized in that, In S2, when using a machine learning-based region growing algorithm, a random forest classifier is used to classify pixels. When training the random forest classifier, the selected features include the pixel's gray value, gradient magnitude, texture features, and spatial location features. Furthermore, the classifier's parameters are updated in real time during the region growing process.

4. The slope support stability prediction method based on remote sensing data according to claim 1, characterized in that, In S3, when the image of the area affected by the support structure is input into the convolutional neural network model, data augmentation processing is performed on the image: when a weight optimization layer for slope texture is added to the channel attention module, the weight optimization layer adopts a structure combining convolutional layer and fully connected layer. First, the local features of slope texture are extracted through the convolutional layer, and then the local features are mapped to global weights through the fully connected layer to adjust the weights of the channel attention module.

5. The slope support stability prediction method based on remote sensing data according to claim 4, characterized in that, When performing data augmentation on the image, S3 adopts a data augmentation method based on generative adversarial networks: by training the generative adversarial network to generate a synthetic image with different texture and lighting conditions from the original image, the training data is expanded. In the weight optimization layer, a multi-head attention structure in the attention mechanism is used to process the texture features of different scales and directions separately, and then the different feature representations are fused through the multi-head attention mechanism.

6. The slope support stability prediction method based on remote sensing data according to claim 1, characterized in that, When establishing the relationship model between deformation probability and actual deformation, S4 adopts a logistic regression model: the proportion of areas with pixel values ​​exceeding a set threshold in the deformation probability map is used as the independent variable, and the actual slope support deformation is used as the dependent variable. The model parameters are estimated by the maximum likelihood estimation method. When making a secondary judgment, the voting mechanism adopts the majority voting principle. If the number of "deformed" categories in adjacent pixels exceeds the set proportion, the pixel is classified as "deformed"; otherwise, it is classified as "undeformed".

7. A system for operating the slope support stability prediction method based on remote sensing data according to any one of claims 1-6, characterized in that, It includes a multi-source remote sensing data acquisition unit (1), a vegetation removal preprocessing unit (2), a support structure identification and area delineation unit (3), a convolutional neural network analysis unit (4), a stability classification and determination unit (5), and a feedback optimization and control unit (6).

Images