An intelligent landslide surface monitoring method based on YADMW image target recognition

Abstract

The present invention relates to the field of geological environmental disaster monitoring, and specifically to a landslide surface intelligent monitoring method based on YADMW image target recognition, the method comprising: using a camera to collect landslide surface image data in multiple angles and multiple time periods, including landslide surface image data in different time periods, different weather conditions and different geological types; making the collected landslide surface image data into a landslide surface image data set; constructing a YADMW image target recognition model, using the landslide surface image data set to train the YADMW image target recognition model to obtain a final YADMW image target recognition model, and using the final YADMW image target recognition model to monitor the landslide surface image data collected in real time. The present invention improves the accuracy of landslide surface anomaly detection.

Description

Technical Field

[0001] The invention relates to the field of geological environmental disaster monitoring, and in particular to an intelligent landslide surface monitoring method based on YADMW image target recognition. Background Art

[0002] Landslides, as a common geological disaster, frequently occur in mountainous and hilly areas, seriously threatening human life and property safety. Traditional landslide monitoring methods rely on manual observation, on-site surveys and simple sensor networks. These methods often have delays and limitations, and cannot timely and accurately capture key changes before landslides occur. Therefore, the use of advanced intelligent detection technology, especially the landslide surface monitoring method combined with target recognition algorithm, can detect signs of landslides in the early stage, improve early warning capabilities, and reduce losses caused by disasters, which has important practical significance. The harm of landslides to society is extensive and far-reaching. Landslides not only cause landslides and road blockages, but may also trigger secondary disasters such as mudslides and floods, further expanding the scope of disaster impact. The collapse of houses, casualties and damage to infrastructure caused by landslides pose a huge threat to social and economic development and human life. Especially in densely populated areas, landslides may cause thousands of people to be displaced and their living conditions to deteriorate. Therefore, advanced technical means are urgently needed to monitor and warn landslides in real time so that effective emergency measures can be taken to reduce the harm of landslides to society.

[0003] Traditional landslide surface detection methods require human intervention, have high requirements for non-professional technicians, and have problems with the accuracy and delay of the detection method, making it impossible to accurately detect landslide signals. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an intelligent landslide surface monitoring method based on YADMW image target recognition, which solves the problem that traditional landslide surface detection methods require manual intervention and cannot accurately detect landslide signals.

[0005] The present invention is achieved in this way.

[0006] A landslide surface intelligent monitoring method based on YADMW image target recognition, the method comprising:

[0007] Use cameras to collect landslide surface image data at different angles, different time periods, different weather conditions and different geological types;

[0008] The collected landslide surface image data are made into a landslide surface image data set;

[0009] Construct a YADMW image target recognition model, use the landslide surface image data set to train the YADMW image target recognition model to obtain the final YADMW image target recognition model, and use the final YADMW image target recognition model to monitor the landslide surface image data collected in real time;

[0010] The YADMW image target recognition model includes: a first MMQANext module connected to the data input terminal through a two-layer Conv module, and output to the first CBFuse module through an eighth Conv module after performing convolution operation in the first MMQANext module and reducing the calculation complexity and redundant parameters of the attention matrix;

[0011] The data input end also continuously sets up a combination of multiple groups of Conv modules and DGConvNet modules after a layer of Conv modules, and outputs to the SPPELAN module through the last combined DGConvNet module. Except for the first DGConvNet module, each of the remaining combined DGConvNet modules outputs data to the CBLinear module, and multiple CBLinear modules output to the first CBFuse module; the first CBFuse module fuses the outputs of multiple CBLinear modules and the output of the eighth Conv module;

[0012] The output data of the first CBFuse module is respectively output to the first Conv-reg module and the first Conv-ds module through the second MMQANext module. The first Conv-reg module and the first Conv-ds module are output to the fifth Concat module and then output a detection result;

[0013] The SPPELAN module is sequentially output through the first Upsample module, the first Concat module, the fifth MMQANext module, the second Upsample module and the second Concat module, and then output to the fourth Conv-reg module and the fourth Conv-ds module after the sixth MMQANext module. The fourth Conv-reg module and the fourth Conv-ds module are output to the eighth Concat module and then output a detection result.

[0014] Furthermore, the second MMQANext module outputs data to the third MMQANext module after passing through the ninth Conv module and the second CBFuse module in sequence, and the output of the third MMQANext module is output to the sixth Concat module via the second Conv-reg module and the second Conv-ds module respectively, and then outputs a detection result;

[0015] The third MMQANext module outputs data to the fourth MMQANext module after passing through the tenth Conv module and the third CBFuse module in sequence. The output of the fourth MMQANext module is output to the seventh Concat module via the third Conv-reg module and the third Conv-ds module, and then outputs a detection result.

[0016] Further, the sixth MMQANext module is connected to the eleventh Conv module, the third Concat module and the seventh MMQANext module in sequence. The third Concat module also receives the output data of the fifth MMQANext module. The output of the seventh MMQANext module is output to the ninth Concat module via the fifth Conv-reg module and the fifth Conv-ds module, and then outputs a detection result.

[0017] Furthermore, the output of the seventh MMQANext module is connected to the twelfth Conv module, the fourth Concat module and the eighth MMQANext module in sequence. The fourth Concat module adds the output of the SPPELAN module and the output of the twelfth Conv module and then transmits them to the eighth MMQANext module. The output of the eighth MMQANext module is output to the tenth Concat module via the sixth Conv-reg module and the sixth Conv-ds module respectively, and then outputs a detection result.

[0018] Furthermore, the collected landslide surface image data are made into a landslide surface image data set, including: target labeling of the landslide surface image; data enhancement operation of the labeled landslide surface image; and target frame clustering using the AKS clustering algorithm.

[0019] Furthermore, the target frame clustering using the AKS clustering algorithm includes:

[0020] An adaptive learning mechanism is used in the AKS clustering algorithm, so that each centroid has an independent learning rate. The learning rate of the landslide surface image dataset is dynamically adjusted according to the dispersion or concentration of the data points in the cluster.

[0021] Dynamic cluster structure adjustment, when the cluster The number of data points in Exceeding the set threshold , and the intra-cluster variance When the cluster exceeds a certain set value, the cluster splitting operation is performed; and Cluster The distance between the centroids Less than the set threshold , and when the variance within the cluster is similar, clusters are merged. Representation Cluster The centroid position of Representation Cluster The centroid position of

[0022] For each data point , according to the data points The distance to the centroid is assigned a weight , used to control the data points Contribution to centroid update;

[0023] The centroids within a cluster are fine-tuned locally, and the distribution of centroids is re-evaluated globally to ensure balance between clusters.

[0024] Furthermore, the WEF Loss loss function adopted by the YADMW image target recognition model is:

[0025] ,

[0026] is the true label, is the predicted value, represents the edge detection operator, and are the edges of the predicted image and the real image, respectively. represents the landslide feature extractor, , , is the weight coefficient of loss, represents the WEFLoss loss function, Represents the index, which is used to traverse each sample in the data set. Represents the total number of samples in the dataset.

[0027] Furthermore, the method for processing data by the DGConvNet module includes:

[0028] Data is input in the form of feature maps and is divided into two paths;

[0029] One way is to use the first point-by-point convolution layer to perform point-by-point convolution operation using a 7x7 convolution kernel; the other way is to use the first depth convolution layer to perform depth convolution operation using a 7x7 convolution kernel;

[0030] The outputs after two-way convolution are layer normalized;

[0031] The feature map after layer normalization is divided into two paths again. One path passes through the second point-by-point convolution layer through 1x1 point-by-point convolution to expand the number of channels; the other path passes through the second depth convolution layer through 1x1 point-by-point depth convolution to expand the number of channels;

[0032] The feature maps of the two expanded channels are processed by the standard convolutional layer after using the GELU activation function and global response normalization;

[0033] The standard convolutional layer compresses the number of channels back to the number of channels at the output through 1x1 convolution.

[0034] Furthermore, in the MMQANext module, the input data enters the first local convolution layer and the second local convolution layer after hierarchical sharing and factor decomposition, with the number of channels being 96. After passing through the first local convolution layer, the data enters the third local convolution layer and the fourth local convolution layer after hierarchical sharing and factor decomposition. The number of channels becomes 288 through channel expansion. The third local convolution layer and the fourth local convolution layer are layer-normalized after channel fusion, with the number of channels being 480. The fifth local convolution layer is layer-normalized after channel fusion with the first local convolution layer and the second local convolution layer after parameter sharing. After being input into the MQA module, the data is channel-fused with the second local convolution layer after parameter sharing, with the number of channels being 288. Finally, the data is output through the sixth local convolution layer, with the number of channels being 96.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows: the method of the present invention significantly improves the detection accuracy of the landslide area, the advanced feature extraction capability enhances the adaptability of the algorithm detection system in complex environments, the real-time processing technology ensures rapid response capability, and the intelligent monitoring process reduces human intervention, thereby improving the automation level and monitoring efficiency of the landslide surface intelligent monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0037] Figure 2 A structural block diagram of the YADMW image target recognition model provided by an embodiment of the present invention;

[0038] Figure 3 A structural framework diagram of a DGConvNet module provided in an embodiment of the present invention;

[0039] Figure 4 This is a structural board framework diagram of the MMQANext module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] See also Figure 1 Combination Figure 2 As shown,

[0042] A landslide surface intelligent monitoring method based on YADMW image target recognition, the method comprising:

[0043] Use high-resolution cameras to collect images of the landslide surface from multiple angles and at multiple time periods. The collected images include landslide surface image data under different time periods, different weather conditions and different geological types;

[0044] The collected landslide surface image data are made into a landslide surface image data set;

[0045] Construct a YADMW image target recognition model, use the landslide surface image data set to train the YADMW image target recognition model to obtain the final YADMW image target recognition model, and use the final YADMW image target recognition model to monitor the landslide surface image data collected in real time;

[0046] Among them, see Figure 2 As shown, the YADMW image target recognition model includes: a first MMQANext module connected to the data input terminal through a two-layer Conv module, and output to the first CBFuse module through an eighth Conv module after the convolution operation of the first MMQANext module and the reduction of the attention matrix calculation complexity and redundant parameters;

[0047] The data input end also continuously sets up a combination of multiple groups of Conv modules and DGConvNet modules after a layer of Conv modules, and outputs to the SPPELAN module through the last combined DGConvNet module. Except for the first DGConvNet module, each of the remaining combined DGConvNet modules outputs data to the CBLinear module, and multiple CBLinear modules output to the first CBFuse module; the first CBFuse module fuses the outputs of multiple CBLinear modules and the output of the eighth Conv module;

[0048] The output data of the first CBFuse module is respectively output to the first Conv-reg module and the first Conv-ds module through the second MMQANext module. The first Conv-reg module and the first Conv-ds module are output to the fifth Concat module and then output a detection result P1;

[0049] The SPPELAN module is sequentially output through the first Upsample module, the first Concat module, the fifth MMQANext module, the second Upsample module and the second Concat module, and then output to the fourth Conv-reg module and the fourth Conv-ds module after the sixth MMQANext module. The fourth Conv-reg module and the fourth Conv-ds module are output to the eighth Concat module and then output a detection result P4.

[0050] Furthermore, the second MMQANext module sequentially passes through the ninth Conv module and the second CBFuse module, and then outputs data to the third MMQANext module. The output of the third MMQANext module is respectively output to the sixth Concat module via the second Conv-reg module and the second Conv-ds module, and then outputs a detection result P2;

[0051] The third MMQANext module outputs data to the fourth MMQANext module after passing through the tenth Conv module and the third CBFuse module in sequence. The output of the fourth MMQANext module is output to the seventh Concat module via the third Conv-reg module and the third Conv-ds module, and then outputs a detection result P3.

[0052] The sixth MMQANext module is connected to the eleventh Conv module, the third Concat module and the seventh MMQANext module in sequence. The third Concat module also receives the output data of the fifth MMQANext module. The output of the seventh MMQANext module is output to the ninth Concat module via the fifth Conv-reg module and the fifth Conv-ds module, and then outputs a detection result P5.

[0053] The output of the seventh MMQANext module is connected to the twelfth Conv module, the fourth Concat module and the eighth MMQANext module in sequence. The fourth Concat module adds the output of the SPPELAN module and the output of the twelfth Conv module to the eighth MMQANext module. The output of the eighth MMQANext module is respectively output to the tenth Concat module via the sixth Conv-reg module and the sixth Conv-ds module, and then outputs a detection result P6. The outputs of the six detection results jointly provide feature maps of different scales, so that the method of the present invention is more accurate in detecting landslide surfaces of different sizes. On the feature map of each scale, multiple bounding box information in each cell will be predicted, including confidence and coordinate position information.

[0054] In the above modules, Conv represents the convolution operation to extract local features of the input image;

[0055] CBFuse represents the basic network convolution structure component, which is used for feature fusion;

[0056] SPPELAN stands for lightweight network;

[0057] CBLinear means enhancing the information flow of the network through reversible connections;

[0058] Conv-reg represents regression convolution, which is used to output the bounding box position and calculate the loss;

[0059] Conv-ds represents classification convolution, which is used to output category probability distribution;

[0060] Upsample represents the upsampling operation;

[0061] Concat means the channel number addition operation;

[0062] Among them, the present invention first establishes a high-resolution landslide surface dataset, obtains images of the landslide area through drones, ground photography, etc., ensures that the image clarity is high enough, and covers various landslide types, terrains and environmental conditions. The collected high-resolution landslide surface images are data labeled, and the high-resolution images are opened using LabelImg image annotation software, and the landslide area boundary box is manually labeled. A corresponding annotation file is generated for each high-resolution image. Here, a label file in YOLO format (txt format) is generated, and each line includes a category ID, the center coordinates of the boundary box, the width, and the height.

[0063] The labeled high-resolution dataset is preprocessed. The first step is to resize the image to a fixed size of 640×640 suitable for the target recognition algorithm. Then, data enhancement operations such as random rotation, scaling, cropping, and flipping are performed to increase the diversity of the data in the established high-resolution dataset. Finally, normalization is performed to normalize the pixel values ​​to between 0 and 1.

[0064] The collected landslide surface image data is made into a landslide surface image data set including:

[0065] After the landslide surface dataset is created, the AKS clustering algorithm (Adaptive Kernel-Selection) is used to cluster the high-resolution dataset annotation boxes. Specifically, it includes:

[0066] (1) Adaptive learning mechanism,

[0067] Traditional clustering directly updates the centroid position in each iteration. The adaptive learning mechanism is introduced in the AKS clustering algorithm, so that each centroid has an independent learning rate, which can dynamically adjust the update step size according to the internal structure of each cluster of the landslide surface dataset. The centroid of the cluster , the update formula is as follows:

[0068] ,

[0069] is the adaptive learning rate, the initial value is fixed, and it is adaptively adjusted with the number of iterations. For the The number of data points in a cluster, Belongs to cluster No. data points. Indicates In the iteration The centroid position of the cluster, Indicates In the iteration The centroid position of the cluster, is the number of iterations.

[0070] According to the dispersion or concentration of the data points in the cluster, the learning rate of the landslide surface dataset is dynamically adjusted. When the data points in the cluster are concentrated, the variance is small, the learning rate is also small, and the centroid is updated slowly; on the contrary, when the data points in the cluster are dispersed, the learning rate is large, which speeds up the update of the centroid. Dynamic adjustment of the learning rate of the landslide surface dataset The calculation formula is as follows:

[0071] ,

[0072] is the variance of the data points within the cluster, indicating the degree of dispersion of the data within the cluster. It is an adjustment factor that controls how fast the learning rate changes.

[0073] (2) Dynamic cluster structure adjustment:

[0074] Dynamic cluster structure adjustment automatically splits or merges clusters according to the internal characteristics of the cluster. The number of data points in Exceeding the set threshold , and the intra-cluster variance When the number of clusters exceeds a certain set value, the cluster splitting operation is performed. and The distance between the centroids Less than the set threshold , and when the variance within the clusters is similar, clusters are merged.

[0075] (3) Noise weight adjustment:

[0076] To reduce the impact of noise and outliers, for each data point , and assign a weight according to its distance to the centroid , the weight is used to control the contribution of the point to the centroid update. The centroid update formula after adding the weight is as follows:

[0077] ,

[0078] (4) Local and global optimization:

[0079] First, the centroids within a cluster are fine-tuned locally; then, the distribution of centroids is re-evaluated globally to ensure balance between clusters.

[0080] The YADMW image target recognition model of the present invention is a new lightweight convolutional neural network architecture, wherein, see Figure 3 As shown, the method for DGConvNet module to process data includes:

[0081] The data is input in the form of feature maps, d represents the dimension, the number of channels is 96, and it is divided into two paths;

[0082] One way is to use the first point-by-point convolution layer to perform point-by-point convolution operation using a 7x7 convolution kernel, and the number of channels is 96; the other way is to use the first depth convolution layer to perform depth convolution operation using a 7x7 convolution kernel, and the number of channels is 96;

[0083] The outputs after two-way convolution are layer normalized (LN);

[0084] The feature map after layer normalization (LN) is divided into two paths again. One path passes through the second point-by-point convolution layer through 1x1 point-by-point convolution to expand the number of channels to 384; the other path passes through the second depth convolution layer through 1x1 point-by-point depth convolution to expand the number of channels to 384;

[0085] The GELU activation function is used for the feature maps with two extended channels ( Figure 3 ) and global response normalization ( Figure 3 The GRN in the figure is then processed by a standard convolutional layer;

[0086] The standard convolutional layer compresses the number of channels back to the number of channels at the output through 1x1 convolution, which is 96 channels.

[0087] Using depthwise separable convolution, the standard convolution is split into depthwise convolution and pointwise convolution, which significantly reduces the number of parameters and computational complexity. In order to solve the problem of feature collapse that may exist in the backbone network, global response normalization (GRN) is used to enhance feature competition between channels.

[0088] See also Figure 4 As shown in the figure, the data input to the MMQANext module enters the first local convolution layer and the second local convolution layer after layered sharing and factor decomposition, with a channel number of 96. After passing through the first local convolution layer, it enters the third and fourth local convolution layers after layered sharing and factor decomposition. The number of channels becomes 288 through channel expansion. The third and fourth local convolution layers are layer-normalized after channel fusion, with a channel number of 480. The fifth local convolution layer is layer-normalized after channel fusion with the first and second local convolution layers after parameter sharing. After being input into the MQA (multi-head attention) module, it is channel-fused with the second local convolution layer after parameter sharing, with a channel number of 288. Finally, it is output through the sixth local convolution layer, with a channel number of 96.

[0089] The low-rank approximation and hybrid sparse attention mechanisms are introduced into the adopted MQA module to further reduce the computational complexity of the attention matrix and the reasoning speed of the MQA module. Secondly, by introducing a hierarchical sharing mechanism, different key or value matrices are shared in the hierarchy to increase the representation capability while reducing parameters. Then, through parameter sharing and factor decomposition techniques, redundant parameters are reduced, making the YADMW image target recognition model more lightweight and suitable for deployment on mobile devices. Finally, the local convolution operation is combined with the MQA module to enhance the YADMW image target recognition model's ability to capture local features.

[0090] In order to further improve the detection accuracy of the model, the YADMW image target recognition model adopts the WEF Loss loss function, which is:

[0091] ,

[0092] is the weight, is the true label, is the predicted value, represents the edge detection operator, and are the edges of the predicted image and the real image, respectively. represents the landslide feature extractor, , , is the weight coefficient of loss, Represents the WEF Loss loss function, Represents the index, which is used to traverse each sample in the data set. Represents the total number of samples in the dataset.

[0093] In the landslide detection task, the frequency and severity of landslides in different regions may vary greatly, so the WEFLoss loss function can increase the model's attention to landslide areas by assigning higher weights to landslide classes. The introduction of boundary-aware loss allows the model to focus more on the boundary detection of landslide areas. The predicted boundary and the true boundary can be extracted, and the edge difference can be calculated to compare the extracted landslide features with the real landslide features. By calculating the difference in the landslide feature space, the YADMW image object recognition model can better capture the unique properties of landslides.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A landslide surface intelligent monitoring method based on YADMW image target recognition, characterized in that: The method includes: Use cameras to collect landslide surface image data at different angles, different time periods, different weather conditions and different geological types; The collected landslide surface image data are made into a landslide surface image data set; Construct a YADMW image target recognition model, use the landslide surface image data set to train the YADMW image target recognition model to obtain the final YADMW image target recognition model, and use the final YADMW image target recognition model to monitor the landslide surface image data collected in real time; The YADMW image target recognition model includes: a first MMQANext module connected to the data input terminal through a two-layer Conv module, and output to the first CBFuse module through an eighth Conv module after performing convolution operation in the first MMQANext module and reducing the calculation complexity and redundant parameters of the attention matrix; The data input end also continuously sets up a combination of multiple groups of Conv modules and DGConvNet modules after a layer of Conv modules, and outputs to the SPPELAN module through the last combined DGConvNet module. Except for the first DGConvNet module, each of the remaining combined DGConvNet modules outputs data to their respective corresponding CBLinear modules, and multiple CBLinear modules output to the first CBFuse module; the first CBFuse module fuses the outputs of multiple CBLinear modules and the output of the eighth Conv module; The output data of the first CBFuse module is respectively output to the first Conv-reg module and the first Conv-ds module through the second MMQANext module. The first Conv-reg module and the first Conv-ds module are output to the fifth Concat module and then output a detection result; The SPPELAN module is sequentially output through the first Upsample module, the first Concat module, the fifth MMQANext module, the second Upsample module and the second Concat module, and then output to the fourth Conv-reg module and the fourth Conv-ds module after the sixth MMQANext module. The fourth Conv-reg module and the fourth Conv-ds module are output to the eighth Concat module and then output a detection result.

2. The method for intelligent monitoring of landslide surface based on YADMW image target recognition according to claim 1 is characterized in that: The second MMQANext module outputs data to the third MMQANext module after passing through the ninth Conv module and the second CBFuse module in sequence. The output of the third MMQANext module is output to the sixth Concat module via the second Conv-reg module and the second Conv-ds module respectively, and then outputs a detection result; The third MMQANext module outputs data to the fourth MMQANext module after passing through the tenth Conv module and the third CBFuse module in sequence. The output of the fourth MMQANext module is output to the seventh Concat module via the third Conv-reg module and the third Conv-ds module, and then outputs a detection result.

3. The intelligent landslide surface monitoring method based on YADMW image target recognition according to claim 2 is characterized in that: The sixth MMQANext module is connected to the eleventh Conv module, the third Concat module and the seventh MMQANext module in sequence. The third Concat module also receives the output data of the fifth MMQANext module. The output of the seventh MMQANext module is output to the ninth Concat module via the fifth Conv-reg module and the fifth Conv-ds module, and then outputs a detection result.

4. The method for intelligent monitoring of landslide surface based on YADMW image target recognition according to claim 3 is characterized in that: The output of the seventh MMQANext module is connected to the twelfth Conv module, the fourth Concat module and the eighth MMQANext module in sequence. The fourth Concat module adds the output of the SPPELAN module and the output of the twelfth Conv module and then transmits them to the eighth MMQANext module. The output of the eighth MMQANext module is output to the tenth Concat module via the sixth Conv-reg module and the sixth Conv-ds module, and then outputs a detection result.

5. The method for intelligent monitoring of landslide surface based on YADMW image target recognition according to claim 1 is characterized in that: The collected landslide surface image data are made into a landslide surface image dataset, including: target labeling of the landslide surface image; data enhancement operation of the labeled landslide surface image; and target frame clustering using the AKS clustering algorithm.

6. The method for intelligent monitoring of landslide surface based on YADMW image target recognition according to claim 5 is characterized in that: The target box clustering using the AKS clustering algorithm includes: An adaptive learning mechanism is used in the AKS clustering algorithm, so that each centroid has an independent learning rate. The learning rate of the landslide surface image dataset is dynamically adjusted according to the dispersion or concentration of the data points in the cluster. Dynamic cluster structure adjustment, when the cluster The number of data points in Exceeding the set threshold , and the intra-cluster variance When the cluster exceeds a certain set value, the cluster splitting operation is performed; and Cluster The distance between the centroids Less than the set threshold , and when the variance within the cluster is similar, clusters are merged. Representation Cluster The centroid position of Representation Cluster The centroid position of For each data point , according to the data points The distance to the centroid is assigned a weight , used to control the data points Contribution to centroid update; The centroids within a cluster are fine-tuned locally, and the distribution of centroids is re-evaluated globally to ensure balance between clusters.

7. A landslide surface intelligent monitoring method based on YADMW image target recognition according to any one of claims 1 to 4, characterized in that: The WEF Loss loss function adopted by the YADMW image target recognition model is: ， is the true label, is the predicted value, represents the edge detection operator, and are the edges of the predicted image and the real image, respectively. represents the landslide feature extractor, , , is the weight coefficient of loss, Represents the WEF Loss loss function, Represents the index, which is used to traverse each sample in the data set. Represents the total number of samples in the dataset.

8. A landslide surface intelligent monitoring method based on YADMW image target recognition according to any one of claims 1 to 4, characterized in that: The method for processing data by the DGConvNet module includes: Data is input in the form of feature maps and is divided into two paths; One way is to use the first point-by-point convolution layer to perform point-by-point convolution operation using a 7x7 convolution kernel; the other way is to use the first depth convolution layer to perform depth convolution operation using a 7x7 convolution kernel; The outputs after two-way convolution are layer normalized; The feature map after layer normalization is divided into two paths again. One path passes through the second point-by-point convolution layer through 1x1 point-by-point convolution to expand the number of channels; the other path passes through the second depth convolution layer through 1x1 point-by-point depth convolution to expand the number of channels; The feature maps of the two expanded channels are processed by the standard convolutional layer after using the GELU activation function and global response normalization; The standard convolutional layer compresses the number of channels back to the number of channels at the output through 1x1 convolution.

9. A landslide surface intelligent monitoring method based on YADMW image target recognition according to any one of claims 1 to 4, characterized in that: For the MMQANext module, the input data enters the first local convolution layer and the second local convolution layer after hierarchical sharing and factor decomposition, with the number of channels being 96. After passing through the first local convolution layer, the data enters the third local convolution layer and the fourth local convolution layer after hierarchical sharing and factor decomposition. The number of channels becomes 288 through channel expansion. The third local convolution layer and the fourth local convolution layer are layer-normalized after channel fusion, with the number of channels being 480. The fifth local convolution layer is layer-normalized after parameter sharing with the first local convolution layer and the second local convolution layer after channel fusion. After being input into the MQA module, the data is channel-fused with the second local convolution layer after parameter sharing, with the number of channels being 288. Finally, the data is output through the sixth local convolution layer, with the number of channels being 96.

Images