A multi-view graph clustering method for unmanned aerial vehicle remote sensing data
By constructing an initial graph structure and using multi-view comparative learning, dynamic bias removal and adaptive enhancement of weak view representations, the problem of insufficient clustering accuracy caused by modal heterogeneity in UAV remote sensing data is solved, achieving efficient and unsupervised multi-view graph clustering results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HAINAN UNIV
- Filing Date
- 2026-02-02
- Publication Date
- 2026-06-05
AI Technical Summary
Multi-view data from UAV remote sensing exhibits modal heterogeneity. Existing clustering methods struggle to effectively mitigate modal bias in dominant views and fully exploit complementary information in weaker views, resulting in insufficient clustering accuracy and generalization ability. This is particularly problematic under unlabeled conditions, where precise analysis is difficult to achieve.
By constructing an initial graph structure, multi-view contrastive learning and dynamic debiased variable generation are used to generate gradient update masks. The original branch and the enhancement branch are processed in parallel to adaptively enhance the weak view representation. The global consensus embedding is optimized through a unique information consistency loss to achieve feature consistency and robustness among views.
It significantly improves the clustering accuracy and generalization ability of UAV remote sensing data, reduces the dependence on high-quality labeled data, and provides an accurate and efficient unsupervised clustering scheme that is suitable for scenarios such as urban planning and environmental monitoring.
Smart Images
Figure CN122156691A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing mapping technology, and in particular to a multi-view graph clustering method for UAV remote sensing data. Background Technology
[0002] With the rapid development of UAV technology, it has become one of the core carriers for remote sensing data acquisition due to its advantages of high mobility, low cost, high resolution and flexible deployment. It is widely used in key fields such as urban planning, environmental monitoring, resource exploration and emergency rescue. UAVs can quickly collect multimodal remote sensing data, including hyperspectral images (HSI), synthetic aperture radar (SAR), digital surface models (DSM), texture feature maps (such as EMP, LBP, Gabor) and other multi-view data. These data describe the physical, structural and texture features of ground objects from different dimensions, providing rich information support for tasks such as accurate ground object identification and land cover classification.
[0003] However, the effective utilization of UAV remote sensing data still faces two major challenges: First, the cost of acquiring high-quality labeled data is extremely high. The amount of multi-view data collected by UAVs is huge, the scenes are varied and the types of ground features are complex. Manual labeling is not only time-consuming and labor-intensive, but also difficult to ensure label consistency. Especially in scenarios with high real-time requirements such as dynamic monitoring and emergency response, the scarcity of labeled data is even more prominent, which seriously limits the application of supervised learning methods. Therefore, unsupervised learning paradigms have become a key technical path for UAV remote sensing data mining. Among them, multi-view graph clustering (MVC) has become one of the core technologies for UAV remote sensing data analysis because of its ability to fuse multi-source heterogeneous data and extract meaningful patterns from the data under unlabeled conditions.
[0004] On the other hand, UAV multi-view data exhibits significant modal heterogeneity. Data from different views vary greatly in terms of feature distribution, data dimensionality, and noise levels. This leads to a situation where, during the training of clustering models, some dominant views with strong discriminative power tend to dominate the learning process and excessively control the learning of feature representations, while the complementary information of weaker views is suppressed and difficult to fully exploit. At the same time, the high-resolution nature of UAV data results in a large amount of detailed noise and redundant information, further exacerbating the heterogeneity of multi-view data. This makes it difficult for existing clustering methods to learn robust shared representations, ultimately affecting clustering accuracy and generalization ability.
[0005] To address the fusion problem of multi-view data, Deep Multi-View Clustering (DMVC) technology has emerged. Utilizing powerful encoders such as Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs), it learns unified shared representations from heterogeneous data sources, significantly improving clustering performance. The introduction of contrastive learning further enhances feature discriminativeness and consistency between views, becoming a crucial technical support for multi-view clustering. However, most existing deep multi-view clustering methods treat all views equally, failing to fully consider the uniqueness of UAV multi-view data. UAV multimodal data exhibits greater heterogeneity, larger data volume, and more complex noise distribution, leading to more pronounced suppression effects of dominant views and insufficient learning of weak views. These methods cannot dynamically adapt to the modal differences in UAV multi-view data, failing to effectively mitigate modal bias caused by dominant views or specifically enhance the representational capabilities of weak views. Ultimately, this results in suboptimal clustering results, failing to meet the practical needs of accurate analysis of UAV remote sensing data. Summary of the Invention
[0006] In view of this, the present invention proposes a multi-view graph clustering method for UAV remote sensing data, which can dynamically alleviate modal bias, adaptively enhance weak view representation, and improve clustering performance.
[0007] The technical solution of this invention is implemented as follows: A multi-view graph clustering method for UAV remote sensing data includes the following steps: Step S1: Construct an initial graph structure from the remote sensing multi-view data collected by the UAV, which includes a set of node feature vectors and an adjacency matrix. Step S2: Based on the initial graph structure, obtain the initial representation of all nodes under the corresponding view data, and perform multi-view comparison learning on all view data according to the initial node representation to obtain the overall multi-view comparison loss; Step S3: Aggregate the initial node representations into aggregated feature representations, calculate the debiased variables of the view data, generate a gradient update mask based on the debiased variables, and obtain the adjusted gradient. Step S4: Using the initial graph structure as input, combined with the debiased variables and the adjusted gradient, the original branch and the enhancement branch are processed in parallel. After fusing the outputs of the two branches, the fused representation of the view data is obtained, and the unique information consistency loss is calculated. Step S5: Integrate the fusion representation of all views, combine the overall multi-view contrast loss, unique information consistency loss and reconstruction loss to construct the overall objective function, and output the global consensus embedding after optimization; Step S6: Use a clustering algorithm to cluster the global consensus embedding and output the final clustering result.
[0008] Preferably, the specific steps of step S1 are as follows: The SLIC algorithm is used to input images for each view in the remote sensing multi-view data acquired by the UAV. Divided into N superpixels Each pixel is a node in the view, where H is the image height, W is the image width, and C is the number of feature channels; Define the feature vector of node i This represents the mean of all pixel features within the corresponding superpixel:
[0009] in For superpixels Number of pixels included The features representing pixel p; An adjacency matrix is constructed based on spatial proximity and feature similarity. :
[0010] in The preset feature similarity threshold, Describing the vector norm, Let be the superpixel of node j.
[0011] Preferably, step S2 includes the following specific steps: Input the set of node feature vectors and the adjacency matrix of each view data into the encoder to obtain the initial representation of all nodes under that view data; For any two distinct view data v and u, construct positive sample pairs and negative sample pairs, and calculate the contrast loss between view data v and u. ; For all view data combinations Calculate the overall multi-view contrast loss .
[0012] Preferably, the contrast loss The expression is:
[0013] in Represents the similarity function, Here, N is the temperature parameter, and N is the number of nodes. and These represent the initial node representations of node i in view data v and u, respectively, forming a positive sample pair. For the initial representation of node j in view data u, and Forming negative sample pairs; The expression for the overall multi-view contrast loss is: .
[0014] Preferably, the specific steps of step S3 are as follows: Initial representation of nodes for each view data v belonging to view set V. Perform global aggregation to obtain the aggregated feature representation of view data v. ; Calculate the debiasing variable for each view data v ,in The expression is:
[0015] in For the aggregated feature representation of view data u, view data , and Representing aggregated features respectively and The norm; The gradient update mask is generated based on the debiased variables, and the gradient weights of the enhanced branch and the original branch are adjusted. The formula is as follows:
[0016] in To enhance the gradient of the branch, The gradient of the original branch. This is the adjusted gradient.
[0017] Preferably, step S4 includes the following specific steps: The initial graph structure is input into the original branch, which performs feature encoding and outputs the original branch representation. ; The initial graph structure is input into the enhancement branch, which updates its parameters based on the adjusted gradient, performs adaptive enhancement encoding on the initial graph structure, and outputs the enhancement branch representation. ; Based on debiased variables By merging the outputs of the two branches, a fused representation of the view data is obtained. :
[0018] Calculate the unique information consistency loss of the enhanced branch. .
[0019] Preferably, the unique information consistency loss The expression is:
[0020] in It represents a distance or divergence measure.
[0021] Preferably, the overall objective function of step S5 is expressed as follows:
[0022] in To reconstruct the loss, For overall multi-view contrast loss, For the loss of consistency of unique information, and This is a hyperparameter.
[0023] Preferably, the reconstruction loss The expression is:
[0024] Where V is a set of views, and view data v belongs to the set of views V. Let i be the feature vector of node i in view data v. for The reconstruction loss corresponds to the reconstruction result of the decoder. It is the Frobenius norm.
[0025] Compared with the prior art, the beneficial effects of the present invention are: This invention presents a multi-view graph clustering method for UAV remote sensing data. First, an initial graph structure is constructed through superpixel segmentation, preserving ground feature information while reducing noise interference and computational complexity. Then, multi-view comparative learning enhances feature consistency between views. Dynamic view de-biasing is combined to suppress excessive dominance of the dominant view in real time, avoiding the loss of complementary information. Next, the original and enhanced branches are processed and fused in parallel to adaptively enhance the representation of weak views, and a unique information consistency loss ensures the effectiveness of the enhanced features. Finally, robust consensus embedding is obtained through multi-loss joint optimization, significantly improving clustering accuracy and generalization ability. This method does not rely on high-quality labeled data and solves the problems of strong heterogeneity in UAV remote sensing multi-view data, dominant view suppression of weak views, and high labeling costs. It provides an accurate and efficient unsupervised clustering solution for scenarios such as urban planning and environmental monitoring. Attached Figure Description
[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0027] Figure 1 This is a flowchart of a multi-view graph clustering method for UAV remote sensing data according to the present invention; Figure 2 This is a comparison chart of clustering visualization results of the XuZhou dataset, which is an embodiment of the multi-view graph clustering method for UAV remote sensing data according to the present invention. Figure 3 Hyperparameters of an embodiment of a multi-view graph clustering method for UAV remote sensing data according to the present invention and Sensitivity analysis plots on the XuZhou, Salinas, and Trento datasets. Detailed Implementation
[0028] To better understand the technical content of this invention, a specific embodiment is provided below, and the invention will be further described in conjunction with the accompanying drawings.
[0029] See Figure 1 The present invention provides a multi-view graph clustering method for UAV remote sensing data, comprising the following steps: Step S1: Construct an initial graph structure from the remote sensing multi-view data collected by the UAV, including a set of node feature vectors and an adjacency matrix. The specific steps are as follows: The SLIC algorithm is used to input images for each view in the remote sensing multi-view data acquired by the UAV. Divided into N superpixels Each pixel is a node in the view, where H is the image height, W is the image width, and C is the number of feature channels; Define the feature vector of node i This represents the mean of all pixel features within the corresponding superpixel:
[0030] in For superpixels Number of pixels included The features representing pixel p; An adjacency matrix is constructed based on spatial proximity and feature similarity. :
[0031] in The preset feature similarity threshold, Describing the vector norm, For the superpixel of node j, by constructing the initial graph structure, we can reduce noise interference and computational complexity while preserving key structural information.
[0032] Step S2: Based on the initial graph structure, obtain the initial representations of all nodes under the corresponding view data. Then, perform multi-view comparison learning on all view data based on the initial node representations to obtain the overall multi-view comparison loss. Specific steps include: Input the set of node feature vectors and the adjacency matrix of each view data into the encoder to obtain the initial representation of all nodes under that view data; For any two distinct view data v and u, construct positive sample pairs and negative sample pairs, and calculate the contrast loss between view data v and u. Comparison of losses The expression is:
[0033] in Represents a similarity function (such as cosine similarity). Here, N is the temperature parameter, and N is the number of nodes. and These represent the initial node representations of node i in view data v and u, respectively, forming a positive sample pair. For the initial representation of node j in view data u, and Forming negative sample pairs; For all view data combinations Calculate the overall multi-view contrast loss The expression is: .
[0034] By employing a contrastive learning framework to learn multi-view representations, a contrastive loss function is constructed for each view to encourage the aggregation of similar node pairs and the separation of heterogeneous node pairs. By maximizing mutual information between views and promoting consistency between modalities, multi-view representations are effectively learned.
[0035] Step S3: Aggregate the initial node representations into aggregated feature representations, calculate the debiasing variables of the view data, generate a gradient update mask based on the debiasing variables, and obtain the adjusted gradient. The specific steps are as follows: Initial representation of nodes for each view data v belonging to view set V. Perform global aggregation to obtain the aggregated feature representation of view data v. ; Calculate the debiasing variable for each view data v ,in The expression is:
[0036] in For the aggregated feature representation of view data u, view data , and Representing aggregated features respectively and The norm; The gradient update mask is generated based on the debiased variables, and the gradient weights of the enhanced branch and the original branch are adjusted. The formula is as follows:
[0037] in To enhance the gradient of the branch, The gradient of the original branch. This is the adjusted gradient.
[0038] This invention proposes a dynamic view debiasing method. For different view data v and u belonging to the view set V, their respective initial node representations can be obtained in step S2, and then the corresponding aggregated feature representations can be obtained through global aggregation. For view data v, the debiasing variable is calculated based on its relative contribution to the overall loss or feature norm. For the dominant view ( Larger), freeze a larger proportion of parameters in the enhancement branch, reducing its learning ability to avoid overfitting, for weak views ( (Smaller) retains more trainable parameters, promoting feature learning and performance compensation. This adaptive freezing is achieved through gradient update masks, which ensures that complementary information from all view data is effectively captured.
[0039] Step S4: Taking the initial graph structure as input, and combining the debiased variables and adjusted gradients, the data is processed in parallel through the original branch and the enhancement branch. The outputs of the two branches are fused to obtain the fused representation of the view data, and the unique information consistency loss is calculated. The specific steps include: The initial graph structure is input into the original branch, which performs feature encoding and outputs the original branch representation. ; The initial graph structure is input into the enhancement branch, which updates its parameters based on the adjusted gradient, performs adaptive enhancement encoding on the initial graph structure, and outputs the enhancement branch representation. ; Based on debiased variables By merging the outputs of the two branches, a fused representation of the view data is obtained. :
[0040] Calculate the unique information consistency loss of the enhanced branch. The expression is:
[0041] in It represents a distance or divergence measure (such as mean square error or cosine distance).
[0042] Further improvements to weak view representations are achieved through view-specific augmentation, where the augmentation branch for each view is modulated by debiased variables to control the degree of feature learning, wherein the augmentation branch representation is obtained through debiased variables. Adaptive scaling is used to fuse the outputs of the two branches. Then, to ensure that the augmented branches retain the complementary and unique information of each view, a unique information consistency loss is introduced. This loss encourages the augmented branch representation to maintain consistency with the original branch representation while preserving view-specific features. This loss can be used to regularize the augmentation process, avoid overfitting, and ensure the informativeness and discriminativeness of the augmented features.
[0043] Step S5: Integrate the fused representations of all views, combine the overall multi-view contrast loss, unique information consistency loss, and reconstruction loss to construct an overall objective function, and output the global consensus embedding after optimization. The expression of the overall objective function is:
[0044] in For the reconstruction loss (with a fixed weight of 1), used to measure the model's (decoder branch's) ability to reconstruct the original node from the learned embeddings, this invention uses the mean squared error averaged over the view and the node:
[0045] Where V is a set of views, and view data v belongs to the set of views V. Let i be the feature vector of node i in view data v. for The reconstruction loss corresponds to the reconstruction result of the decoder. It is the Frobenius norm. For overall multi-view contrast loss, For the loss of consistency of unique information, and The hyperparameter is used to control the trade-offs between components. By jointly optimizing this objective function, the framework can effectively learn multi-view representations, dynamically identify dominant modalities, and adaptively enhance weak views, thereby improving clustering performance.
[0046] Step S6: Use a clustering algorithm to cluster the global consensus embedding and output the final clustering result.
[0047] To verify the effectiveness of this invention, the following comparative experiments were conducted: Datasets and baseline methods: Three publicly available remote sensing multi-view datasets were used: Trento, XuZhou, and Salinas. The Trento dataset contains HSI and LiDAR bimodal data, totaling 30,214 samples, divided into 6 classes; the XuZhou dataset covers four complementary views: HSI, EMP, LBP, and Gabor, totaling 68,877 samples, divided into 9 classes; and the Salinas dataset integrates HSI, EMP, and Gabor trimodal data, totaling 54,129 labeled samples, covering 16 land cover types.
[0048] The comparative experiment selected 11 advanced multi-view clustering methods, including 6 general MVC methods (EMVCC, SSGCC, MM24, TGRS24, AAAI25, TCSVT25) and 5 remote sensing-specific MVC methods (SEC-LSRM, CDD, TKDE25), as well as two ablation variants. All methods were compared fairly under the same experimental configuration.
[0049] Implementation details: The experiments were implemented using Python and PyTorch frameworks, with an NVIDIA GeForce RTX4090 GPU (24GB VRAM) and 128GB of system memory. To reduce random fluctuations, each experiment was repeated 10 times, and results are expressed as mean ± standard deviation. The baseline method uses the official open-source code and reproduces the reported performance.
[0050] Model hyperparameter configuration: learning rate 0.001, weight decay 0.0001, batch size 256, training epochs 200. The SLIC algorithm is used for superpixel segmentation to generate graph nodes. Hyperparameter sensitivity analysis is used to determine the contrastive loss weights. And consistency loss weight All values were 0.01. Three commonly used metrics were used to evaluate clustering performance: accuracy (ACC), Kappa coefficient, and normalized mutual information (NMI).
[0051] The clustering performance and visualization results are shown in Table 1: Table 1. Performance comparison of this embodiment with other methods on three remote sensing datasets.
[0052] As shown in Table 1, the quantitative results demonstrate that the multi-view clustering method of this invention generally outperforms the traditional k-means algorithm. The framework proposed in this embodiment achieves optimal performance across all metrics for all datasets, with ACC improvements of 3.88% (Xu Zhou), 3.16% (Salinas), and 1.01% (Trento) compared to the suboptimal baseline methods. This verifies the effectiveness of dynamic view debiasing and specific view enhancement mechanisms in mitigating dominant modality bias and strengthening weak view representations.
[0053] To verify the effectiveness of view de-biasing (D) and view enhancement (E), an ablation experiment was conducted in this embodiment. In Table 1, w / oD represents the removal of the view de-biasing module, and w / oE represents the removal of a specific view enhancement module. The experimental results show that: Removing the view debiasing module (w / oD) resulted in a significant decrease in clustering accuracy (e.g., the ACC of the XuZhou dataset dropped from 80.87% to 77.69%), demonstrating its crucial role in mitigating modal bias. Removing specific view enhancement modules (w / oE) also led to performance degradation (ACC dropped to 77.83% on the XuZhou dataset), validating their importance in enhancing weak view representations; The complete model (OURS) consistently outperformed both ablation variants, demonstrating the complementary advantages of the two modules and that their synergy is key to achieving optimal performance.
[0054] Figure 2 The visualization results of clustering on the XuZhou dataset are presented (GT represents the ground truth label map). It can be seen that the clustering results in this embodiment highly match the ground truth labels, effectively capturing the spatial distribution and boundary characteristics of different land cover types. In contrast, the baseline method exhibits more misclassification and incoherent clustering in multi-class adjacent areas. This visualization further supports the quantitative results, validating the superiority of this method in achieving accurate and meaningful clustering in remote sensing multi-view data.
[0055] To verify the hyperparameters (Comparison of loss weights) and Sensitivity analysis was performed to examine the impact of consistency loss weights on model performance, and the results are as follows: Figure 3 As shown, the analysis results indicate that: and When the size is too small, the model cannot fully utilize the consistency between views and the regularization constraints of the augmented branches in contrastive learning, resulting in poor performance; Too large and Imposing excessive constraints may distort the inherent information of features or cause training instability, which will also lead to a decrease in performance. when and Within the range of 0.001 to 0.1, this embodiment maintains stable high performance with only minor fluctuations; The optimal hyperparameters are =0.01sum =0.01, achieving peak performance on all datasets, validating the importance of view debiasing and balancing optimization of specific view enhancement modules.
[0056] This embodiment proposes a multi-view graph clustering method for UAV remote sensing data. It constructs a Debiased and Enhanced Multi-View Graph Clustering (DTE-MVGC) framework. By introducing a dynamic view debiasing module, it alleviates the dominance of some modalities. At the same time, it designs an adaptive view enhancement module to strengthen the representation of weak views. This framework can effectively capture the complementary information of all modalities, improve the robustness of clustering results, and show significant advantages. Extensive experiments on three datasets (including ablation studies, hyperparameter analysis, and visualization results) further verify the effectiveness and superiority of the debiasing.
[0057] 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multi-view graph clustering method for UAV remote sensing data, characterized in that, Includes the following steps: Step S1: Construct an initial graph structure from the remote sensing multi-view data collected by the UAV, which includes a set of node feature vectors and an adjacency matrix. Step S2: Based on the initial graph structure, obtain the initial representation of all nodes under the corresponding view data, and perform multi-view comparison learning on all view data according to the initial node representation to obtain the overall multi-view comparison loss; Step S3: Aggregate the initial node representations into aggregated feature representations, calculate the debiased variables of the view data, generate a gradient update mask based on the debiased variables, and obtain the adjusted gradient. Step S4: Using the initial graph structure as input, combined with the debiased variables and the adjusted gradient, the original branch and the enhancement branch are processed in parallel. After fusing the outputs of the two branches, the fused representation of the view data is obtained, and the unique information consistency loss is calculated. Step S5: Integrate the fusion representation of all views, combine the overall multi-view contrast loss, unique information consistency loss and reconstruction loss to construct the overall objective function, and output the global consensus embedding after optimization; Step S6: Use a clustering algorithm to cluster the global consensus embedding and output the final clustering result.
2. The multi-view graph clustering method for UAV remote sensing data according to claim 1, characterized in that, The specific steps of step S1 are as follows: The SLIC algorithm is used to input images for each view in the remote sensing multi-view data acquired by the UAV. Divided into N superpixels Each pixel is a node in the view, where H is the image height, W is the image width, and C is the number of feature channels; Define the feature vector of node i This represents the mean of all pixel features within the corresponding superpixel: in For superpixels Number of pixels included The features representing pixel p; An adjacency matrix is constructed based on spatial proximity and feature similarity. : in The preset feature similarity threshold, Describing the vector norm, Let be the superpixel of node j.
3. The multi-view graph clustering method for UAV remote sensing data according to claim 1, characterized in that, The specific steps of step S2 include: Input the set of node feature vectors and the adjacency matrix of each view data into the encoder to obtain the initial representation of all nodes under that view data; For any two distinct view data v and u, construct positive sample pairs and negative sample pairs, and calculate the contrast loss between view data v and u. ; For all view data combinations Calculate the overall multi-view contrast loss .
4. The multi-view graph clustering method for UAV remote sensing data according to claim 3, characterized in that, The contrast loss The expression is: in Represents the similarity function, Here, N is the temperature parameter, and N is the number of nodes. and These represent the initial node representations of node i in view data v and u, respectively, forming a positive sample pair. For the initial representation of node j in view data u, and Forming negative sample pairs; The expression for the overall multi-view contrast loss is: 。 5. The multi-view graph clustering method for UAV remote sensing data according to claim 1, characterized in that, The specific steps of step S3 are as follows: Initial representation of nodes for each view data v belonging to view set V. Perform global aggregation to obtain the aggregated feature representation of view data v. ; Calculate the debiasing variable for each view data v ,in The expression is: in For the aggregated feature representation of view data u, view data , and Representing aggregated features respectively and The norm; The gradient update mask is generated based on the debiased variables, and the gradient weights of the enhanced branch and the original branch are adjusted. The formula is as follows: in To enhance the gradient of the branch, The gradient of the original branch. This is the adjusted gradient.
6. The multi-view graph clustering method for UAV remote sensing data according to claim 1, characterized in that, The specific steps of step S4 include: The initial graph structure is input into the original branch, which performs feature encoding and outputs the original branch representation. ; The initial graph structure is input into the enhancement branch, which updates its parameters based on the adjusted gradient, performs adaptive enhancement encoding on the initial graph structure, and outputs the enhancement branch representation. ; Based on debiased variables By merging the outputs of the two branches, a fused representation of the view data is obtained. : Calculate the unique information consistency loss of the enhanced branch. .
7. A multi-view graph clustering method for UAV remote sensing data according to claim 6, characterized in that, The unique information consistency loss The expression is: in It represents a distance or divergence measure.
8. The multi-view graph clustering method for UAV remote sensing data according to claim 1, characterized in that, The overall objective function of step S5 is expressed as follows: in To reconstruct the loss, For overall multi-view contrast loss, For the loss of consistency of unique information, and This is a hyperparameter.
9. A multi-view graph clustering method for UAV remote sensing data according to claim 8, characterized in that, The reconstruction loss The expression is: Where V is a set of views, and view data v belongs to the set of views V. Let i be the feature vector of node i in view data v. for The reconstruction loss corresponds to the reconstruction result of the decoder. It is the Frobenius norm.