Mountain area broken road associated area scene classification method based on optical remote sensing image
By combining topology modeling and directional pruning with the MFGL-Net model, the problem of identifying and classifying dead-end road association areas in mountainous areas was solved, achieving high-precision scene classification, which is suitable for the identification and monitoring of special features in mountainous areas.
Patent Information
- Application Number
- CN202511334880.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-18
- Publication Date
- 2026-02-10
AI Technical Summary
Existing technologies for identifying associated areas of dead-end roads in mountainous areas suffer from problems such as inaccurate identification of associated areas, insufficient scene classification accuracy, and a lack of systematic solutions. They are unable to effectively utilize the unique information of dead-end roads as road endpoints and are easily affected by background noise in complex mountainous environments.
A method based on optical remote sensing imagery is adopted to determine the endpoint of dead-end roads through topological modeling. Local image patches are cropped by directional cropping based on road azimuth angles, and the MFGL-Net scene classification model is used to classify scenes by fusing global visual features and semantic relationships between ground features.
It achieves high-precision identification and accurate classification of dead-end road association areas in complex mountainous environments, improves the accuracy of scene classification, and provides a complete end-to-end solution suitable for the identification and monitoring of special features in mountainous areas.
Smart Images
Figure CN121505318A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of optical remote sensing image analysis technology, specifically relating to a method for classifying mountainous dead-end road association areas based on optical remote sensing images. Background Technology
[0002] A "dead-end road" refers to a road whose endpoint does not connect to other established road networks. In remote sensing imagery, these roads, due to their unique topological structure, often indicate the special nature of the endpoint area. In remote mountainous areas far from residential activity, the endpoint of a dead-end road is often associated with certain special types of terrain features (such as tunnels through mountains, mine entrances, underground engineering entrances, and special-purpose non-civilian buildings and facilities). These areas may involve illegal activities such as illegal mining, logging, and drug production, or important military facilities. Therefore, discovering and identifying such special terrain features using high-resolution optical remote sensing imagery is of great significance, and dead-end roads can serve as key indicative clues.
[0003] In the field of road breakpoint identification technology, a large number of studies have been conducted in recent years on road extraction based on optical remote sensing images (Xu et al., 2018; Chen et al., 2021; Sheikh et al., 2022; Duan et al., 2025) and intersection node identification (Yang et al., 2024; Wei et al., 2020). The results have been widely applied to fields such as road network analysis (Wang et al., 2022; Jenelius et al., 2006) and post-disaster assessment (Wang et al., 2015; Aghababaei et al., 2021), which also laid the technical foundation for this study. Currently, road node detection methods are mainly divided into two categories: one is based on the road skeleton structure, using morphological refinement algorithms to extract the road centerline and combining eight-neighbor traversal or graph structure analysis to extract road nodes (Fan et al., 2020; Xue et al., 2024); the other is based on deep learning methods, which enhance the ability to model the topology of roads by fusing convolutional neural networks and graph neural networks, implicitly modeling road nodes, including endpoints, while constructing topological connections. For example, Roadtracer (Bastani et al., 2018) uses iterative search guided by a CNN-based decision function to detect road nodes; SAM-Road (Hetang et al., 2024) extracts road intersections through non-maximum suppression. These methods have achieved accurate extraction of road nodes, providing strong support for the extraction of breakpoints and associated areas of dead-end roads. However, related research has not paid attention to the unique information provided by dead-end roads as road endpoints, and has neglected the indicative role of dead-end roads in regional scenes.
[0004] To determine the purpose of the breakpoint association region, it is necessary to classify the images of these regions. Currently, the mainstream image classification methods have shifted from traditional methods based on low-level features (Santos et al., 2010; Han et al., 2002; Guo et al., 2010) and mid-level features (Yang et al., 2007; Gao et al., 2010) to methods based on deep learning networks (Wang et al., 2018; Cheng et al., 2017; Ma et al., 2021). Classic networks such as ResNet (He et al., 2016) and EfficientNet (Tan et al., 2019) are widely used in classification tasks for various types of land surface scenes. Attention modules such as SENet (Hu et al., 2018) and CBAM (Woo et al., 2018) effectively improve the model's ability to perceive key information, while the introduction of the Transformer (Dosovitskiy et al., 2020) architecture further enhances the model's ability to model long-distance semantic dependencies. These methods possess advantages such as strong feature representation capabilities, high classification accuracy, and good generalization performance, providing technical support for scene classification of dead-end road associated regions. However, the extent of dead-end road associated regions is ambiguous, easily affected by interference from nearby unrelated targets, and the targets are sparse and irregularly distributed, making it difficult for mainstream remote sensing scene classification methods to distinguish easily confused scenes based on subtle differences.
[0005] Despite advancements in road node identification and remote sensing image classification, some key shortcomings remain. These shortcomings are manifested in the following aspects:
[0006] In road identification, existing research mainly focuses on the connectivity and topological integrity of road networks, such as traffic flow analysis or post-disaster damage assessment. These studies treat road endpoints as ordinary topological nodes, failing to delve into dead-end roads as a special type of road endpoint. They neglect the fact that dead-end roads act as "directional cues," with their endpoint locations and the direction of road extension indicating surrounding contextual information. Consequently, they cannot accurately define and analyze specific areas associated with dead-end roads, nor can they effectively utilize this information to indicate the type of regional scene.
[0007] In image classification, while mainstream remote sensing scene classification methods perform well on general datasets, they face unique challenges when applied to mountainous areas with dead-end roads. These areas are typically ambiguous in scope, and target features are sparse and irregularly distributed, making it difficult for existing models to effectively capture their inherent semantic relationships. Traditional global feature extraction methods struggle to accurately distinguish these subtle differences, easily leading to misclassification. Furthermore, the complex terrain, varying lighting conditions, and occlusion in mountainous areas further exacerbate the difficulty of identification.
[0008] Due to the unique and rare nature of dead-end road connection zones, there are currently no large-scale, finely labeled public datasets available for research. This makes it difficult for existing technologies to solve the aforementioned problems through large-scale data training, resulting in a lack of systematic solutions specifically for this application domain.
[0009] In summary, existing technologies have the following shortcomings and drawbacks when using optical remote sensing images to identify special features in mountainous areas:
[0010] (1) Inaccurate identification of associated areas: Existing methods fail to make full use of the unique information of dead-end roads as the end of roads, and cannot accurately and directionally identify the context areas related to dead-end roads. They are easily affected by background noise and irrelevant ground features.
[0011] (2) Insufficient scene classification accuracy: When applied to mountainous environments with cluttered backgrounds and sparse targets, mainstream remote sensing scene classification methods rely too much on global visual features and are difficult to effectively distinguish easily confused ground features that are highly similar in detail and semantic relationship.
[0012] (3) Lack of systematic solutions: Existing technologies lack a complete technical framework that integrates "identification of dead-end road associated areas" and "classification of complex scenes".
[0013] In summary, previous research has laid the technical foundation for dead-end road detection and associated area scene classification, but a complete technical system for dead-end road applications has not yet been formed, and a systematic solution is still lacking. Therefore, this invention aims to solve the key technical problem of how to automatically, efficiently, and accurately identify associated areas of dead-end roads and determine their scene types from optical remote sensing images. Summary of the Invention
[0014] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method for classifying dead-end road association areas in mountainous areas based on optical remote sensing images. Using this method, dead-end road association areas can be automatically, efficiently and accurately identified from optical remote sensing images, and their scene types can be determined.
[0015] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0016] A scene classification method for dead-end road association areas in mountainous areas based on optical remote sensing imagery includes the following steps:
[0017] S1. Perform topology modeling based on road network data, and determine the endpoint of dead-end roads based on node connectivity analysis.
[0018] S2. A local image block is oriented and cropped from the optical remote sensing image. The local image block is a rectangular window centered on the end of the dead-end road and whose main axis direction is consistent with the road extension direction.
[0019] S3. Input the local image patch cropped in step S2 into the MFGL-Net scene classification model to predict the scene type of the dead-end road associated region.
[0020] Furthermore, in the above-described method for classifying mountainous dead-end road association areas based on optical remote sensing imagery, step S1 specifically includes:
[0021] S11. Model the road network data as an undirected or directed graph;
[0022] S12. Classify the node types according to the connectivity of each node in the graph;
[0023] S13. Traverse all nodes in the road graph and filter out all nodes with a connectivity of 1. These nodes are the endpoints of dead-end roads.
[0024] Furthermore, in the above-mentioned method for classifying mountain dead-end road association areas based on optical remote sensing images, step S2 determines the extension direction of the road by calculating the spatial azimuth of the dead-end road endpoint segment.
[0025] Furthermore, in the above-described method for classifying mountainous dead-end road association areas based on optical remote sensing imagery, the MFGL-Net scene classification model in step S3 includes a visual stream, a semantic stream, a dual-stream fusion module, and a linear classifier. The visual stream extracts global visual features from the input image, and the semantic stream constructs and analyzes semantic relationships between ground features. Then, the two feature streams are input into the dual-stream fusion module for splicing and fusion. Finally, the fused features are input into the linear classifier for prediction of the final scene type, thereby achieving accurate classification of mountainous dead-end road association areas.
[0026] Furthermore, in the above-described method for classifying mountain dead-end road association areas based on optical remote sensing images, the visual stream uses ResNet-101 as the backbone network. ResNet-101 contains 101 convolutional and pooling layers, which can extract abstract features from low to high levels from the input image layer by layer.
[0027] Furthermore, in the above-described method for classifying mountainous dead-end road association areas based on optical remote sensing imagery, the ResNet stream is configured with a global average pooling layer and two fully connected network layers at its end. The feature map output by the last residual block is first compressed into a vector with a spatial dimension of 1 by the global average pooling layer, and then the vector undergoes a nonlinear transformation through the two fully connected network layers. Finally, the ResNet stream outputs a 256-dimensional global visual feature vector as the visual input of the dual-stream fusion module.
[0028] Furthermore, in the above-described method for classifying mountainous dead-end road association areas based on optical remote sensing imagery, the semantic flow is based on GCN, constructing a semantic graph G(V, E, A) using 11 typical land cover types defined in the dataset as nodes, where V represents the set of nodes, E represents the set of edges, and A represents the adjacency matrix, where the adjacency matrix A∈R 11×11 Based on the statistical results of high-frequency co-occurrence of land features in the dataset, a predefined method is used to encode the semantic prior relationships between land feature categories;
[0029] In the semantic graph, each node v i The initial features include semantic attributes and geometric features, specifically: using word embedding vectors x emb ∈R 128 The semantic attributes representing land cover categories are obtained through a pre-trained model or a bag-of-words model; the center coordinates, width, and height of the bounding box are encoded into geometric features x using a multilayer perceptron. igeo ∈R 128 After concatenating semantic attributes and geometric features, the initial features of the nodes are generated through a nonlinear mapping layer.
[0030] Furthermore, in the above-described method for classifying mountainous dead-end road association areas based on optical remote sensing images, the semantic flow graph convolution process employs a self-connected adjacency matrix to improve training stability and uses a normalized Laplacian operator to achieve message passing. Finally, ground feature-aware pooling is used to sum the features of all nodes and then mapped to 256-dimensional global semantic features through a fully connected layer as the semantic input of the dual-stream fusion module.
[0031] Furthermore, as described above, the method for classifying mountainous dead-end road association areas based on optical remote sensing imagery involves inputting the 256-dimensional global visual features extracted by the ResNet stream and the 256-dimensional global semantic features extracted by the GCN stream into the dual-stream fusion module for concatenation, resulting in a 512-dimensional fused feature. After dimensionality reduction and nonlinear transformation by a fully connected layer, the fused feature is input into the linear classifier to predict the probability of each scene class.
[0032] Furthermore, in the above-mentioned method for classifying mountain dead-end road association areas based on optical remote sensing images, the Adam optimizer is used to minimize the cross-entropy loss function during training; to avoid overfitting, a Dropout layer is added to the network to randomly discard a certain proportion of parameters.
[0033] Compared with existing technologies, the scene classification method for dead-end road association areas in mountainous areas based on optical remote sensing images provided by this invention has the following beneficial effects:
[0034] (1) The method of identifying associated areas based on the end of a dead-end road and the orientation of the road azimuth angle can effectively focus on the key context information in front of it, and directionally and accurately crop out the image block that is most likely to contain the target scene, thereby greatly reducing background interference and improving the positioning accuracy of associated areas in mountainous environments with mixed ground features and blurred boundaries, thus solving the problem of insufficient directional sensitivity of traditional buffer methods.
[0035] (2) The MFGL-Net dual-stream scene classification model proposed in this invention, which integrates global visual features of images and semantic relationships of ground objects, can simultaneously utilize the visual information of images and the semantic information of ground objects to effectively distinguish highly similar scenes, significantly improve classification accuracy, and solve the problem that traditional remote sensing scene classification models have difficulty distinguishing similar scenes in mountainous environments with cluttered backgrounds, sparse targets and irregular distribution. The accuracy rate in typical experimental areas reaches 92.78%, which is significantly better than the mainstream baseline model.
[0036] (3) This invention integrates the solutions to the two core technical challenges mentioned above and provides a topology-guided scene classification framework for dead-end road association areas. This framework takes a geospatial perspective, using dead-end roads as the entry point. First, it accurately locates the association area using its topological information and azimuth angle, and then performs multimodal feature fusion classification on the area. This technical framework not only solves the technical problems of individual links, but more importantly, it provides an end-to-end complete solution from remote sensing imagery to final scene classification, which has significant engineering application value.
[0037] This invention provides a novel approach for identifying special ground and underground features in mountainous environments based on optical remote sensing imagery, offering feasible technical support for spatial scene classification in mountainous areas. This technical framework can provide effective technical support for the identification and dynamic monitoring of specific types of targets of interest in key areas, and is of great significance for combating illegal activities in mountainous regions and maintaining national security in border areas. Attached Figure Description
[0038] Figure 1 This is a flowchart of a method for classifying the scene of dead-end road association area in mountainous areas based on optical remote sensing imagery, provided in an embodiment of the present invention.
[0039] Figure 2 A topology-guided classification framework for dead-end road association areas;
[0040] Figure 3 This is a schematic diagram for extracting the endpoint of a dead-end road based on node connectivity.
[0041] Figure 4 A schematic diagram illustrating the identification of dead-end road association areas based on road azimuth angle;
[0042] Figure 5 A schematic diagram of the architecture of a dual-stream scene classification model that integrates global visual features and semantic relationships between ground features;
[0043] Figure 6 The experimental results for the study area are shown in the examples. Detailed Implementation
[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0045] To achieve accurate identification and classification of dead-end road scenarios under complex terrain conditions, this invention provides a method for classifying dead-end road association areas in mountainous regions based on optical remote sensing imagery. This method includes a topology analysis and spatial azimuth-based directional cropping method, along with an MFGL-Net scene classification model. First, the dead-end road endpoints are accurately extracted from vector road network data. Then, their spatial orientation is used to define semantically related neighboring regions. Finally, the neighboring regions are input into the MFGL-Net scene classification model to predict the scene type of the dead-end road association area.
[0046] Figure 1 The flowchart of a scene classification method for dead-end road association areas in mountainous areas based on optical remote sensing imagery provided in an embodiment of the present invention is shown. The method includes the following steps:
[0047] S1. Perform topology modeling based on road network data and determine the endpoint of dead-end roads based on node connectivity analysis. Specifically, this includes:
[0048] S11. Formalize the road network as an undirected or directed graph G=(V,E), where V represents the set of nodes, representing the endpoints and intersections of the roads; and E represents the set of edges, representing road segments.
[0049] Preferably, the road network is formally modeled as an undirected graph. Using an undirected graph model and combining it with road azimuth angles to determine direction is a simple and efficient step-by-step approach. While a directed graph could also be used, this would add unnecessary complexity for determining topological breakpoints. Using a directed graph would complicate the decision-making criteria. In one specific embodiment, the graph modeling of the road network is implemented using Python and its NetworkX library.
[0050] S12. Classify the node types according to the connectivity of each node in the graph. For any node v in the graph... i Its connectivity degree deg(v) i The value is defined as the number of road edges connected to the node, i.e.: deg(v i) =∣{e∈E∣v i ∈e}∣. Classify node types based on their connectivity values: if deg(vi If deg(v) = 1, it indicates that the node connects to only one road and is usually considered the end of a dead-end road; if deg(v) = 1, it indicates that the node connects to only one road and is usually considered the end of a dead-end road. i If )≥2, it indicates that the node is located in the middle of the road or at an intersection, and is a normally connected node.
[0051] S13. Traverse all nodes in the road graph and filter out the endpoints with a connectivity of 1, thus completing the initial extraction of the dead-end road.
[0052] This dead-end road identification method does not require complex image processing and relies solely on the topological information of the road network, thus possessing the advantages of high efficiency and robustness.
[0053] S2. Select a local image block from the optical remote sensing image. The local image block is a rectangular window centered on the end of the dead-end road and whose main axis direction is consistent with the road extension direction.
[0054] The road's extension direction is determined by calculating the spatial azimuth of the dead-end road segment, and a rectangular window is constructed with the dead-end road's endpoint as the center and its main axis aligned with the road's extension direction. This rectangular window is used to directionally crop local image patches from high-resolution optical remote sensing imagery, serving as input for subsequent scene classification. This method addresses the problem of inaccurate localization of associated areas caused by background interference in complex mountainous environments, a problem inherent in existing technologies.
[0055] S3. Input the local image patch cropped in step S2 into the MFGL-Net scene classification model to predict the scene type of the dead-end road associated region, thereby achieving accurate classification of the scene of the dead-end road associated region in the mountainous area.
[0056] This invention constructs an MFGL-Net scene classification model that integrates global visual features and semantic relationships between ground features. This model is used to classify scenes in cropped images of associated regions. The model employs a two-stream architecture: a visual stream extracts global visual features from the input image, while a semantic stream constructs and analyzes the semantic relationships between ground features. The two-stream architecture and training process of the MFGL-Net scene classification model are described in detail below.
[0057] Overall Model Architecture: The MFGL-Net scene classification model adopts a two-stream architecture, including a visual stream, a semantic stream, a two-stream fusion module, and a linear classifier. The visual stream is responsible for extracting global visual features from the input image, while the semantic stream constructs a semantic graph based on word embeddings of land cover categories and their co-occurrence statistics. It then aggregates node neighborhood information through a graph convolutional network to capture the contextual semantic relationships between land cover features. Finally, the two feature streams are input into the two-stream fusion module for concatenation and fusion, and then input into the linear classifier for final scene type prediction, thereby achieving accurate classification of scenes in mountainous dead-end road association areas. This method solves the problem that existing models struggle to distinguish easily confused scenes based on subtle differences in cluttered backgrounds, sparse and irregularly distributed targets in mountainous environments.
[0058] The working process of visual flow: To effectively extract global visual representations from remote sensing images, this invention preferably uses ResNet-101 as the backbone network. This network alleviates the gradient vanishing problem in deep model training through residual connections, demonstrating excellent performance in various visual tasks. ResNet-101 contains 101 convolutional and pooling layers, capable of extracting abstract features from low-level to high-level features from the input image layer by layer.
[0059] In this embodiment, a Global Average Pooling (GAP) layer and two fully connected layers are introduced at the end of the ResNet stream for feature compression and nonlinear enhancement. Specifically, the feature map output by the last residual block is first compressed into a vector with a spatial dimension of 1 by the GAP layer, and then this vector undergoes a nonlinear transformation through two fully connected layers; finally, the ResNet stream outputs a 256-dimensional global visual feature vector as the visual input to the dual-stream fusion module.
[0060] The working process of semantic flow: To capture the semantic association information of ground features in high-resolution remote sensing imagery, this embodiment designs a semantic flow based on graph convolutional network (GCN). This semantic flow constructs a semantic graph G(V, E, A) using 11 typical ground feature classes defined in the dataset as nodes, where V represents the set of nodes, E represents the set of edges, and A represents the adjacency matrix, where A∈R. 11×11 It is predefined based on the statistical results of high-frequency co-occurrence of land features in the dataset, and is used to encode the semantic prior relationships between land feature categories. Each node v i The initial features are composed of semantic attributes and geometric features: on the one hand, word embedding vectors x are used. emb ∈R 128 The semantic attributes representing land cover categories can be obtained through pre-trained models or bag-of-words models; on the other hand, the center coordinates (x, y, z) of their bounding boxes can be utilized. i yi Width w i and height h i Encoded as geometric features x by a multilayer perceptron (MLP) igeo ∈R 128 After the two are concatenated, the initial features of the nodes are generated through a nonlinear mapping layer.
[0061] The graph convolution process employs a self-connected adjacency matrix to improve training stability and uses a normalized Laplacian operator to achieve message passing. This embodiment fully models the semantic dependencies between ground features through two layers of GCN. Finally, entity-aware pooling is used to sum the features of all nodes and then mapped to 256-dimensional global semantic features through a fully connected layer.
[0062] Finally, the 256-dimensional global visual features extracted from the ResNet stream and the 256-dimensional global semantic features extracted from the GCN stream are concatenated to obtain a 512-dimensional fused feature. This fused feature is then subjected to dimensionality reduction and nonlinear transformation by a fully connected layer and input into a linear classifier to predict the probability of each scene class.
[0063] During training, the Adam optimizer is preferably used to minimize the cross-entropy loss function, with an initial learning rate of 0.0005 and a batch size of 8. To avoid overfitting, a Dropout layer (p=0.5) is added to the network to randomly discard a certain proportion of parameters. After the model training is complete, the model that performs best on the test set is used for the final prediction.
[0064] The effectiveness of this invention is verified through experiments in three typical mountainous environments. These environments encompass diverse geographical features and application contexts, including military border defense zones, mineral resource zones, and forestry resource zones.
[0065] Example 1: Identification of Sensitive Targets Above and Below Ground in Mountainous Environments Using Dead-End Roads as Clues
[0066] This study area is a typical test area with concealed characteristics. Using the method of this invention, we identified 42 dead-end roads in this area, including 23 dead-end roads associated with buildings, 13 dead-end roads associated with mountains, and 6 dead-end roads associated with areas with isolation attributes. The above results demonstrate that this invention has high accuracy and effectiveness in identifying concealed targets associated with mountains in complex geographical environments.
[0067] Example 2: Refined Identification and Monitoring Application of Mining Area Scenes Using Dead-End Roads as Clues
[0068] Applying the method of this invention to typical mining areas aims to achieve refined identification and monitoring of mining scenes using disconnected cable traces. Compared with the broad and coarse labeling of OSM, this invention can more precisely identify hidden or small features within the mining area, such as associated buildings and mine entrances.
[0069] Furthermore, this invention can also assist in the supervision of illegal mining. By spatially overlaying the identification results of dead-end roads with legal mining rights data, auxiliary clues can be provided for judging illegal mining activities. This data can also provide a more accurate data foundation for assessing the scale of mining areas and predicting production capacity.
[0070] Example 3: Application and Value of Dead-End Roads in Forest Areas
[0071] This invention can use dead-end roads as clues for early warning of illegal activities. By regularly monitoring remote sensing images, small-scale illegal activities such as illegal mining and logging that may exist near the end of a dead-end road can be detected and warned in a timely manner.
[0072] Experimental results show that the MFGL-Net scene classification model provided in this invention significantly outperforms several mainstream benchmark models in the task of classifying dead-end road associated areas. Compared to the single ResNet101 model, the MFGL-Net scene classification model improves accuracy by 3.15 percentage points, reaching 92.78%. This demonstrates the superiority of the technical solution that integrates global visual features and semantic relationships of ground features when handling complex remote sensing scenes. Furthermore, the MFGL-Net scene classification model also performs well in metrics such as recall and F1 score, verifying its effectiveness in identifying sparse target ground features.
[0073] The present invention provides a scene classification method for mountainous dead-end road association areas based on optical remote sensing imagery, offering a new approach to identifying special above-ground and underground features in mountainous environments and providing feasible technical support for spatial scene classification in mountainous areas. This method can provide effective technical support for the identification and dynamic monitoring of specific types of targets of interest in key areas, and is of great significance for combating illegal activities in mountainous regions and maintaining national security in border areas.
[0074] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention is also intended to include these modifications and variations.
Claims
1. A method for classifying dead-end road association areas in mountainous areas based on optical remote sensing imagery, comprising the following steps: S1. Perform topology modeling based on road network data and determine the endpoint of dead-end roads based on node connectivity analysis; S2. A local image block is oriented and cropped from the optical remote sensing image. The local image block is a rectangular window centered on the end of the dead-end road and whose main axis direction is consistent with the road extension direction. S3. Input the local image patch cropped in step S2 into the MFGL-Net scene classification model to predict the scene type of the dead-end road associated region.
2. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to claim 1, characterized in that, Step S1 specifically includes: S11. Model the road network data as an undirected or directed graph; S12. Classify the node types according to the connectivity of each node in the graph; S13. Traverse all nodes in the road graph and filter out all nodes with a connectivity of 1. These nodes are the endpoints of dead-end roads.
3. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to claim 2, characterized in that, In step S2, the direction of road extension is determined by calculating the spatial azimuth of the end line segment of the dead-end road.
4. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to any one of claims 1-3, characterized in that, The MFGL-Net scene classification model described in step S3 includes a visual stream, a semantic stream, a dual-stream fusion module, and a linear classifier. The visual stream extracts global visual features from the input image, and the semantic stream constructs and analyzes the semantic relationships between ground features. Then, the two feature streams are input into the dual-stream fusion module for splicing and fusion. Finally, the fused features are input into the linear classifier for prediction of the final scene type, thereby achieving accurate classification of scenes in mountainous dead-end road associated areas.
5. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to claim 4, characterized in that, The visual stream uses ResNet-101 as the backbone network. ResNet-101 contains 101 convolutional and pooling layers, which can extract abstract features from the input image layer by layer from low level to high level.
6. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to claim 5, characterized in that, The ResNet stream is configured with a global average pooling layer and two fully connected network layers at the end. The feature map output by the last residual block is first compressed into a vector with a spatial dimension of 1 by the global average pooling layer, and then the vector is nonlinearly transformed by the two fully connected network layers. Finally, the ResNet stream outputs a 256-dimensional global visual feature vector as the visual input of the dual-stream fusion module.
7. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to claim 6, characterized in that, The semantic flow is based on GCN, constructing a semantic graph G(V, E, A) using 11 typical land features defined in the dataset as nodes, where V represents the set of nodes, E represents the set of edges, and A represents the adjacency matrix, where A∈R. 11×11 Based on the statistical results of high-frequency co-occurrence of land features in the dataset, a predefined method is used to encode the semantic prior relationships between land feature categories; In the semantic graph, each node v i The initial features include semantic attributes and geometric features, specifically: using word embedding vectors x emb ∈R 128 The semantic attributes representing land cover categories are obtained through a pre-trained model or a bag-of-words model; the center coordinates, width, and height of the bounding box are encoded into geometric features x using a multilayer perceptron. igeo ∈R 128 After concatenating semantic attributes and geometric features, the initial features of the nodes are generated through a nonlinear mapping layer.
8. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to claim 7, characterized in that, The semantic flow graph convolution process employs a self-connected adjacency matrix to improve training stability and uses a normalized Laplacian operator to achieve message passing. Finally, ground-presence-aware pooling is used to sum the features of all nodes and then mapped through a fully connected layer to a 256-dimensional global semantic feature, which serves as the semantic input to the dual-stream fusion module.
9. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to claim 8, characterized in that, The 256-dimensional global visual features extracted from the ResNet stream and the 256-dimensional global semantic features extracted from the GCN stream are input into the dual-stream fusion module for concatenation, resulting in a 512-dimensional fused feature. The fused feature is then subjected to dimensionality reduction and nonlinear transformation by a fully connected layer and input into the linear classifier to predict the probability of each scene class.
10. The method for classifying mountainous dead-end road association areas based on optical remote sensing imagery according to any one of claims 5-9, characterized in that, During training, the Adam optimizer is used to minimize the cross-entropy loss function; to avoid overfitting, a Dropout layer is added to the network to randomly discard a certain proportion of parameters.