Road extraction network and method based on wavelet mixed attention and hierarchical feature collaborative enhancement

Abstract

The application provides a road extraction network and method based on wavelet mixed attention and hierarchical feature collaborative enhancement, and belongs to the technical field of remote sensing. The application takes D-LinkNet as a basic U-shaped backbone architecture, and in view of the technical pain points of the existing remote sensing road extraction technology, such as loss of coded down-sampling detail information, insufficient cross-level feature interaction and fusion, insufficient modeling capability of multi-scale slender roads and broken roads, and easy occurrence of missed detection, false detection and structure fracture in road extraction, three core improvement modules are designed, which are a wavelet mixed attention module WHAM, a hierarchical collaborative channel attention module HCCAM and a double-path interactive multi-scale pooling module DPIMSPM. Through collaborative optimization of the three modules, the application significantly improves the connectivity, structural integrity and segmentation accuracy of remote sensing road extraction in complex scenes. The network of the application can be widely applied in the fields of city planning surveying and mapping, traffic dynamic monitoring, automatic driving geographic information collection, national remote sensing survey and the like.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, specifically to a road extraction network and method based on wavelet hybrid attention and hierarchical feature synergistic enhancement. Background Technology

[0002] Remote sensing road extraction is a core technology that uses computer vision to automatically analyze remote sensing images, identify, and segment road networks. It is a fundamental and core component of geographic information updating, intelligent transportation, and autonomous driving perception, and has extremely high engineering application value. Existing remote sensing road extraction technologies are mainly divided into two categories: traditional manual feature extraction methods and intelligent extraction methods based on deep learning.

[0003] Traditional road extraction methods primarily rely on manually designed road spectral, textural, and geometric features, using fixed rules to identify and segment roads. These methods are highly dependent on manual design experience, have limited feature representation capabilities, cannot adapt to complex scenarios such as vegetation occlusion, building shadows, small side roads, and dense intersections, and exhibit extremely poor generalization ability, making it difficult to meet the engineering requirements of large-scale, high-precision remote sensing road extraction.

[0004] With the development of deep learning technology, models such as convolutional neural networks and Transformers have been widely used in remote sensing road extraction tasks, resulting in a large number of improved algorithms such as D-LinkNet, CoANet, DA-RoadNet, and TransRoadNet. These algorithms improve road extraction accuracy through methods such as dilated convolution, striped convolution, attention mechanisms, and global modeling using Transformers. However, existing mainstream technologies still have three core inherent defects that severely limit the effectiveness of road extraction in complex scenarios: First, there is a significant loss of detail information at the encoding end. Existing network encoders generally employ operations such as continuous downsampling and max pooling to expand the receptive field and extract high-level semantic features. However, this process causes irreversible loss of high-frequency detail information of roads in remote sensing images, ultimately leading to blurred road edges, broken small branches, and missing intersection structures in the model-extracted images, significantly reducing the completeness and precision of road extraction. Existing attention optimization schemes mostly focus on single-dimensional optimization in the spatial and channel domains, failing to distinguish between low-frequency features of road structure and high-frequency features of details in the frequency domain, and thus cannot specifically compensate for the detail loss caused by downsampling.

[0005] Second, cross-level feature fusion is inefficient. Existing U-shaped network architectures rely on skip connections to achieve the transfer and compensation of shallow features from the encoder to deep features from the decoder. However, they generally use simple feature concatenation and pixel-by-pixel addition to fuse features. Shallow features are rich in spatial details but weak in semantic information, while deep features have strong semantic representation capabilities but lack spatial details. Direct fusion will produce serious semantic gaps and scale misalignment problems, and lacks a cross-level feature collaborative calibration mechanism, failing to fully leverage the complementary advantages of shallow and deep features, resulting in degradation of road details and limited segmentation accuracy.

[0006] Third, there is insufficient capability for multi-scale complex road modeling. Remote sensing roads are characterized by their elongated strip shape, multi-scale distribution, and susceptibility to local occlusion and breakage. Existing multi-scale feature extraction modules mostly adopt a single convolution mode, which cannot simultaneously adapt to the long-range dependency modeling of road elongated geometry and curved and broken roads. In complex scenes, problems such as road discontinuity, false background detection, and local missed detection are very likely to occur, making it difficult to meet the needs of accurate extraction of roads in multiple scenes and at multiple scales.

[0007] In summary, existing technologies cannot simultaneously address the core challenges of lost detailed information, insufficient cross-level feature interaction, and weak multi-scale complex road modeling capabilities, resulting in deficiencies in accuracy, completeness, and robustness. Therefore, this invention proposes a remote sensing road extraction network and method based on wavelet hybrid attention and hierarchical feature synergy enhancement. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a road extraction network and method based on wavelet hybrid attention and hierarchical feature synergy enhancement. It aims to overcome the shortcomings of existing remote sensing road extraction technologies, such as loss of coding details, insufficient cross-hierarchical feature fusion, and poor multi-scale complex road modeling capabilities. Through innovative designs such as frequency domain high and low frequency feature synergy enhancement, cross-hierarchical channel attention calibration, and dual-path multi-scale feature complementary fusion, it significantly improves the extraction completeness of road edges, small side roads, dense intersections, and broken roads in complex scenes, suppresses the problems of missed detections and false detections, and while significantly improving segmentation accuracy, it reasonably controls the model computational complexity, achieves a balance between accuracy and inference efficiency, and meets the needs of engineering implementation.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement uses D-LinkNet network as the basic backbone architecture. The overall network architecture consists of encoder, decoder and skip connection branches connecting encoder and decoder. The network integrates three core modules: the wavelet hybrid attention module WHAM embedded in the encoder downsampling node, the hierarchical collaborative channel attention module HCCAM configured in the skip connection branch, and the dual-path interactive multi-scale pooling module DPIMSPM deployed in the intermediate feature processing layer of the network. The aforementioned wavelet hybrid attention module uses dual-tree complex wavelet transform (DTCWT) to split the input features into low-frequency components and multi-directional high-frequency components. It uses the low-frequency components as prior information to adaptively enhance high-frequency detail features and completes feature reconstruction through inverse dual-tree complex wavelet transform (IDTCWT). The aforementioned hierarchical collaborative channel attention module performs global pooling and cross-level channel cascading on multi-level features of the encoder to generate cross-scale attention weights, thereby completing multi-scale feature weighting calibration and resolution unification. The aforementioned dual-path interactive multi-scale pooling module sets up a bar convolution branch and a multi-scale deformable attention dual parallel feature extraction path. The output features of the two paths are fused by the complementary weighted fusion module CWFM to output the final features. By relying on the collaboration of three major modules, the complete extraction of multi-scale, slender, and intersecting roads in remote sensing images can be achieved, suppressing defects such as missed detection of road edges, broken small branches, and false detection of background.

[0010] The processing steps of the wavelet hybrid attention module (WHAM) described above include: S1. The input coding features are decomposed by DTCWT to obtain a single low-frequency feature component and multiple high-frequency sub-band components containing real and imaginary parts; S2. The high-frequency components are compressed into channels by two-stage cascaded 1×1 convolutions and activated by Sigmoid to generate the first attention weights. The weights are multiplied by the original high-frequency components to complete the initial high-frequency feature enhancement. S3, low-frequency components, and high-frequency components after initial enhancement are respectively subjected to 3×3 convolution operation and then added element by element to achieve frequency domain feature fusion. The fused features are then subjected to 1×1 convolution and Sigmoid to generate second attention weights, which are used for secondary adaptive modulation of high-frequency components. S4. The modulated high-frequency component and the original low-frequency component are fed together into the IDTCWT inverse transform to reconstruct the feature map with enhanced output details.

[0011] The above-mentioned hierarchical collaborative channel attention module HCCAM is adapted to the three-layer adjacent encoding features of the U-shaped network. The input module processing flow is as follows: Global average pooling (GAP) is performed on the input features of each level to extract channel semantic vectors. All channel vectors are concatenated along the channel dimension to obtain a joint channel representation. The joint channel representation is fed into a lightweight attention network and normalized by Softmax to generate cross-level attention weight coefficients. The weight coefficients are used to weight and modulate the corresponding level features, and all features are scaled to the same resolution through bilinear interpolation.

[0012] In the aforementioned hierarchical collaborative channel attention module HCCAM, the fused features after resolution unification are output in three multi-resolution convolutional branches. The three output features are adapted to the different levels of input of the decoder, and simultaneously supplement the decoder with three types of feature information: shallow details, middle structure, and high-level semantics.

[0013] The aforementioned dual-path interactive multi-scale pooling module DPIMSPM is divided into two parallel feature extraction branches: the first branch is a multi-scale strip convolution branch, and the second branch is a multi-scale deformable attention branch. The strip convolution branch uses six different asymmetric convolution kernels ({1×3, 3×1, 1×5, 5×1, 1×7, 7×1}) in parallel to extract features, adapting to the strip-shaped geometric shape of remote-sensed roads. The output feature channels of each convolution are concatenated and then dynamically fused using lightweight attention to obtain the strip features. .

[0014] The aforementioned multi-scale deformable attention branch constructs attention units in parallel using multiple sets of deformable convolutions with kernels of different sizes. It adaptively adjusts the feature sampling position using learnable offsets to achieve long-distance context modeling of curved, occluded, and broken roads. The multi-scale deformable feature channels are concatenated and then fused with lightweight dynamic weights to obtain deformable features. .

[0015] The above-mentioned complementary weighted fusion module CWFM implementation and The fusion method is as follows: the two feature streams are concatenated channel by channel and then fed into a depthwise separable convolution, combined with Sigmoid activation to generate a complementary weight map. The complementary weights are then used to apply the complementary weights to the following functions: The weighted, fine-structured regions are dominated by bar features, while the global semantic regions are dominated by deformable features. The weighted summation outputs the final features of the module.

[0016] The above network training phase uses the Adam optimizer, an initial learning rate of 0.0002, and a batch size of 4. If the loss does not decrease for 3 consecutive rounds, the learning rate is halved, and the maximum number of epochs is 200. Training is terminated early if the loss is higher than the historical best loss or the learning rate is lower than 0.0000005 for 6 consecutive rounds.

[0017] The network loss function described above uses a weighted joint loss function of binary cross-entropy (BCE) and Dice loss. BCE is used to alleviate the imbalance between road and background pixel categories in remote sensing images, while Dice loss optimizes the segmentation accuracy of small target roads with fine branches.

[0018] The application method of the road extraction network based on wavelet hybrid attention and hierarchical feature co-enhancement described above includes the following steps: S1. Image preprocessing: The original remote sensing images are uniformly cropped to 512×512 pixels using a non-overlapping cropping method. For areas where the image edges are smaller than this size, edge pixel padding is used to fill in the gaps. Data augmentation operations such as horizontal flipping, vertical flipping, and random scaling are performed on the training set images to improve the model's generalization ability.

[0019] S2. Model Training: Based on the PyTorch framework, using the CHN6-CUG and Massachusetts public remote sensing road datasets, the WHNet network is iteratively trained according to the preset training parameters, and the optimal weight model of the validation set is saved.

[0020] S3. Image Inference and Extraction: The preprocessed remote sensing image to be detected is input into the trained optimal WHNet model. The model outputs a binary road segmentation map through forward inference, completing the fully automatic and accurate extraction of remote sensing roads. The output results can be directly applied to scenarios such as urban surveying, traffic monitoring, and autonomous driving geographic information collection.

[0021] The road extraction network and method based on wavelet hybrid attention and hierarchical feature synergy enhancement mentioned in this invention have the following significant advantages: 1. Precise preservation of frequency domain details, solving the problem of lost road details. This invention is the first to create a wavelet hybrid attention module, which achieves high- and low-frequency feature separation through dual-tree complex wavelet transform. Relying on the prior knowledge of low-frequency structure to guide the adaptive enhancement of high-frequency details, it differs from the traditional spatial domain attention optimization scheme. It accurately compensates for the loss of details of road edges, small branches, and intersections caused by encoder downsampling at the frequency domain level, effectively solving the problems of road blurring and branch breakage in traditional models.

[0022] 2. Cross-level feature fusion to eliminate feature fusion defects. The hierarchical collaborative channel attention module of this invention realizes cross-scale adaptive weighted calibration of three-layer encoded features, effectively eliminating the semantic gap and scale misalignment problem between shallow and deep features, giving full play to the complementary advantages of shallow details, mid-level structure and deep semantic features, greatly improving the feature fusion quality of skip connections and strengthening the integrity of road structure.

[0023] 3. Dual-path multi-scale modeling, adapting to complex road features. The dual-path interactive multi-scale pooling module of this invention combines the advantages of strip convolution and deformable attention. It can adapt to the long strip-shaped geometric shape of roads and adaptively model the long-range dependencies of curved, occluded, and broken roads. By dynamically fusing the features of the two paths through complementary weights, it takes into account both local details and global semantics, significantly improving the ability to extract multi-scale roads in complex scenes.

[0024] 4. Significantly improved accuracy, excellent robustness and generalization. Test results on two authoritative public datasets, CHN6-CUG and Massachusetts, show that the four core metrics of this invention—F1, IoU, precision, and recall—are all superior to six mainstream comparison algorithms. The extraction performance of fine branches, occluded roads, and dense intersections is significantly improved, and the false positive and false negative rates are greatly reduced. The model has extremely strong generalization ability and scene adaptability.

[0025] 5. Balance between accuracy and computational load, with strong engineering applicability. The model parameters and floating-point operations of this invention are at an industry average level. Compared with Transformer-type algorithms, it significantly reduces computational overhead, and compared with lightweight CNN models, it significantly improves accuracy. It balances high accuracy and high efficiency, making it suitable for engineering application scenarios such as batch automated processing of remote sensing images and real-time road network monitoring. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the overall architecture of the road extraction network of the present invention; Figure 2 This is a schematic diagram of the wavelet hybrid attention module of the present invention; Figure 3 This is a schematic diagram of the hierarchical collaborative channel attention module of the present invention; Figure 4 This is a schematic diagram of the structure of the dual-path interactive multi-scale pooling module of the present invention; Figure 5 The visualization results are shown for different networks on the CHN6-CUG dataset in the examples. Figure 6 The above are visualizations of different networks on the Massachusetts dataset in the examples. Detailed Implementation

[0027] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0028] A road extraction network (WHNet) and method based on wavelet hybrid attention and hierarchical collaborative feature enhancement are proposed. The method suppresses the loss of detail information by collaboratively enhancing high-frequency features with low-frequency features. A wavelet hybrid attention module (WHAM) is constructed, and a hierarchical collaborative channel attention module (HCCAM) is proposed to achieve efficient information transfer and collaborative enhancement between encoded features at different levels. A dual-path interactive multi-scale pooling module (DPIMSPM) is designed to enhance the modeling ability of road features at different scales, thereby improving the connectivity and structural integrity of extracted roads in complex scenes. The network structure is as follows: Figure 1 As shown in the image.

[0029] The road extraction network WHNet uses D-LinkNet as its basic architecture, such as Figure 2 As shown, the structure of the wavelet hybrid attention module (WHAM) is as follows: In the encoding stage of remote sensing road extraction tasks, while continuous downsampling expands the receptive field and enhances high-level semantics, it inevitably causes irreversible loss of high-frequency details, leading to blurred road edges and broken narrow branches. To overcome the damage to detail information caused by traditional downsampling strategies (such as max pooling), the wavelet hybrid attention module (WHAM) is embedded in a key node after feature downsampling and before deep semantic modeling. Based on the dual-tree complex wavelet transform (DTCWT), it decomposes features into multi-directional high-frequency subbands and low-frequency components. The core mechanism of WHAM lies in leveraging the structural stability of low-frequency features to collaboratively enhance high-frequency features: by applying a channel attention mechanism to high-frequency components to adaptively capture details, and simultaneously introducing low-frequency components as prior guidance, it effectively suppresses noise interference introduced during the downsampling process. This collaborative strategy of "low-frequency guidance and high-frequency enhancement" significantly improves the network's ability to perceive and retain key high-frequency features of roads, thereby effectively mitigating the loss of detail information at the encoding stage and enhancing the extraction accuracy of fine road structures.

[0030] The WHAM workflow is as follows: First, the input feature map is decomposed by DTCWT to generate a low-frequency component. and a set of high-frequency components containing real and imaginary parts. Subsequently, an attention mechanism was designed within the module to enhance this set of high-frequency components, which contain both real and imaginary parts. High-frequency components First, channel compression and feature calibration are performed using two cascaded 1×1 convolutional layers, followed by a sigmoid activation function to generate an attention weight map. This weight map is then compared with high-frequency components. Multiplication yields the enhanced high-frequency components. Its mathematical expression is as follows: (1) (2) Where DTCWT is the dual-tree complex wavelet transform; B is the product operation; B is bilinear interpolation; Conv is the convolution operation. σ It is the Sigmoid activation function; In the low-frequency-guided high-frequency enhancement stage, the low-frequency component is first... and high frequency components A 3×3 convolution is applied to each component, and then the two convolutions are added element-wise to fuse low-frequency semantic information with high-frequency structural details. Next, a 1×1 convolutional layer is used to compress the fused features through channels, followed by a sigmoid activation function to generate an attention weight map. This weight map is then compared with the high-frequency components. Element-wise multiplication is performed to achieve adaptive modulation of high-frequency features. Finally, the weighted high-frequency components and low-frequency components are input together into the dual-tree complex wavelet inverse transform (IDTCWT) to reconstruct the enhanced output. Its mathematical expression is as follows: (3) (4) (5) Where iDTCWT is the inverse complex wavelet transform of a dual-tree system; It is a product operation; Conv is a convolution operation. σ It is the Sigmoid activation function.

[0031] like Figure 3 As shown, the structure of the hierarchical collaborative channel attention module HCCAM is as follows: In remote sensing road extraction tasks, the elongated shape and structural complexity of road targets require models to possess both high-resolution detail and strong semantic representation capabilities. While traditional U-shaped networks compensate for detail loss by transferring encoder features to the decoder through skip connections, they often employ simple concatenation or addition operations, ignoring the semantic gap and scale differences between features at different levels. Specifically, low-level features are rich in spatial detail but semantically weak, high-level features are semantically strong but lack spatial information, and mid-level features serve a transitional role. Without an effective collaborative mechanism, direct fusion can easily lead to semantic inconsistencies and spatial alignment deviations, resulting in detail degradation. To address this issue, the Hierarchical Collaborative Channel Attention Module (HCCAM) achieves efficient information transfer and collaborative enhancement between encoded features at different levels. By introducing a channel attention mechanism, HCCAM can adaptively recalibrate the channel weights of multi-level features, suppress irrelevant noise, and highlight key road features, thereby eliminating semantic mismatch during cross-level information transfer. This hierarchical collaborative strategy not only optimizes feature fusion quality but also significantly improves the model's ability to preserve the integrity and detail of the road structure.

[0032] First, for multi-scale input features The channel context is modeled. Specifically, global average pooling is applied to the features at each level to aggregate their spatial dimensionality information, thereby extracting the global semantic response of the channel dimension at each scale. This operation can effectively measure the relative importance of features at different levels to the final prediction. Subsequently, the channel description vectors obtained from GAP at each level are concatenated along the channel dimension to form a joint channel representation that integrates multi-scale contextual information, providing a basis for subsequent attention weight generation or feature recalibration.

[0033] (6) (7) Where Concat is a cascading operation; GAP is global average pooling. Subsequently, the module introduces a lightweight attention generation network that adaptively learns the cross-layer weight distribution among features at three scales and normalizes it using the Softmax function to generate cross-scale attention vectors. This mechanism enables the network to dynamically focus on feature layers containing key road structures based on the input scene. Building upon this, features at each scale are weighted and modulated according to their corresponding attention weights, and then uniformly upsampled to the target resolution using bilinear interpolation. This effectively mitigates feature alignment bias caused by scale inconsistencies, thereby avoiding information fragmentation during cross-scale fusion and improving the coherence and integrity of the road structure.

[0034] (8) (9) (10) (11) Concat is a cascading operation; B is the product operation; B is bilinear interpolation; Conv is the convolution operation. Finally, the fused features are input into the multi-resolution convolutional branch, which generates three scale outputs corresponding to the decoder, respectively. These outputs are used to simultaneously enhance high-resolution details, mid-scale structures, and low-resolution semantic representations, thereby further improving the connectivity and structural consistency of road extraction.

[0035] (12) Concat is a cascading operation; B is the product operation; B is bilinear interpolation; Conv is the convolution operation.

[0036] like Figure 4 As shown, the structure of the dual-path interactive multi-scale pooling module DPIMSPM is as follows: Remotely sensed roads exhibit significant multi-scale characteristics and long linear features. Single-scale feature modeling struggles to capture both local details and global coherence, often leading to road fragmentation or insufficient multi-scale road extraction in complex scenarios. To address this, the Dual-path interactive multi-scale pooling module (DPIMSPM) aims to enhance the network's ability to extract features from roads at different scales and fractured roads through a dual-path collaborative interaction mechanism. Specifically, the module first processes the input features... E i The path is divided into two complementary paths. One path introduces a multi-scale deformable attention mechanism, utilizing multiple sets of different receptive fields. （K 1 ,K 2 ,K 3 ） Parallel modeling of contextual information and adaptive adjustment of sampling positions using learnable offsets enhance the modeling ability for complex geometries such as bends and fractures. To better adapt to the geometric characteristics of roads, another path employs strip convolutional kernels of different sizes (1×3, 3×1, 1×5, 5×1, 1×7, and 7×1) to construct a multi-scale feature extraction module, enabling the network to handle the slender strip-like structures of roads more naturally. Based on this, the two paths are dynamically fused into... and The module then enters the Convolutional Weighted Feature Modulation (CWFM) module. CWFM dynamically generates the fusion weights. By employing a complementary weighting mechanism, an adaptive balance between global context and local details is achieved, effectively mitigating local ambiguity caused by occlusion and background interference. Meanwhile, the deformable attention module adopts a lightweight design, directly generating queries, keys, and offsets through convolution, thus improving the modeling capability for long-range dependencies while controlling computational overhead.

[0037] First, the input feature map is processed. E i The network is divided into channels and fed into multi-scale strip convolution branches and multi-scale deformable convolutional attention branches. Different sizes of strip convolution kernels (1×3, 3×1, 1×5, 5×1, 1×7, 7×1) are used in the multi-scale strip convolution branch to construct the multi-scale feature extraction module, enabling the network to handle the long, thin strip structures of roads more naturally. The network then performs layer-by-layer processing on each channel, followed by adaptive feature fusion through dynamic weights. Its mathematical expression is as follows: (13) The kernel sizes are as follows: (14) (15) Where Conv is the convolution operation; Weights are generated using lightweight attention; In the multi-scale deformable convolutional attention branch, deformable convolutional kernels of different sizes are used to construct deformable attention. These kernels are then cascaded across channels, and adaptive fusion to obtain features is achieved through dynamic weights. Its mathematical expression is as follows: (16) Where Conv is the convolution operation; Weights are generated using lightweight attention; Multi-scale strip convolution branches Attention branch of multi-scale deformable convolution The complementary weighted fusion module (CWFM) is used for adaptive fusion. First, the features from the two branches are concatenated along the channel dimension, then fed into a lightweight convolutional block, and a weight map is generated using the sigmoid activation function. Based on this, the weight map is further processed... and The final output is obtained by applying complementary weights to each region. This design ensures that the fusion process is adaptively balanced in terms of content: regions with strong semantic information requirements are weighted by... Dominated, while areas of fracture or fine structure are dominated by Dominant. The final output calculation is as follows: (17) (18) Stack performs a channel-wise concatenation operation; DWConv is a depthwise separable convolution operation. σ It is the Sigmoid activation function.

[0038] Example: Experiments were conducted on the network structure and method of the present invention to verify its technical effects, including: I. Experimental Environment The experiment was implemented using the PyTorch deep learning framework, running on a 64-bit Windows 11 operating system, and using Python 3.8. Considering the GPU resource limitations of the device, a non-overlapping cropping method was used to crop the original image to 512×512 pixels. For edge regions smaller than this size, padding was performed using edge pixels. To improve the model's generalization ability and robustness, various data augmentation operations were performed on the training set samples, including horizontal flipping, vertical flipping, and scaling.

[0039] The Adam optimizer is used for training, with an initial learning rate of 0.0002 and a batch size of 4. If the loss function does not decrease for three consecutive epochs, the learning rate is halved. The maximum number of epochs is set to 200. Training stops if the loss exceeds the current minimum loss or the learning rate falls below 0.0000005 for six consecutive epochs.

[0040] To address the imbalance between road labels and background in high-resolution remote sensing imagery, a joint loss function combining binary cross-entropy loss (BCE) and Dice loss is employed as the overall loss function. Its expression is: ; L represents the overall loss function, P is the predicted image output by the network, and Y is the true label. The BCE loss addresses class imbalance, while the Dice loss effectively optimizes segmentation performance for small target regions. The combination of these two losses enhances the model's ability to identify road regions.

[0041] II. Experimental Parameter Settings To ensure the fairness and integrity of the experiment and the data, this study strictly followed the partitioning scheme provided by the official dataset and did not screen or modify the data samples in any way. Considering the GPU resource limitations of the device, a non-overlapping cropping method was used to crop the original images to a size of 512×512 pixels. For edge regions that were smaller than this size, they were padded with edge pixels. To improve the generalization ability and robustness of the model, various data augmentation operations were performed on the training set samples, including horizontal flipping, vertical flipping, and scaling.

[0042] The Adam optimizer is used for training, with an initial learning rate of 0.0002 and a batch size of 4. If the loss function does not decrease for three consecutive epochs, the learning rate is halved. The maximum number of epochs is set to 200. Training stops if the loss exceeds the current minimum loss or the learning rate falls below 0.0000005 for six consecutive epochs.

[0043] To address the imbalance between road labels and background in high-resolution remote sensing imagery, a joint loss function combining binary cross-entropy loss (BCE) and Dice loss is employed as the overall loss function. Its expression is: ; L represents the overall loss function, P is the predicted image output by the network, and Y is the true label. The BCE loss addresses class imbalance, while the Dice loss effectively optimizes segmentation performance for small target regions. The combination of these two losses enhances the model's ability to identify road regions.

[0044] To verify the effectiveness of the proposed network, six road extraction methods—CARNet, RCFSNet, MSMDFFNet, DBRANet, D-LinkNet, and TransRoadNet—were used as comparative experiments. Precision (P), recall (R), F1-Score (F1), and intersection-over-union (IoU) were employed as evaluation metrics. ; ; ; ; TP, FP, TN, and FN represent the number of true positives, false positives, true negatives, and false negatives, respectively. In road extraction tasks, F1 score and IoU are two key evaluation metrics. The F1 score is the harmonic mean of precision and recall, used to comprehensively measure the overall performance of the model. This metric balances false positives and false negatives, avoiding focusing on only a single evaluation dimension, thus providing a more comprehensive performance assessment. In road extraction tasks, a higher F1 score indicates that the model can extract road centerlines more accurately while reducing false positives and false negatives. IoU primarily measures the overlap between the model's predictions and the actual road labels; it is defined as the ratio of the intersection to the union of the predicted road region and the actual road region. A higher IoU value means that the model's predictions match the actual roads better, reflecting more accurate road extraction results.

[0045] III. Datasets and Evaluation Metrics This experiment selects two publicly available road extraction datasets to verify the feasibility of the proposed method: the CHN6-CUG road dataset and the Massachusetts road dataset. The two road datasets are selected from different geographical regions and have high spatial resolution. The basic information about these two road datasets will be introduced below.

[0046] The CHN6-CUG road dataset is a representative large-scale urban satellite remote sensing imagery road dataset, covering typical cities with different levels of urbanization, city size, development stage, urban structure, and historical and cultural background. The dataset includes 4511 remote sensing images, each 512 pixels × 512 pixels with a spatial resolution of 0.5m. 3608 images were used for model training, and 903 images were used for testing and evaluation.

[0047] The Massachusetts Roads Dataset is a remote sensing road dataset covering various scenarios in urban, suburban, and rural areas. The dataset contains 1171 remote sensing images, each 1500 pixels × 1500 pixels with a spatial resolution of 1 meter. The original dataset includes 1108 training images, 14 validation images, and 49 test images.

[0048] To more comprehensively evaluate the overall performance of the model, in addition to traditional accuracy metrics, this experiment further introduces performance metrics such as model parameter count (Params), floating-point operations (FLOPs), and extraction time / inference speed. Params measure the total number of learnable parameters in the model, reflecting model size and storage overhead; a smaller parameter count generally indicates a more lightweight model. FLOPs measure the number of floating-point operations required for one forward propagation, reflecting computational complexity and resource consumption. Extraction time or inference speed measures the model's efficiency in real-world road extraction tasks, typically expressed as processing time per image or frames per second (FPS). By introducing these metrics, different methods can be comprehensively analyzed from multiple perspectives, including model accuracy, complexity, and real-time performance, more fully validating the effectiveness and practicality of the proposed method in remote sensing image road extraction tasks. To ensure the fairness and consistency of the experimental results, the FLOPs of all models were statistically analyzed under a batch size of 4.

[0049] IV. Experimental Results and Analysis like Figure 5 As shown, six sets of road extraction results from the CHN6-CUG dataset are presented: (a) and (b) represent the original image and the ground truth label, respectively; (c)~(k) represent CARNet, RCFSNet, MSMDFFNet, DBRANet, D-LinkNet, and TransRoadNet, respectively. Figure 5As can be seen, for the CHN6-CUG dataset, the proposed method achieves road extraction results that are closest to the real labels in most scenarios. In particular, the proposed WHNet demonstrates strong stability and integrity in road recognition at different scales, and its boundary extraction is more continuous and accurate, indicating a significant advantage in adapting to complex road structures. In contrast, other comparative methods have obvious shortcomings. For example, in the region shown in row 1, most comparative methods exhibit significant missed detections at road edges, resulting in incomplete road contours, while the proposed method only shows a few false positives, with clearer overall boundaries and significantly improved extraction accuracy. In the regions shown in rows 2 and 4, when facing complex road structures with breaks and dense intersections, other methods generally suffer from discontinuous road connections or local missing sections, while the proposed method effectively maintains the continuity and integrity of the road structure. This advantage is mainly due to the effective design of the DPIMSPM module in multi-directional feature modeling and break region compensation, thereby enhancing the model's ability to express complex topological structures. In the results shown in row 3, RCFSNet, MSMDFFNet, D-LinkNet, and DBRANet all exhibited significant false negatives, while our proposed method demonstrated greater robustness in identifying fine structures and recovering weak response regions. In the region shown in row 5, our method achieved more complete road boundary delineation and more natural edge transitions, while effectively reducing false positives and false negatives, further validating the model's robustness and generalization ability in complex scenarios.

[0050] like Figure 6 The image shows six sets of road extraction results from the Massachusetts dataset: (a) and (b) represent the original image and the ground truth label, respectively; (c) to (k) represent CARNet, RCFSNet, MSMDFFNet, DBRANet, D-LinkNet, and TransRoadNet, respectively. Figure 6As can be seen, for the Massachusetts dataset, our method still outperforms other methods in terms of visual performance. The visualization results show that WHNet can depict the overall road network structure more completely, especially showing greater stability in densely distributed road areas. For narrow branches and locations with breaks, the model recovers road continuity more clearly and expresses boundaries more accurately. In the first row of examples, most of the compared methods exhibited false positives or false negatives to varying degrees, while our method only had a few false positives, resulting in cleaner and more coherent overall segmentation, demonstrating higher extraction accuracy. In the third and fifth rows, although most models could identify the main road structure, they had many false negatives in detailed areas, while our method performed better in preserving small structures and restoring connectivity. In the fourth row of results, MSMDFFNet and DBRANet showed relatively obvious false positives, while RCFSNet had some false negatives; the other methods and our method could generally detect the complete road structure, but our method was more robust in terms of detail integrity and noise suppression.

[0051] To more objectively and comprehensively evaluate the performance of different models in the road extraction task, Tables 1 and 2 list the quantitative evaluation metrics of each model on the CHN6-CUG dataset and the Massachusetts dataset, including precision, recall, F1 score, and intersection-over-union ratio (IoU).

[0052] As shown in Table 1, the proposed WHNet achieves the best overall performance on this dataset. WHNet outperforms other methods in key evaluation metrics such as Precision, Recall, F1, and IoU, demonstrating strong feature representation and target recognition capabilities. Specifically, in the CHN6-CUG road extraction experiment, WHNet's F1 score reached 74.13%, at least 1.89 percentage points higher than other methods; its IoU reached 61.75%, significantly higher than the second-best method by 1.6 percentage points, indicating superior region overlap and segmentation consistency. These quantitative results demonstrate that WHNet maintains high accuracy while exhibiting good overall stability and generalization ability in road extraction tasks.

[0053] Table 1. Accuracy metrics of different network models on the CHN6-CUG dataset.

[0054] Table 2. Accuracy metrics of different network models on the Massachusetts dataset.

[0055] Table 2 summarizes the quantitative experimental results of each method on the Massachusetts dataset. As shown in Table 2, the proposed WHNet achieves the best overall performance on this dataset. In key evaluation metrics such as Precision, Recall, F1, and IoU, WHNet outperforms other comparative methods, demonstrating strong feature representation and object recognition capabilities. Specifically, in the Massachusetts road extraction experiment, WHNet's F1 score reaches 77.23%, at least 1.46 percentage points higher than other methods; its IoU reaches 64.82%, significantly higher than the second-best method by 1.49 percentage points, showing superior region overlap and segmentation consistency. In summary, the quantitative results demonstrate that WHNet maintains high accuracy while exhibiting good overall stability and generalization ability in the road extraction task.

[0056] V. Ablation Experiment To verify the effectiveness of the wavelet hybrid attention module WHAM, the cross-layer feature attention module HCCAM, and the dual-path interactive multi-scale pooling module DPIMSPM in this method, ablation experiments were conducted on the CHN6-CUG and Massachusetts datasets. The experimental results are shown in Table 3: "√" indicates that the module is retained, and "×" indicates that the module is removed.

[0057] Table 3 Ablation Experiment Results

[0058] Note: √ in the table indicates inclusion, and bold numbers represent the optimal value.

[0059] As can be seen from the experimental results in Table 3, when any of the key modules proposed in this paper are removed, the road extraction performance of the model on both datasets decreases to varying degrees. This fully demonstrates that the wavelet hybrid attention module WHAM, the hierarchical collaborative channel attention module HCCAM, and the dual-path interactive multi-scale pooling module DPIMSPM all play irreplaceable and important roles in the overall framework, and together they constitute the key support for performance improvement.

[0060] Specifically, in the three ablation experiments, the model performance also showed a significant decline when the WHAM module was removed. The F1-Score decreased by 0.38% and 0.49% on the two datasets, respectively, and the IoU decreased by 0.58% and 0.41%, respectively. This result indicates that the wavelet hybrid attention module can reconstruct and enhance features from the frequency domain perspective. By strengthening edge and detail information in high-frequency components and utilizing low-frequency components to provide structural priors, it effectively improves the model's ability to characterize fine roads, boundary regions, and complex textures.

[0061] After removing the HCCAM module, the F1-Score on the two datasets decreased by 0.71% and 0.58%, respectively, and the IoU decreased by 0.70% and 0.76%, respectively. This indicates that in the skip connection process, relying solely on simple feature transfer is insufficient to fully leverage the complementary advantages of multi-level features. However, the hierarchical collaborative channel attention module achieves hierarchical information transfer and collaborative enhancement by explicitly modeling the dependencies between cross-level features, thus playing a crucial role in the understanding of complex road topology and semantic representation.

[0062] When the DPIMSPM module was removed, the model's F1-Score decreased by 0.82% and 0.92% on the two datasets, respectively, and the IoU decreased by 0.97% and 0.85%, respectively. This result indicates that the dual-path interactive multi-scale pooling module, by fusing multi-scale deformable attention and strip convolutional features, effectively enhances the model's ability to model complex road structures and long-range dependencies, while also alleviating road breakage and information loss problems to some extent, thus significantly improving the ability to express multi-scale features.

[0063] In summary, the three modules optimize the network from different perspectives, such as multi-scale modeling, cross-layer feature interaction, and frequency domain information enhancement. They complement and synergize with each other, jointly improving the model's performance in terms of road structure integrity, detail preservation, and overall segmentation accuracy, thus verifying the rationality and effectiveness of the proposed method.

[0064] VI. Complexity Analysis The complexity of different models is analyzed, mainly comparing the number of model parameters (parameters) and the number of floating-point operations (FLOPs).

[0065] Table 4. FLOPs and parameter count of different models

[0066] The specific results are shown in Table 4. As can be seen from the comparison results in Table 4, the proposed model is at a lower-middle level in terms of parameter size compared to other existing methods. The overall parameter count is slightly higher than D-LinkNet, but lower than models such as RCFSNet and MSMDFFNet. In terms of FLOPs, this method is significantly lower than TransRoadNet and also better than the computationally expensive RCFSNet. Compared to some lightweight models such as CARNNet and D-LinkNet, although the computational cost is increased, the proposed method significantly outperforms all the compared methods in terms of road extraction accuracy.

Claims

1. A road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement, characterized in that, Using the D-LinkNet network as the basic backbone architecture, the overall network architecture consists of an encoder, a decoder, and jump connection branches connecting the encoder and decoder; The network integrates three core modules: the wavelet hybrid attention module WHAM embedded in the encoder downsampling node, the hierarchical collaborative channel attention module HCCAM configured in the skip connection branch, and the dual-path interactive multi-scale pooling module DPIMSPM deployed in the intermediate feature processing layer of the network. The wavelet hybrid attention module uses dual-tree complex wavelet transform (DTCWT) to split the input features into low-frequency components and multi-directional high-frequency components. It uses the low-frequency components as prior information to adaptively enhance the high-frequency detail features and completes feature reconstruction through inverse dual-tree complex wavelet transform (IDTCWT). The hierarchical collaborative channel attention module performs global pooling and cross-level channel cascading on the multi-level features of the encoder to generate cross-scale attention weights to complete multi-scale feature weighting calibration and resolution unification. The dual-path interactive multi-scale pooling module sets up a bar convolution branch and a multi-scale deformable attention dual parallel feature extraction path. The output features of the two paths are fused by the complementary weighted fusion module CWFM to output the final features. By relying on the collaboration of three major modules, the complete extraction of multi-scale, slender, and intersecting roads in remote sensing images can be achieved, suppressing defects such as missed detection of road edges, broken small branches, and false detection of background.

2. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 1, characterized in that, The processing steps of the wavelet hybrid attention module (WHAM) include: S1. The input coding features are decomposed by DTCWT to obtain a single low-frequency feature component and multiple high-frequency sub-band components containing real and imaginary parts; S2. The high-frequency components are compressed into channels by two-stage cascaded 1×1 convolutions and activated by Sigmoid to generate the first attention weights. The weights are multiplied by the original high-frequency components to complete the initial high-frequency feature enhancement. S3, low-frequency components, and high-frequency components after initial enhancement are respectively subjected to 3×3 convolution operation and then added element by element to achieve frequency domain feature fusion. The fused features are then subjected to 1×1 convolution and Sigmoid to generate second attention weights, which are used for secondary adaptive modulation of high-frequency components. S4. The modulated high-frequency component and the original low-frequency component are fed together into the IDTCWT inverse transform to reconstruct the feature map with enhanced output details.

3. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 1, characterized in that, The HCCAM (Hierarchical Collaborative Channel Attention Module) adapts to the three-layer adjacent coding features of the U-shaped network. The input module processing flow is as follows: Global average pooling (GAP) is performed on the input features of each level to extract channel semantic vectors. All channel vectors are concatenated along the channel dimension to obtain a joint channel representation. The joint channel representation is fed into a lightweight attention network and normalized by Softmax to generate cross-level attention weight coefficients. The weight coefficients are used to weight and modulate the corresponding level features, and all features are scaled to the same resolution through bilinear interpolation.

4. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 3, characterized in that, In the aforementioned hierarchical collaborative channel attention module HCCAM, the fused features after resolution unification are output in three multi-resolution convolutional branches. The three output features are adapted to the different levels of input of the decoder, and simultaneously supplement the decoder with three types of feature information: shallow details, middle structure, and high-level semantics.

5. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 1, characterized in that, The dual-path interactive multi-scale pooling module DPIMSPM is divided into two parallel feature extraction branches: the first branch is a multi-scale strip convolution branch, and the second branch is a multi-scale deformable attention branch. The strip convolution branch uses six long strip asymmetric convolution kernels {1×3, 3×1, 1×5, 5×1, 1×7, 7×1} to extract features in parallel, which is suitable for the long strip geometric shape of remote sensing roads. The output feature channels of each convolution are concatenated and then subjected to lightweight attention to generate dynamic fusion weights. The weighted fusion is then used to obtain the bar feature. .

6. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 5, characterized in that, The multi-scale deformable attention branch uses multiple sets of deformable convolutions with different sized kernels to construct attention units in parallel. It relies on learnable offsets to adaptively adjust the feature sampling position, thereby achieving long-distance context modeling of curved, occluded, and broken roads. Deformable features are obtained by splicing multi-scale deformable feature channels and then fusing them with lightweight dynamic weights. .

7. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 6, characterized in that, The complementary weighted fusion module CWFM is implemented. and The fusion method is as follows: the two feature streams are concatenated channel by channel and then fed into a depthwise separable convolution, combined with Sigmoid activation to generate a complementary weight map. The complementary weights are then used to apply the complementary weights to the following functions: The weighted, fine-structured regions are dominated by bar features, while the global semantic regions are dominated by deformable features. The weighted summation outputs the final features of the module.

8. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 1, characterized in that, The network training phase uses the Adam optimizer, an initial learning rate of 0.0002, and a batch size of 4. If the loss does not decrease for three consecutive rounds, the learning rate is halved. The maximum number of epochs is 200. Training is terminated early if the loss is higher than the historical best loss or the learning rate is lower than 0.0000005 for six consecutive rounds.

9. The road extraction network based on wavelet hybrid attention and hierarchical feature synergy enhancement according to claim 8, characterized in that, The network loss function described uses a weighted joint loss function of binary cross-entropy (BCE) and Dice loss. BCE is used to alleviate the imbalance between road and background pixel categories in remote sensing images, while Dice loss optimizes the segmentation accuracy of small target roads with fine branches.

10. The method for road extraction using the wavelet hybrid attention and hierarchical feature co-enhancement network as described in any one of claims 1-9, characterized in that, Includes the following steps: S1. Image preprocessing: The original remote sensing images are uniformly cropped to 512×512 pixels using a non-overlapping cropping method. For areas where the image edges are smaller than this size, edge pixel padding is used to fill in the gaps. Data augmentation operations such as horizontal flipping, vertical flipping, and random scaling are performed on the training set images to improve the model's generalization ability. S2. Model Training: Based on the PyTorch framework, using the CHN6-CUG and Massachusetts public remote sensing road datasets, the WHNet network is iteratively trained according to the preset training parameters, and the optimal weight model of the validation set is saved. S3. Image Inference and Extraction: The preprocessed remote sensing image to be detected is input into the trained optimal WHNet model. The model outputs a binary road segmentation map through forward inference, completing the fully automatic and accurate extraction of remote sensing roads. The output results can be directly applied to scenarios such as urban surveying, traffic monitoring, and autonomous driving geographic information collection.

Images