Remote sensing image road extraction method fusing multi-scale and road topology correlation characteristics

Abstract

The invention provides a remote sensing image road extraction method fusing multi-scale and road topological correlation characteristics, which comprises the following steps: firstly, establishing a remote sensing image road extraction model, and then importing preprocessed data into the model to output a result; according to the remote sensing image road extraction model, a backbone network is replaced by a lightweight MobileNetv3, so that the calculation overhead and the model complexity are reduced. Secondly, designing a multi-layer cascade cavity space pyramid pooling module, fusing a convolution attention mechanism, and enhancing the extraction capability and topological connectivity of fine and weak road features by expanding a receptive field; and finally, in combination with a mixed loss function of Focal Loss and Dice Loss, the problem of class imbalance between the road and the background is relieved, environmental interference is inhibited, and road feature extraction is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of geographic information science and technology, and in particular relates to a method for extracting roads from remote sensing images that integrates multi-scale and road topological features. Background Technology

[0002] Road extraction has long been a research hotspot in the application of high-resolution remote sensing data. High-resolution optical remote sensing images cover a wide area, and intelligent road extraction based on such images and combined with semantic segmentation technology has significant application value in urban planning, electronic map production, and emergency disaster response. Current research focuses on how to accurately and comprehensively extract road networks from large-scale remote sensing images. However, due to the influence of diverse road surface materials, lighting variations, and building occlusion, automatic road extraction and network construction remain highly challenging tasks in the field of remote sensing image analysis. Existing road extraction methods can be mainly divided into traditional methods and deep learning-based methods.

[0003] Traditional road extraction methods can be broadly categorized into pixel-based and object-oriented approaches. Pixel-based methods primarily rely on the spectral features of pixels for road extraction, including spectral analysis, thresholding, and edge detection. While these methods can effectively extract basic road networks from clear, homogeneous remote sensing images, they generally lack consideration for the feature space background and texture structure, are often significantly affected by salt-and-pepper noise, and are sensitive to shadows of other features or regions with similar texture features to roads, leading to classification errors. Object-oriented extraction methods, on the other hand, treat roads as independent objects and establish corresponding information models. Compared to pixel-based methods, they exhibit better noise resistance and applicability. Typical algorithms include watershed segmentation, region growing algorithms, and support vector machine-based algorithms. Although these object-oriented classification methods utilize the overall information of the target and have better noise resistance, they are more prone to confusion with spatially adjacent and similar-shaped features, resulting in a "sticking" phenomenon and lower accuracy. Overall, traditional methods have certain limitations in generalization ability.

[0004] Deep learning enables automatic feature learning and image recognition. Its adaptability and robustness allow neural networks to autonomously extract geospatial information. Deep learning models can extract road features end-to-end, overcoming the problems of low automation and poor generalization ability of traditional methods. When processing large-scale remote sensing imagery, they significantly outperform traditional methods in terms of accuracy and efficiency. Current deep learning methods are mainly divided into two categories: Convolutional Neural Networks (CNNs) and Transformers. Among them, CNNs possess powerful feature learning and pattern recognition capabilities, effectively processing complex remote sensing imagery. Ronneberger et al.'s U-Net enhances edge detail capture capabilities through a symmetric encoder-decoder structure; Chaurasia et al.'s LinkNet combines lightweight design with residual connections; and Google's DeepLab series of models in semantic segmentation have achieved significant breakthroughs in handling complex image segmentation tasks. The DeepLabV3+ model uses the Xception network as its backbone architecture, significantly improving computational efficiency and feature extraction capabilities through depthwise separable convolutions. Chao et al.'s DOCSwin Trans, employing cross-window deformation attention and adaptive resampling strategies, performs excellently on complex road imagery. However, Transformer-based models face challenges due to their large number of parameters and high computational requirements caused by multi-head self-attention mechanisms and stacked architectures. SegFormer captures the global topology of roads through a hierarchical Transformer encoder and achieves multi-scale fusion while reducing hardware requirements with a lightweight decoder. D-SegNeXt innovatively introduces depthwise separable convolution and dynamic channel attention, which significantly improves the ability to distinguish small targets such as rural roads. HF-HRNet proposed by Zhang et al. uses cascaded dilated convolution and multi-scale enhancement modules, and performs excellently in human pose estimation and semantic segmentation.

[0005] Despite significant progress in deep learning methods, challenges remain: CNNs have limited receptive fields, making it difficult to model global topology; Transformers have high computational complexity; and multi-scale fusion can easily lead to semantic gaps. Furthermore, factors such as building shadows, vegetation variations, road topological complexity, and spectral similarity in high-resolution imagery can cause classification ambiguity, and the trade-off between feature representation and computational efficiency remains an unresolved issue. Summary of the Invention

[0006] The purpose of this invention is to provide a remote sensing image road extraction method that integrates multi-scale and road topology features. This method solves the problems of poor connectivity, insufficient fusion, and incomplete boundaries caused by diverse materials, lighting variations, occlusion, and complex topology in high-resolution images, and achieves high-precision, robust, and automated extraction and topology construction of road networks.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for road extraction from remote sensing images that integrates multi-scale and road topological correlation features, comprising the following steps:

[0008] S1. Establish a high-resolution remote sensing image road extraction model that integrates multi-scale and road topology features;

[0009] S2. Standardize and resize the semantic segmentation data of remote sensing images;

[0010] S3. Import the image data obtained in S2 into the high-resolution remote sensing image road extraction model obtained in S1 for extraction, and output the prediction results.

[0011] Furthermore, S1 specifically includes the following steps:

[0012] S101. Improve the backbone network of DeepLabV3+ by replacing the original backbone network with the MobileNetV3 backbone network and introducing a squeeze-excitation module.

[0013] S102. Establish a multi-layered cascaded hollow spatial pyramid pooling module that integrates convolutional attention modules;

[0014] S103. Establish a composite loss function of Focal Loss and Dice Loss.

[0015] Furthermore, the MobileNetV3 backbone network described in S101 includes depthwise separable convolutional layers; the depthwise separable convolution decomposes the convolution operation into two independent steps: channel-wise convolution and pointwise convolution; wherein channel-wise convolution achieves lightweight filtering by applying a single convolutional kernel to each input channel; and pointwise convolution constructs new features by calculating a linear combination of the input channels.

[0016] Furthermore, the squeeze-excitation module described in S101 generates channel weight vectors through global average pooling operations, and learns the adaptive weight coefficients of each feature channel through a fully connected layer.

[0017] Furthermore, the MobileNetV3 backbone network described in S101 uses the h-swish activation function instead of the traditional ReLU6 activation function:

[0018]

[0019] Where x is the input feature value.

[0020] Furthermore, the multi-layer cascaded hollow spatial pyramid pooling module includes four 3×3 depth-separable convolutional layers and a convolutional attention module. The depth-separable convolutional layers include a hole convolution operation, and the hole expansion rate increases stepwise.

[0021] Furthermore, the multi-layered cascaded void space pyramid pooling module specifically comprises:

[0022]

[0023] in, y is a convolution operation with kernel size k and dilation rate v, and y is the output of a multi-layer cascaded hollow spatial pyramid pooling module;

[0024] The dilated convolution operation specifically involves:

[0025]

[0026] Where x is the one-dimensional input, K is the kernel size, v is the dilation rate, w(k) represents the k-th parameter in the kernel, y′(i) is the output of the one-dimensional dilated convolution, and i is the index symbol;

[0027] The size of the receptive field R of the dilated convolution is:

[0028] R = (k-1) × (v-1) + k (4)

[0029] For cascaded porous convolutional layers R1, R2, ... R n , n represents the nth dilated convolutional layer being cascaded, with a cumulative receptive field R. max The calculation formula is:

[0030] R max =R1+R2+…R n -(n-1) (5).

[0031] Furthermore, the convolutional attention module is embedded in each layer of the convolutional operation of the remote sensing image road extraction model; it is applied again after the feature fusion operation to form a multi-stage attention enhancement structure that runs through the forward propagation path of the network.

[0032] Furthermore, the convolutional attention module includes a channel attention module and a spatial attention module; the channel attention submodule determines the weight vector M through global average pooling and a fully connected layer. c (F):

[0033] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (6)

[0034] Where F represents the input feature map; AvgPool and MaxPool represent average pooling and max pooling, respectively; MLP represents a multilayer perceptron; and σ represents the sigmoid activation function.

[0035] The spatial attention submodule uses 7×7 convolution to generate a spatial weight matrix M. s (F):

[0036] M s (F)=σ(f 7×7 ([AvgPool(F);MaxPool(F)])) (7)

[0037] Among them, f 7×7 It is a 7×7 convolution.

[0038] Furthermore, the composite loss function of Focal Loss and Dice Loss is specifically as follows:

[0039] L = L d +L f (8)

[0040] Among them, L d L represents Dice loss. f Indicates Focal loss;

[0041] The Dice loss, DiceLoss, is:

[0042] DiceLoss=1-(2*|X∩Y| / (|X|+|Y|)) (9)

[0043] Where X represents the segmented region of the predicted result, Y represents the ground truth region, and X∩Y represents the overlapping part between the predicted result and the ground truth region;

[0044] The Focal loss, FocalLoss, is:

[0045] Focal Loss=-α*(1-pt)∧γ*log(pt) (12)

[0046] Where pt represents the classification confidence level; α is the positive sample weight coefficient; and γ is the hard sample focusing factor.

[0047] Compared with existing technologies, the advantages of this invention are as follows: This invention proposes a novel multi-level cascaded void spatial pyramid module (MCASPP), the core of which lies in introducing dense skip connections and a progressive void ratio configuration strategy (r = 6, 12, 18, 24) to construct a continuously expanding global receptive field. This module achieves progressive contextual modeling through a cross-level feature reuse mechanism, effectively enhancing the model's scale adaptability and structure perception capability for road targets. Simultaneously, a convolutional attention module (CBAM) is embedded synchronously in the MCASPP and decoder stages to construct a dual-domain attention collaborative mechanism, strengthening the activation of road texture features and suppressing non-road structural interference in the channel and spatial domains respectively, thereby significantly improving the accuracy and boundary integrity of road recognition in complex backgrounds.

[0048] In terms of network architecture design, this invention replaces the original Xception backbone network in DeepLabV3+ with a lightweight MobileNetV3 network that integrates a compression-enhanced attention mechanism (SE) and an h-swish activation function. This improved method, by introducing an adaptive channel attention mechanism and advanced activation functions, significantly reduces the number of model parameters and computational complexity while enhancing the feature representation ability of low-contrast, weakly textured road regions, and improving the model's robustness in complex remote sensing scenarios such as uneven lighting and shadow occlusion.

[0049] Furthermore, this invention addresses the problems of unbalanced pixel distribution, blurred boundaries, and severe occlusion in remote sensing images by proposing a Focal-Dice joint loss function. This loss function integrates the weighting mechanism of Focal Loss for hard-to-classify samples with the optimization capability of Dice Loss for region continuity. By dynamically adjusting the training weights of easily confused and hard-to-classify samples, it enhances the model's attention to road topological break regions and blurred boundary regions, thereby alleviating the gradient vanishing problem caused by class imbalance and improving the model's convergence stability and the connectivity and completeness of road extraction. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a network structure diagram of this implementation method;

[0052] Figure 2 This is a schematic diagram of a depth-separable convolutional layer in this embodiment;

[0053] Figure 3 This is a structural diagram of the squeeze-excitation attention mechanism in this embodiment;

[0054] Figure 4 This is a comparison diagram of the MCASPP module structure in this embodiment and the traditional ASPP model structure; where (a) is the ASPP model structure; and (b) is the MCASPP module structure in this embodiment.

[0055] Figure 5 This is a schematic diagram of the maximum receptive field that MCASPP can obtain in this embodiment;

[0056] Figure 6 This is a schematic diagram of the convolutional attention module structure in this embodiment;

[0057] Figure 7 This is a qualitative analysis of the results extracted from the DeepGlobe dataset by different methods; where (a) is region 1, (b) is region 2, (c) is region 3, (d) is region 4, (e) is region 5, and (f) is region 6.

[0058] Figure 8 This is a qualitative analysis of the extraction results of different methods on the CHN6-CUG dataset; where (a) is region 1, (b) is region 2, (c) is region 3, (d) is region 4, and (e) is region 5.

[0059] Figure 9 The effects of different modules on the segmentation results are shown; (a) is the image, (b) is the real road, (c) is ASPP, (d) is MCASPP, and (e) is MCASPP+CBAM. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] This embodiment provides a method for road extraction from remote sensing images that integrates multi-scale and road topology features; it achieves high-precision road extraction in complex scenes while maintaining topological continuity.

[0062] To achieve the above objectives, this implementation establishes a remote sensing image road extraction model. It achieves accurate perception and geometric representation of road features by integrating a dynamic multi-scale feature fusion mechanism and an attention-guided optimization strategy. First, the DeepLabV3+ backbone network is improved by optimizing it into a lightweight MobileNetV3 structure with an integrated attention mechanism, reducing computational overhead and model complexity, effectively compressing computational resources and improving feature extraction performance. Second, a multi-level Cascade Atrous Spatial Pyramid Pooling (MCASPP) structure is established, integrating convolutional attention modules, to expand the receptive field while enhancing the extraction capability and connectivity of subtle road features. Finally, a loss function combining Focal Loss and Dice Loss is used to effectively mitigate the class imbalance between road and background samples, suppress environmental interference, and enhance the extraction capability of road features.

[0063] Based on the aforementioned remote sensing image road extraction model, this embodiment provides a remote sensing image road extraction method: such as Figure 1 As shown, the input image is first processed by the MobileNetV3 backbone network to extract deep and shallow features. The deep features are then processed sequentially by the MCASPP module, with a Convolutional Block Attention Module (CBAT) embedded after each convolutional layer to enhance multi-scale semantic representation and better preserve road detail information. Subsequently, a 1×1 convolution operation is performed to integrate the deep features. The shallow features are refined in the decoder using the CBAT to minimize information loss, followed by another 1×1 convolution operation. The processed deep features are then upsampled by a factor of 4, fused with the processed shallow features, and compressed using a 3×3 convolution. Finally, bilinear interpolation with a factor of 4 upsampling generates a final prediction result of the same size as the original image, which is then output. This implementation systematically optimizes feature extraction, context modeling, and attention-guided processing, effectively integrating multi-level semantic information and significantly improving road extraction performance.

[0064] In some exemplary implementations, the remote sensing image semantic segmentation data used above includes the DeepGlobe Road Extraction Dataset and the CHN6-CUG Road Dataset. The DeepGlobe Road Extraction Dataset originates from the road dataset of the DeepGlobe Satellite Imagery Understanding Challenge. This dataset contains high-resolution satellite imagery covering various geographic regions, seasonal variations, and weather conditions, encompassing urban blocks, rural areas, highways, and mountainous terrain, and is widely considered a benchmark dataset for road extraction tasks. The dataset contains 6,226 labeled images with a spatial resolution of 0.5 meters, randomly divided in an 8:1:1 ratio into 4,314 training images, 623 validation images, and 623 test images.

[0065] The CHN6-CUG road dataset is a large-scale satellite image set covering multiple urban areas in China. All images maintain a spatial resolution of 0.5 meters and are labeled with various road types, including railways, highways, urban roads, and rural paths. It contains 2,400 training images, 300 test images, and 300 validation images. In this implementation, the loaded images and their corresponding labels are normalized and resized. Input images are resized to 512×512 pixel JPG format, while ground reality labels are in PNG format, with background pixels labeled as 0 and road areas labeled as 1.

[0066] In some specific implementations, the MobileNetV3 backbone network reduces computational load by employing depthwise separable convolutional layers, thereby improving the performance of the remote sensing image road extraction model. Unlike traditional convolutional layers, depthwise separable convolution decomposes the convolution operation into two independent steps: channel-wise convolution and pointwise convolution. Channel-wise convolution achieves lightweight filtering by applying a single convolutional kernel to each input channel, while pointwise convolution constructs new features by calculating a linear combination of the input channels. Assume the width, height, and number of channels of the input feature matrix are W, H, and C, respectively, and the width, height, and number of convolutional kernels are K. w K h and K n (Step size set to 1, output depth to K) h If the parameter computation of ordinary convolution is P1, then the computational cost P1 is:

[0067] P1=W×H×C×K w ×K h ×K n (1)

[0068] The computational cost of depthwise separable convolution parameters, P2:

[0069] P2=W×H×C×K w ×K h×1+W×H×C×K n ×1×1 (2)

[0070] This allows us to calculate:

[0071]

[0072] According to formula (3), as the number of output channels and the number of convolution kernels increase, the number of parameters in depthwise separable convolution will decrease significantly. The specific calculation process of depthwise separable convolution in this embodiment is as follows: Figure 2 As shown.

[0073] Furthermore, this implementation introduces a squeeze-and-excitation (SE) module to improve the robustness of road feature extraction based on a lightweight architecture using the MobileNetV3 backbone network. For example... Figure 3 As shown, the squeeze-excitation module described in this embodiment generates channel weight vectors through global average pooling and learns adaptive weight coefficients for each feature channel through a fully connected layer, thereby achieving dynamic enhancement of feature channels that are strongly correlated with road texture, effectively suppressing background noise and interference from irrelevant structures, and significantly improving the model's feature discrimination ability and semantic expression accuracy in complex remote sensing scenarios.

[0074] In some possible implementations, the MobileNetV3 backbone network uses the h-swish activation function instead of the traditional ReLU6 activation function:

[0075]

[0076] Where x is the input feature value;

[0077] This activation function effectively avoids gradient saturation, a common problem in depthwise separable convolutions, while reducing inference latency on computing devices. Furthermore, MobileNetV3 enhances its adaptability to narrow roads and low-contrast scenes by adaptively adjusting the channel expansion ratio.

[0078] This implementation improves upon the existing Atrous Spatial Pyramid Pooling (ASPP) structure by proposing a Multi-level Cascaded Atrous Spatial Pyramid Pooling (MCASPP) module. This overcomes the limitations of traditional ASPP in remote sensing image road extraction tasks, such as limited ability to model multi-scale contextual information, low feature utilization, and difficulty in constructing continuous road topology structures. Figure 4As shown in (a), the traditional ASPP structure usually consists of three parallel dilated convolution branches with different dilation rates (r = 6, 12, 18) and a global average pooling branch. Although this structure can capture multi-scale semantic information to a certain extent, it still has defects such as discontinuous receptive field coverage and insufficient feature fusion hierarchy when processing high-resolution remote sensing images, which can easily lead to loss of road details and topological breaks.

[0079] Therefore, this embodiment proposes a multi-layered cascaded void space pyramid pooling module (such as...) Figure 4 (b) As shown, a cascaded stacked design is adopted, which introduces a dense connection mechanism to achieve cross-layer feature reuse and information sharing, forming a progressive feature dependency relationship between adjacent dilated convolutional layers. Specifically, the multi-layer cascaded dilated spatial pyramid pooling module includes four 3×3 depth-separable convolutional layers, each including a dilated convolution operation with a progressively increasing dilation rate (r = 6, 12, 18, 24), thereby constructing a continuously expanding receptive field and effectively enhancing the model's ability to perceive road targets at different scales. In addition, to improve the learning efficiency of key road features, this embodiment of the multi-layer cascaded dilated spatial pyramid pooling module also includes a convolutional attention module that integrates an attention mechanism. This module collaboratively optimizes feature responses in the channel and spatial domains, strengthens the activation performance of road texture regions, and suppresses background noise and irrelevant structural interference, thereby significantly improving the accuracy and topological connectivity of road extraction.

[0080] In some specific implementations, the multi-layered cascaded void space pyramid pooling module is specifically represented as follows:

[0081]

[0082] in, y is a convolution operation with kernel size k and dilation rate v, and y is the output of a multi-layer cascaded hollow spatial pyramid pooling module.

[0083] Dilated convolution expands the receptive field without changing the feature map resolution, enabling each convolution output to integrate the expanded spatial context; the dilated convolution operation of the depthwise separable convolutional layer can be represented as:

[0084]

[0085] Here, the one-dimensional input is defined as x, K is the kernel size, v is the dilation rate, w(k) represents the k-th parameter in the kernel, y′(i) is the output of the one-dimensional dilated convolution, and i is the index symbol.

[0086] The formula for calculating the receptive field size of dilated convolution is:

[0087] R = (k-1) × (v-1) + k (7)

[0088] For cascaded porous convolutional layers R1, R2, ..., R n , where n represents the nth dilated convolutional layer being cascaded, and the formula for calculating the cumulative receptive field is:

[0089] R max =R1+R2+…R n -(n-1) (8)

[0090] According to formulas (7) and (8), the maximum receptive field of ASPP is significantly improved after adopting multi-layer cascaded expansion of the receptive field; for example Figure 5 As shown, if R1 and R2 represent convolutions with dilation rates of 6 and 12, respectively, and their receptive fields are 13 and 25, then when these two convolutions are concatenated, their original ASPP receptive fields are...

[0091] R(6,12)=13+25-(2-1)=37;

[0092] The maximum receptive field in this implementation method is:

[0093] R(6,12,18)=R(6,12)+R(18)

[0094] =37 + 37 - (2 - 1)

[0095] =R(6)+R(12)+R(18)

[0096] =13+25+37-(3-1)

[0097] =73

[0098] In this implementation, the ASPP with multi-layer cascaded expansion of the receptive field reaches a maximum receptive field of 73, exceeding the original ASPP's 37. According to formula (8), the receptive field achievable by MCASPP is as follows: Figure 5 As shown.

[0099] In some specific implementations, the convolutional attention module significantly improves the semantic segmentation accuracy of the remote sensing image road extraction model by prioritizing key feature channels during the encoding stage. This implementation embeds the convolutional attention mechanism into each layer of the high-resolution remote sensing image road extraction model that integrates multi-scale and road topological features. This mechanism is further applied after feature fusion (such as skip connections), forming a multi-stage attention enhancement structure that runs through the network's forward propagation path. Through these improvements, the remote sensing image road extraction model can dynamically optimize the channel weight distribution of the feature map, enhancing its ability to learn road target features while effectively suppressing interference from complex background areas and irrelevant ground features, thus improving the model's ability and robustness in depicting road topology and boundary details.

[0100] like Figure 6 In this embodiment, the convolutional attention module includes a channel attention module and a spatial attention module, highlighting key features in two dimensions to obtain more accurate recognition results. The channel attention submodule calculates the weight vector through global average pooling (GAP) and a fully connected layer:

[0101] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (9)

[0102] Where F represents the input feature map; AvgPool / MaxPool are global pooling operations (average pooling and max pooling), where AvgPool calculates the average value of all pixels in each channel of the feature map, and MaxPool extracts the maximum value of all pixels in each channel of the feature map; MLP stands for Multilayer Perceptron; and σ is the Sigmoid activation function. The channel attention mechanism adaptively enhances the channel responses related to the road (such as linear texture and grayscale continuity) while suppressing the activation of interfering factors such as vegetation and buildings.

[0103] The spatial attention submodule uses 7×7 convolutions to generate the spatial weight matrix:

[0104] M s (F)=σ(f 7×7 ([AvgPool(F);MaxPool(F)])) (10)

[0105] Where [AvgPool(F); MaxPool(F)] represents channel cascading; f 7×7 The convolution is 7×7. This mechanism focuses network attention on regions that conform to road morphological features while suppressing isolated noise activation. In intersection scenarios, spatial attention can focus on the core area of ​​road intersections, preventing topological fragmentation. MCASPP constructs a continuous receptive field through cascaded dilated convolutions, while the convolutional attention module adaptively enhances road regions and attenuates background noise. Together, they achieve accurate road feature perception consistent with the extension characteristics of road strips.

[0106] The Dice loss function is widely used in image segmentation tasks, especially for binary classification problems. In remote sensing road extraction tasks, since road pixels account for only a very small proportion of the entire image, the Dice loss can effectively highlight the overlap between the predicted and ground truth regions. This implementation enhances edge preservation by maximizing the intersection-union ratio (IU / U) metric, thereby mitigating overfitting and improving the model's generalization performance. The Dice loss is:

[0107] DiceLoss=1-(2*|X∩Y| / (|X|+|Y|)) (11)

[0108] Where X represents the segmented region of the predicted result, Y represents the ground truth region, and X∩Y represents the overlap between the predicted result and the ground truth region. Focal loss addresses class imbalance by assigning higher weights to difficult-to-classify samples. In road extraction tasks, due to the sparse distribution of road pixels, this mechanism guides the model to focus on difficult samples, thereby improving detection accuracy. The Focal loss is:

[0109] Focal Loss=-α*(1-pt)∧γ*log(pt) (12)

[0110] Where pt represents the classification confidence level; in this implementation, the positive sample weight coefficient α = 0.5 and the difficult sample focusing factor γ = 2 are set.

[0111] To address these issues, this implementation combines the advantages of Dice and Focal losses, proposing an improved loss function to resolve class imbalance, boundary ambiguity, and gradient vanishing problems in road extraction. The improved loss function is defined as follows:

[0112] L = L d +L f (13)

[0113] Among them, L d L represents Dice loss. f This indicates Focal loss.

[0114] The main remote sensing image semantic segmentation datasets used in this implementation include the DeepGlobe Road Extraction dataset and the CHN6-CUG Road dataset. The DeepGlobe Road Extraction dataset originates from the road dataset of the DeepGlobe Satellite Image Understanding Challenge. This dataset contains high-resolution satellite imagery covering various geographical regions, seasonal variations, and meteorological conditions, encompassing urban blocks, rural areas, highways, and mountainous terrain, and is widely considered a benchmark dataset for road extraction tasks. The dataset contains 6,226 labeled images with a spatial resolution of 0.5 meters, randomly divided into 4,314 training images, 623 validation images, and 623 test images in an 8:1:1 ratio. The CHN6-CUG Road dataset is a large-scale satellite image set covering multiple urban areas in China. All images maintain a spatial resolution of 0.5 meters and are labeled with various road types, including railways, highways, urban roads, and rural paths. It contains 2,400 training images, 300 test images, and 300 validation images. During the experiment, the loaded images and their corresponding labels were normalized and resized. The input image was resized to a 512×512 pixel JPG format, while the ground reality labels were in PNG format, with background pixels marked as 0 and road areas marked as 1.

[0115] This implementation uses a deep learning framework built on PyTorch, Python 3.7.10, and CUDA 11.7. The server configuration is as follows: the operating system is Windows Server*2019, using a 64-bit operating platform; the CPU includes two Intel(R) Xeon(R) Silver 4210 CPUs; the GPU is an NVIDIA GeForce RTX 2080Ti with 11GB of GDDR6 memory; the random access memory (RAM) includes four 32GB DDR4 2400 MHz memory modules, totaling 128GB of memory; the main training parameters are shown in Table 1.

[0116] Table 1 Key parameters for model training

[0117] Parameter name Numeric / Type Input image size 512×512 Training batch size 4 Maximum training cycle 200 Optimizer Adam Maximum learning rate 5e-4 Learning rate adjustment methods Cos num_workers 2 momentum 0.9

[0118] Evaluation metrics are used to evaluate the detection performance of semantic segmentation models. In the field of target extraction from remote sensing images, commonly used evaluation metrics include accuracy, average pixel precision (mPA), and average intersection-union ratio (mIoU). Their calculation formulas are shown in equations (14) to (16), respectively:

[0119]

[0120] Where TP represents the number of correctly identified road pixels, FP represents the number of incorrectly identified road background pixels, FN represents the number of incorrectly identified road pixels as background pixels, and n+1 and n are the category index values.

[0121]

[0122] Where k+1 is the number of categories, p ii p is the number of pixels correctly classified. ij Let $j$ be the number of pixels that are predicted to be class $i$. $PA$ represents the ratio between the number of correctly classified pixels in each class and the total number of pixels. $mPA$ represents the average of the $PA$ calculated for each class.

[0123]

[0124] Where TN represents the number of background pixels that were correctly identified.

[0125] To further validate and evaluate the method of this implementation, the DeepGlobe and CHN6-CUG datasets were selected. The DeepSegment implementation of this method was compared with several practical and efficient methods, including U-Net, DeepLabV3+, NL-LinkNet, D-SegNeXt, and Segformer. These models showed varying performance across different metrics in the experiments on the DeepGlobe and CHN6-CUG datasets, as shown in Table 2. Segformer-b0 had the lowest extraction accuracy because it's the most lightweight version, with only 3.72M parameters and a single image inference time of just 26ms, designed for real-time inference and mobile deployment. Segformer with B5 as its backbone, however, enhances contextual inference modeling by increasing the receptive field, achieving a 6.55% improvement in mIoU compared to B0 on ​​the DeepGlobe dataset. U-Net wasn't designed specifically for road extraction, resulting in relatively low accuracy on both datasets. DeepLabV3+, leveraging the multi-scale spatial feature learning capabilities of its ASPP module, outperformed U-Net in capturing road features, achieving relatively high accuracy on the CHN6-CUG dataset. However, its large parameter count limits its ability to handle road continuity. D-LinkNet, designed for high-resolution satellite imagery road extraction, won first place in the CVPR DeepGlobe 2018 Road Extraction Challenge. NL-LinkNet improves upon this by embedding a non-local attention module at the end of the encoder and employing adaptive dilation rate convolution to enhance its ability to model road topological continuity. However, experiments show that it still has limitations in modeling road hierarchical topological relationships. On the DeepGlobe dataset, D-SegNeXt achieves good extraction results, second only to the model proposed in this implementation. On the CHN6-CUG dataset, D-SegNeXt has the best accuracy at 96.52%, demonstrating its suitability for road extraction tasks. The DeepSegment method in this implementation improves road detail extraction and road network integrity by fusing a dilated spatial pyramid pool and an attention mechanism. On the DeepGlobe dataset, its mPA, accuracy, and mIoU are 89.26%, 98.2%, and 81.62%, respectively; on the CHN6-CUG dataset, they are 88.59%, 96.16%, and 82.52%, respectively. Compared to other methods, the model in this implementation shows improvements across various metrics. Furthermore, compared to mainstream networks, it has a lower parameter count, consumes fewer resources, and has a faster inference time.

[0126] Table 2 Quantitative Analysis of Extraction Results from Different Models

[0127]

[0128] In addition to the quantitative analysis of the experimental results mentioned above, this implementation method selects a relatively typical scene from the dataset for visual qualitative analysis. While these models extracted relatively complete roads in the DeepGlobe and CHN6-CUG datasets, they lacked processing of road details and all contained a small number of errors. The method in this implementation method has the lowest error rate. The extraction results are as follows: Figure 7 (a)~(f) and Figure 8 As shown in (a) to (e), the road surface extraction results of U-Net, DeepLabV3+, NL-LinkNet, and Segformer all exhibit multiple instances of misclassification, omissions, and disconnections, and their noise resistance is relatively weak, making them susceptible to interference from class imbalance factors. D-SegNeXt possesses a certain ability to integrate contextual semantic information and can handle road extraction tasks in complex scenes, but the extracted roads still have multiple gaps or omissions. In areas obscured by vegetation and in some complex scenes, such as... Figure 7 In the field paths shown in (b) and (f), only DeepSegment and D-SegNeXt extracted the road portion relatively completely; Figure 8 In (a), only DeepSegment, D-SegNeXt, and Deeplabv3+ extracted the vegetation-covered paths; in the DeepGlobe dataset, U-Net, DeepLabV3+, NL-LinkNet, and Segformer showed low attention to fine road features, with Deeplabv3+ exhibiting more misclassifications, such as... Figure 7 (a), (c); However, in the complex urban environment of the CHN6-CUG dataset, U-Net, DeepLabV3+, NL-LinkNet, and Segformer exacerbate inter-class confusion when the spectral information of roads and surrounding features is similar, making misclassification and underclassification more likely. For example... Figure 8 In (c), U-Net and DeepLabV3+ misclassify a main road into two; Figure 8 In (d) and (e), main roads were missed in classification. Therefore, compared with many current mainstream methods, the method of this implementation can extract the road structure and its topological relationship features more completely, and can handle road extraction tasks in various complex geographical scenarios.

[0129] To verify the performance of each improvement step in this implementation, we designed five scenarios on the DeepGlobe dataset under the same environment. Experiment 1: DeepLabV3+ original network; Experiment 2: DeepLabV3+ model using MobileNetV3 as the backbone network; Experiment 3: Based on Experiment 2, replacing the ASPP module with the MCASPP module; Experiment 4: Based on Experiment 3, inserting the CBAM attention mechanism; Experiment 5: Based on Experiment 4, using our improved loss function. Specific ablation experiment designs and results are shown in Table 3.

[0130] Table 3 Ablation Experiment Design

[0131]

[0132] A comparison of Experiments 1 and 2 shows that using the lightweight MobilenetV3 backbone network slightly reduces the accuracy of road extraction. However, the method in this implementation maximizes model performance by sacrificing the complex backbone network and introducing appropriate attention mechanisms and structural optimizations. A comparison of Experiments 2 and 3 shows that using MCASPP is more adaptable to complex road extraction tasks, and Experiment 4 significantly improves road extraction accuracy by fusing CBAM and MCASPP. Experiment 5 demonstrates the effectiveness of our improved loss function in road extraction tasks. To further verify the effectiveness of the proposed method, this implementation qualitatively analyzes the following two scenarios, and the results are as follows: Figure 9 As shown in (a)-(e). Panel (a) shows the original image data, (b) shows the ground truth labels, (c) shows the results using the ASPP module, (d) shows the results using the MCASPP module, and (e) shows the results combining the MCASPP module and the convolutional block attention module.

[0133] Figure 9 The comparison between (c) and (d) shows that the proposed combination of MCASPP and CBAM effectively integrates feature information from detail and semantic branches, thereby enabling the extraction of a more comprehensive and continuous road region. Figure 9 In (c), although the ASPP module roughly captures the road structure through its multi-scale spatial feature learning ability, significant segmentation errors and omissions still exist due to the limitations of its receptive field and relatively low dilation convolution expansion rate. Furthermore, ASPP easily confuses road segments with similar spectral features to the urban background. In contrast, Figure 9 The multi-level cascaded ASPP module in (d) alleviates this problem to some extent. However, in Region 2, due to the complexity of the road environment, it is still difficult to accurately extract the complete road structure. Figure 9(e) demonstrates the integration of the MCASPP module with CBAM, which significantly enhances the model's ability to learn road features, particularly in enhancing details and edge features, resulting in more accurate perception and geometric representation of road elements.

[0134] This implementation combines convolution operations with different dilation rates (r = 3, 6, 12, 18, 24) through experiments. Studies show that dilation rates exceeding 24 lead to diminishing returns in feature modeling and increased computational complexity. Therefore, this implementation limits the maximum dilation rate to 24. As shown in Table 4, the proposed MCASPP outperforms other methods in accuracy performance, improving mIoU by 1.06% compared to the original ASPP architecture while only adding 12 milliseconds of computational latency, demonstrating an effective trade-off between efficiency and accuracy. The results validate that cascaded ASPP (6, 12, 18, 24) is superior to the original ASPP configuration. While cascaded ASPP (3, 6, 12, 18, 24) also offers high accuracy and efficiency, its model structure is more complex and it exhibits overfitting; therefore, this approach was not chosen.

[0135] Table 4 Ablation experiments of ASPP cascaded convolutions with different expansion rates

[0136]

[0137] To verify the impact of inserting attention mechanisms on the model's feature extraction capability, we used MobileNetV3 as the backbone feature extraction network and, based on the MCASPP module, inserted different attention mechanisms after convolution and after shallow features for experiments. This implementation verified the improvement in algorithm extraction accuracy by inserting several attention mechanisms: SE Net, ECA Net, CA Net, and CBAM Net. The results are shown in Table 5. Experimental results show that the model accuracy is improved after inserting attention mechanisms. Introducing CBAM Net results in higher accuracy compared to introducing SE Net and ECA Net, with mPA, accuracy, and mIoU of 89.21%, 97.91%, and 80.92%, respectively.

[0138] Table 5 Comparison of road extraction results with different attention mechanisms

[0139] Comparative experimental group mPA / % Accuracy / % mIoU / % SENet 88.96 97.55 79.38 ECANet 89.04 97.64 80.66 CANet 88.71 97.23 80.19 CBAMNet 89.21 97.91 80.92

[0140] This implementation proposes a lightweight road extraction network, DeepSegment, that integrates multi-layered cascaded dilated spatial pyramid pooling and attention mechanisms to address the problems of high computational complexity, weak road topological connectivity, and susceptibility to shadows and occlusion from surrounding objects in complex remote sensing scenarios. By using MobileNetV3 as the backbone network and combining dynamic channel calibration and gradient optimization strategies, the number of parameters is significantly reduced and the model inference speed is accelerated while maintaining the network's semantic representation capabilities. The MCASPP module constructs a continuously expanding global receptive field through cascaded dilated convolutions and dense skip connections, effectively solving the problem of multi-scale feature discretization and fusion in the traditional ASPP module, and significantly improving the extraction accuracy of fine roads and occluded areas. The CBAM dual-domain attention mechanism further enhances the model's ability to learn road morphological features by strengthening the road texture channel response and suppressing background noise. The Focal-Dice joint loss function is used to balance class weights and boundary attention, effectively alleviating the gradient vanishing problem caused by the low proportion of road pixels. Experiments on the DeepGlobe and CHN6-CUG datasets demonstrate that DeepSegment improves the accuracy and completeness of extraction results compared to mainstream models, while also meeting real-time processing requirements. This implementation provides an efficient and reliable solution for automated road extraction from high-resolution remote sensing imagery, with significant application value, particularly in natural resource surveys, urban road network updates, and emergency disaster relief planning. Future work will explore integrating other remote sensing data sources and their feature information, such as multispectral imagery, LiDAR point clouds, and morphological features, to further enhance the model's robustness under extreme lighting and occlusion scenarios.

[0141] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0142] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for road extraction from remote sensing images that integrates multi-scale and road topological features, characterized in that, Includes the following steps: S1. Establish a high-resolution remote sensing image road extraction model that integrates multi-scale and road topology features; S2. Standardize and resize the semantic segmentation data of remote sensing images; S3. Import the image data obtained in S2 into the high-resolution remote sensing image road extraction model obtained in S1 for extraction, and output the prediction results.

2. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 1, characterized in that, S1 specifically includes the following steps: S101. Improve the backbone network of DeepLabV3+ by replacing the original backbone network with the MobileNetV3 backbone network and introducing a squeeze-excitation module. S102. Establish a multi-layered cascaded hollow spatial pyramid pooling module that integrates convolutional attention modules; S103. Establish a composite loss function of Focal Loss and Dice Loss.

3. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 2, characterized in that, The MobileNetV3 backbone network described in S101 includes depthwise separable convolutional layers; the depthwise separable convolution decomposes the convolution operation into two independent steps: channel-wise convolution and pointwise convolution; channel-wise convolution achieves lightweight filtering by applying a single convolutional kernel to each input channel; pointwise convolution constructs new features by calculating a linear combination of the input channels.

4. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 2, characterized in that, The squeezing-excitation module described in S101 generates channel weight vectors through global average pooling operations and learns adaptive weight coefficients for each feature channel through a fully connected layer.

5. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 2, characterized in that, The MobileNetV3 backbone network described in S101 uses the h-swish activation function instead of the traditional ReLU6 activation function: Where x is the input feature value.

6. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 2, characterized in that, The multi-layer cascaded void spatial pyramid pooling module includes four 3×3 depth-separable convolutional layers and a convolutional attention module. The depth-separable convolutional layers include void convolution operations, and the void expansion rate increases stepwise.

7. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 6, characterized in that, The multi-layered cascaded void space pyramid pooling module is specifically as follows: in, y is a convolution operation with kernel size k and dilation rate v, and y is the output of a multi-layer cascaded hollow spatial pyramid pooling module; The dilated convolution operation specifically involves: Where x is the one-dimensional input, K is the kernel size, v is the dilation rate, w(k) represents the k-th parameter in the kernel, y′(i) is the output of the one-dimensional dilated convolution, and i is the index symbol; The size of the receptive field R of the dilated convolution is: R = (k-1) × (v-1) + k (4) For cascaded porous convolutional layers R1, R2, ... R n , n represents the nth dilated convolutional layer being cascaded, with a cumulative receptive field R. max The calculation formula is: R max =R1+R2+…R n -(n-1) (5)。 8. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 6, characterized in that, The convolutional attention module is embedded in each layer of the remote sensing image road extraction model; it is applied again after the feature fusion operation to form a multi-stage attention enhancement structure that runs through the forward propagation path of the network.

9. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 8, characterized in that, The convolutional attention module includes a channel attention module and a spatial attention module; the channel attention submodule determines the weight vector M through global average pooling and a fully connected layer. c (F): M c (F)<σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (6) Where F represents the input feature map; AvgPool and MaxPool represent average pooling and max pooling, respectively; MLP represents a multilayer perceptron; and σ represents the sigmoid activation function. The spatial attention submodule uses 7×7 convolution to generate a spatial weight matrix M. s (F): M s (F)=σ(f 7×7 ([AvgPool(F);MaxPool(F)])) (7) Among them, f 7×7 It is a 7×7 convolution.

10. The method for road extraction from remote sensing images that integrates multi-scale and road topological features according to claim 2, characterized in that, The composite loss function of Focal Loss and Dice Loss is as follows: L=L d +L f (8) Among them, L d L represents Dice loss. f Indicates Focal loss; The Dice loss, DiceLoss, is: DiceLoss=1-(2*|X∩Y| / (|X|+|Y|)) (9) Where X represents the segmented region of the predicted result, Y represents the ground truth region, and X∩Y represents the overlapping part between the predicted result and the ground truth region; The Focal loss, FocalLoss, is: Focal Loss=-α*(1-pt) ∧ γ*log(pt) (12) Where pt represents the classification confidence level; α is the positive sample weight coefficient; and γ is the hard sample focusing factor.

Images