A Road Extraction Method Based on Visible Light Remote Sensing Images

Abstract

This invention discloses a road extraction method based on visible light remote sensing images, belonging to the field of remote sensing image road extraction technology. The method includes the following steps: 1. Acquiring a remote sensing image dataset; 2. Preprocessing the remote sensing image dataset; 3. Constructing a road extraction model based on an improved SAM model; 4. Training and optimizing the parameters of the improved SAM model to obtain the optimal model; 5. Inputting the remote sensing image to be tested into the trained road extraction model to obtain the road extraction result. The method based on the improved SAM model of this invention can effectively enhance the ability to represent thin-line road features and model the spatial continuity of roads, thereby improving the accuracy and structural integrity of the road extraction results.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and computer vision technology, specifically to a road extraction method based on visible light remote sensing images. Background Technology

[0002] Road extraction from remote sensing images is a key technology in remote sensing information processing and geospatial data acquisition, and it has significant application value in areas such as map updating, traffic planning, urban management, and disaster assessment. With the development of remote sensing imaging technology and computer vision technology, remote sensing images are constantly evolving towards higher spatial and temporal resolution, providing an increasingly rich data foundation for the refined and intelligent extraction of road targets.

[0003] Currently, methods for road extraction from remote sensing images mainly include traditional feature-based methods and deep learning-based methods. Traditional methods rely heavily on edge, texture, grayscale, and morphological features, which often struggle to reliably extract complete road information in complex backgrounds and with diverse road shapes. In contrast, deep learning methods can automatically learn multi-level road features, achieving better results in road extraction.

[0004] However, existing methods still have certain limitations in complex scenarios. Roads in remote sensing images are typically characterized by their elongated distribution, significant width variations, complex orientations, and susceptibility to occlusion by trees, buildings, and shadows. These features can lead to insufficient response of thin-line road features, unclear road boundary delineation, and unstable road structure connections, thus affecting the accuracy and completeness of road extraction results. Therefore, enhancing the representational ability of thin-line road features and improving the modeling effect of road spatial continuity to improve the accuracy and structural integrity of road extraction from remote sensing images has become an urgent problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a road extraction method based on visible light remote sensing images, which solves the problems of insufficient representation of thin-line road features and weak ability to model road spatial continuity in existing methods, thereby improving the accuracy and structural integrity of road extraction results.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A remote sensing image dataset is acquired, which contains road targets and their corresponding annotation information, for training and testing the model.

[0007] The remote sensing image dataset is preprocessed to improve the consistency, adaptability, and model training stability of the sample data. The preprocessing includes image cropping, resizing, normalization, and data augmentation operations. The data augmentation operations include geometric transformations and non-geometric transformations, where geometric transformations include rotation and mirroring, and non-geometric transformations include hue, saturation, and brightness adjustments.

[0008] The preprocessed remote sensing image is input into the image encoder of the improved SAM model to extract multi-level image feature information. The image encoder is used to obtain road texture features, edge features and semantic features at different scales, providing feature support for subsequent road area identification and structure restoration.

[0009] The improved SAM model introduces a Dense Pyramid Feature Enhancement Module, which includes Deformable Convolution and Strip Pooling. This module aggregates and interacts features at different levels at multiple scales, enhancing the model's ability to represent road targets of varying widths (especially thin-line roads) while improving the fusion of local road details and contextual information.

[0010] The improved SAM model introduces a Criss-Cross Attention Module to model the dependencies of feature maps in the horizontal and vertical directions, enhance the ability to express long-distance associations in the road extension direction, improve the ability to maintain road spatial continuity, and alleviate the road breakage problem caused by road occlusion and shadow interference.

[0011] The features processed by the dense pyramid feature enhancement module and the cross-attention module are input into the mask decoder to generate road segmentation results. The mask decoder is used to restore the spatial distribution of road areas and output a road extraction result map corresponding to the input remote sensing image.

[0012] In the improved SAM model, a hybrid loss function, Road Loss, is introduced for network training and parameter updates, thus optimizing the network. After training, the remote sensing image to be tested is input into the trained model, and the road extraction results are output. The hybrid loss function, Road Loss, is composed of a weighted sum of a binary cross-entropy loss function and a Dice loss function, thereby enhancing the model's supervision of road area prediction results and improving the convergence ability and road segmentation accuracy during model training.

[0013] The beneficial effects of this invention are as follows: by introducing a dense pyramid feature enhancement module into the improved SAM model, and combining it with deformable convolution and strip pooling modules, the model's ability to represent thin-line roads, curved roads, and multi-scale road targets can be improved; by introducing a cross-attention module, the model's ability to model long-distance road dependencies and spatial continuous structures can be enhanced; the synergistic effect of the two can effectively improve the problems of insufficient representation of thin-line road features and insufficient ability to maintain road spatial continuity during the road extraction process, thereby improving the accuracy and structural integrity of the road extraction results. Attached Figure Description

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0015] Figure 1 This is a flowchart of a road extraction method based on visible light remote sensing images in an embodiment of the present invention; Figure 2 This is a schematic diagram of the improved SAM road extraction model structure of a road extraction method based on visible light remote sensing images in an embodiment of the present invention; Figure 3 This is a schematic diagram of a dense pyramid feature enhancement module for a road extraction method based on visible light remote sensing images in an embodiment of the present invention; Figure 4 This is a schematic diagram of a strip pooling structure for a road extraction method based on visible light remote sensing images in an embodiment of the present invention; Figure 5 This is a schematic diagram of the cross-attention module structure of a road extraction method based on visible light remote sensing images in an embodiment of the present invention. Detailed Implementation

[0016] 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.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application.

[0018] This embodiment provides a road extraction method based on visible light remote sensing images, mainly addressing problems such as missed extraction, road breaks, and local missegmentation caused by insufficient representation of thin-line road features and inadequate ability to maintain road spatial continuity in complex scenes. This method is based on the SAM model, introducing a dense pyramid feature enhancement module and a cross-attention module to improve the accuracy and structural integrity of the road extraction results.

[0019] This embodiment uses the CHN6-CUG public road dataset, which contains remote sensing image data of roads in typical urban scenes in China. The images are selected from high-resolution remote sensing images of the core urban areas of six cities—Beijing, Wuhan, Shanghai, Shenzhen, Hong Kong, and Macau—from the Google Earth platform. They cover various scenes, including main roads, secondary roads, winding roads, intersecting roads, and complex backgrounds, and can well reflect the characteristics of road targets in actual remote sensing images under conditions of scale changes, background interference, and occlusion.

[0020] In the data preprocessing stage, the remote sensing image dataset is first subjected to image cropping, resizing, and normalization to meet network input requirements and improve training stability. Specifically, the original images are cropped into fixed-size image patches while maintaining their spatial correspondence with the labels; normalization is used to reduce the differences in numerical distribution between different images, improving model convergence efficiency. To further enhance the model's adaptability and generalization ability to complex scenes, data augmentation operations, including geometric and non-geometric transformations, are performed on the sample images during the training phase to improve the model's robustness to changes in road direction, lighting conditions, and background interference.

[0021] After preprocessing, the remote sensing image is input into the image encoder of the improved SAM model for feature extraction. The image encoder extracts multi-level image feature information from the input image, including road edge features, texture features, and semantic features. Shallow features are primarily used to preserve road boundaries and local texture details, while deep features are mainly used to enhance the semantic distinction between road areas and background areas, thus providing feature support for subsequent road area recognition and structure restoration.

[0022] In the improved SAM model, a dense pyramid feature enhancement module is introduced to perform multi-scale aggregation and interaction of features at different levels. For example... Figure 3As shown, this module employs a dual-path parallel structure of densely dilated convolutional cascade branches and strip pooling branches, and performs feature fusion at the output end through a concatenation operation. The densely dilated convolutional cascade branches use five sets of 3×3 deformable convolutions with sequentially increasing dilation rates (d=3, 6, 12, 18, 24). The outputs of each branch are reused as inputs for subsequent branches and concatenated level by level, achieving collaborative fusion of multi-scale features and effectively mitigating the grid effect. The deformable convolutions introduce learnable sampling offsets, improving the feature extraction capability for curved and irregular roads. The strip pooling branches use 1×N and N×1 long strip pooling kernels to establish long-distance contextual dependencies along the horizontal and vertical directions, respectively, effectively capturing the overall road orientation. The outputs of the two branches are concatenated and fused along the channel dimension, allowing the model to simultaneously consider both local road details and global structure, improving road extraction accuracy.

[0023] After completing the dense pyramid feature enhancement, the features are input into the cross-attention module for road spatial continuity modeling. For example... Figure 5 As shown, this module enables rapid information propagation along the same row or column by performing attention calculations in the horizontal and vertical directions of the feature map. By stacking two cross-attention operations, an indirect long-distance association can be established between any two pixel locations, thereby achieving near-global context modeling and significantly improving the ability to maintain the continuity of road space.

[0024] The enhanced features, processed by the dense pyramid feature enhancement module and the cross-attention module, are input into the mask decoder to generate road segmentation results. The mask decoder performs spatial reconstruction and pixel-level classification on the enhanced feature information, progressively restoring the high-dimensional feature mapping to a road prediction map corresponding to the input image. Simultaneously, the cue encoder receives the default cue embedding vector as cue input and outputs cue embedding features, which, along with the image features extracted by the image encoder, are fed into the mask decoder to maintain the integrity of the overall structure of the improved SAM model. In the final output road prediction map, road regions are classified as target categories, and non-road regions are classified as background categories, resulting in a road extraction map with a relatively complete structure and clear boundaries.

[0025] During the model training phase, this embodiment introduces a hybrid loss function, Road Loss, into the improved SAM model for network training and parameter updates. The hybrid loss function Road Loss is a weighted sum of the binary cross-entropy loss function and the Dice loss function. Its specific formula is as follows: Where α and β are weighting coefficients, and α > 0 and β > 0.

[0026] The binary cross-entropy loss function is used to constrain the pixel-level classification results of the road region and the background region, improving the accuracy of the model's classification of each pixel. Its specific formula is as follows: Where N represents the total number of pixels, yi represents the true label of the i-th pixel, and pi represents the predicted probability that the i-th pixel belongs to the road category.

[0027] The Dice loss function measures the overlap between the predicted results and the true labels, enhancing the model's ability to learn the overall road structure. It is particularly helpful in improving the segmentation of thin linear roads and small-scale road regions. Its specific formula is as follows: Here, ε is a smoothing term used to avoid the denominator being zero.

[0028] This invention achieves collaborative modeling of local road detail features and global continuous structural information by introducing a dense pyramid feature enhancement module and a cross-attention module into the improved SAM model. The dense pyramid feature enhancement module effectively improves the model's ability to represent thin-line roads and multi-scale road targets, while the cross-attention module enhances the model's ability to model long-distance road dependencies and spatial continuous structures. The synergistic effect of the two helps to improve the accuracy and structural integrity of the road extraction results.

[0029] The experimental environment for this experiment was as follows: Intel Core i9-12900H processor, GeForce RTX3090 graphics card, 1024 GB PCIe 4.0 NVMe solid-state drive, Windows 10 Professional operating system, and Python 3.8 and PyTorch 2.0 development environment.

[0030] The above embodiments are merely one specific implementation of the present invention, and their descriptions are quite specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and all such modifications and improvements fall within the protection scope of the present invention.

Claims

1. A road extraction method based on visible light remote sensing images, characterized in that, Includes the following steps: S1: Obtain the remote sensing image dataset; S2: Preprocess the remote sensing image dataset; S3: Construct a road extraction model based on an improved SAM model, which includes an image encoder, a dense pyramid feature enhancement module, a cross-attention module, a cue encoder, and a mask decoder; S4: Train and optimize the parameters of the improved SAM model to obtain the trained road extraction model; S5: Input the remote sensing image to be tested into the trained road extraction model to obtain the road extraction results.

2. The road extraction method based on visible light remote sensing images according to claim 1, characterized in that, The remote sensing image dataset uses the CHN6-CUG public road dataset, and the corresponding road labels were obtained manually.

3. The road extraction method based on visible light remote sensing images according to claim 1, characterized in that, Data preprocessing includes image cropping, resizing, normalization, and data augmentation.

4. The road extraction method based on visible light remote sensing images according to claim 1, characterized in that, The image encoder is used to extract road texture features, edge features, and semantic features at different scales, thereby providing feature support for subsequent road area recognition and structure restoration.

5. The road extraction method based on visible light remote sensing images according to claim 1, characterized in that, The dense pyramid feature enhancement module includes deformable convolution and strip pooling modules. The deformable convolution is used to adaptively adjust the convolution sampling position according to the road geometry to enhance the feature extraction capability for curved roads, intersecting roads, and roads with irregular shapes. The strip pooling module is used to obtain long-distance contextual information extending laterally or longitudinally to enhance the modeling capability for road linear structures and continuously extending features.

6. The road extraction method based on visible light remote sensing images according to claim 1, characterized in that, The cross-attention module is used to model the dependencies of feature maps in the horizontal and vertical directions to enhance the ability to express long-distance associations in the road extension direction and improve the ability to maintain the spatial continuity of the road.

7. The road extraction method based on visible light remote sensing images according to claim 1, characterized in that, The improved SAM model is trained and its parameters are optimized using a hybrid loss function, Road Loss, which is a weighted sum of a binary cross-entropy loss function and a Dice loss function.

Images